Learning quadrupedal locomotion over challenging terrain
Open accessAbstract
A learning-based locomotion controller enables a quadrupedal ANYmal robot to traverse challenging natural environments.
AI snapshot
Written from the paper’s full text.
TL;DR
The paper presents a reinforcement-learning controller for blind quadrupedal locomotion that uses only proprioceptive history and transfers zero-shot from simulation to challenging natural terrain. It was deployed on two generations of ANYmal robots across mud, snow, vegetation, water, rubble, and the DARPA Subterranean Challenge Urban Circuit without environment-specific tuning.
Key findings
- In moss, mud, and vegetation, the presented controller achieved higher average speeds than a state-of-the-art baseline: 0.452, 0.338, and 0.248 m/s versus 0.199, 0.197, and not reported for vegetation; mechanical cost of transport was lower: 0.423, 0.692, and 1.23 versus 0.625, 0.931, and not reported.
- The controller showed zero-shot generalization to conditions never encountered in training, including deformable terrain, dynamic footholds, and overground impediments; it drove two ANYmal-B robots in four 60-minute DARPA Subterranean Challenge Urban Circuit missions with a zero failure rate.
- It developed an emergent foot-trapping reflex: it overcame a 16.8 cm step, with maximum foot clearance on flat terrain of 12.9 cm and 13.6 cm for LF/RF legs increasing to 22.5 cm and 18.5 cm during foot-trapping; with a 10 kg payload (22.7% of total weight) it traversed steps up to 13.4 cm, while the baseline could not traverse any steps with the payload.
- Ablations showed longer proprioceptive memory improved robustness: under a 50 N lateral force for 5 s, TCN-100 deviation was 35.5% lower than TCN-1; privileged training and adaptive terrain curriculum were both important for learning and performance.
Methods
- Two-stage privileged learning: a teacher MLP was trained via TRPO in simulation with privileged terrain and contact information, then distilled via DAgger into a proprioceptive student temporal convolutional network (TCN) using a history of proprioceptive observations; the controller used PMTG motion primitives, analytic inverse kinematics, and joint PD control, with no exteroceptive input.
- Training used RaiSim with learned actuator models, domain randomization, and an adaptive terrain curriculum that used particle filtering over parameterized hills, steps, and stairs; the student policy was deployed directly on physical ANYmal-B and ANYmal-C robots at 400 Hz onboard.
- Evaluation included natural-environment deployments, DARPA Subterranean Challenge missions, indoor debris and step tests, payload tests, foot-slippage tests, and ablation studies with 5 random seeds and 100 trials per diagnostic test.
Limitations
- The paper states that the controller only exhibits the trot gait, narrower than gaits seen in nature, and relies solely on proprioception, so blind locomotion is inherently limited—for example, it will walk off a cliff if commanded, and its gait is conservative because it must feel out the environment.
- The paper states that training uses only rigid terrains and a small set of procedurally generated profiles, with no compliance or overground obstructions such as vegetation.
- The text makes evident that the quantitative baseline comparison understates the difference between controllers because baseline failures were reset by a human operator and catastrophic failures were not factored into the speed and efficiency measurements.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 62 cited sources checked · 0 supported · 0 flagged
- Findingown data§ 1. INTRODUCTION · p. 3
The controller reliably trots through mud, sand, rubble, thick vegetation, snow, running water, and a variety of other off-road terrain.
“The controller reliably trots through mud, sand, rubble, thick vegetation, snow, running water, and a variety of other off-road terrain.”
Quote in the paperSubstantial claim - Findingown data§ 1. INTRODUCTION · p. 3
Realistic modeling and parameter randomization were not sufficient to achieve robust locomotion on rough terrain.
“We use these ideas as well, but have found that they were not sufficient to achieve robust locomotion on rough terrain.”
Quote in the paperSubstantial claim - Findingown data§ Natural environments · p. 5
The presented controller achieves higher locomotion speed than the baseline in all conditions: moss, mud, and vegetation.
“The presented controller achieves higher locomotion speed in all conditions.”
Quote in the paperSubstantial claim - Findingcitation§ 1. INTRODUCTION · p. 3
The controller was used to drive two generations of ANYmal quadrupeds in a variety of conditions that are beyond the reach of prior published work in legged robotics.
“The controller was used to drive two generations of ANYmal quadrupeds [20] in a variety of conditions that are beyond the reach of prior published work in legged robotics.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [20] 20. M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V. Tsounis, J. Hwangbo, K. Bodie, P . Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, M. A. Höpflinger, ANYmal - a highly mobile and dynamic quadrupedal robot, IEEE/RSJ International Conference on Intelligent Robots and Systems , 38-44 (IEEE, 2016). source
Not checked in the free run.
- Findingown data§ Natural environments · p. 3
The presented controller has been deployed in diverse natural environments including steep mountain trails, creeks with running water, mud, thick vegetation, loose rubble, snow-covered hills, and a damp forest.
“The presented controller has been deployed in diverse natural environments, as shown in Fig. 1 and Movies 1 and S1. These include steep mountain trails, creeks with running water, mud, thick vegetation, loose rubble, snow-covered hills, and a damp forest.”
Quote in the paperSubstantial claim
- Findingreasoning§ 1. INTRODUCTION · p. 3
Privileged information, namely ground-truth knowledge of the terrain and the robot's contact with it, enables the teacher policy to quickly achieve high performance.
“The privileged information enables the policy to quickly achieve high performance.”
Quote in the paperSubstantial claim - Findingown data§ Further analysis of emergent behavior · p. 12
The saliency value at the foot-trapping event is kept high while stepping up.
“In Fig. 6C we can see that the saliency value at the foot-trapping event (FT) is kept high while stepping up.”
Quote in the paperNot a substantial claim - Findingreasoning§ Adaptive terrain curriculum · p. 12
Throughout the training process, the mean episode length is shorter for the model being trained without adaptive curriculum.
“Throughout the training process, the mean episode length is shorter for the model being trained without adaptive curriculum (Fig. 5J).”
Quote in the paperSubstantial claim - Findingown data§ Further analysis of emergent behavior · p. 12
The estimated elevations and normal vectors adapt to the step during the foot-trapping reflex.
“The estimated elevations and normal vectors adapt to the step during the foot-trapping reflex (iii+iv).”
Quote in the paperNot a substantial claim - Findingown data§ Indoor experiments · p. 5
The learned controller manifests a foot-trapping reflex: the policy identifies the trapping of the foot purely from proprioceptive observations and lifts the foot over the obstacle.
“The learned controller manifests a foot-trapping reflex, as shown in Fig. 3B and Movie S3. The policy identifies the trapping of the foot purely from proprioceptive observations and lifts the foot over the obstacle.”
Quote in the paperSubstantial claim - Findingown data§ Memory in proprioceptive control · p. 12
Controllers with longer memory are more robust to external disturbances.
“Fig. 5D shows that controllers with longer memory are more robust to external disturbances.”
Quote in the paperSubstantial claim - Findingown data§ Adaptive terrain curriculum · p. 12
A teacher trained without adaptive curriculum plateaus at a lower reward level.
“Fig. 5I shows that a teacher trained without adaptive curriculum plateaus at a lower reward level.”
Quote in the paperSubstantial claim - Findingown data§ Privileged training · p. 12
The baseline TCN-20 policy trained directly without privileged training fails the diagnostic tests: it is incapable of locomoting on a slope or traversing a step.
“Fig. 5E shows that the baseline fails the diagnostic tests: it is incapable of locomoting on a slope or traversing a step.”
Quote in the paperSubstantial claim - Findingown data§ Indoor experiments · p. 6
The baseline is incapable of traversing any steps under any command speed with the payload.
“The baseline is incapable of traversing any steps under any command speed with the payload.”
Quote in the paperSubstantial claim - Findingown data§ Privileged training · p. 12
The baseline does not reach comparable reward during training as the teacher MLP architecture with privileged information or the proprioceptive TCN-20 architecture trained via privileged learning.
“Fig. 5F shows that the baseline does not reach comparable reward during training as the teacher MLP architecture with privileged information or the proprioceptive TCN-20 architecture (same as the baseline, no privileged information) trained via privileged learning.”
Quote in the paperSubstantial claim - Findingown data§ Adaptive terrain curriculum · p. 12
The success rates on the testing terrains are significantly lower when trained without the adaptive curriculum.
“The success rates on the testing terrains are significantly lower when trained without the adaptive curriculum, as shown in Fig. 5H.”
Quote in the paperSubstantial claim - Findingown data§ Memory in proprioceptive control · p. 12
Memory length doesn't have a strong effect in the uniform slope setting.
“As shown in Fig. 5B, memory length doesn't have a strong effect in the uniform slope setting.”
Quote in the paperNot a substantial claim - Findingreasoning§ A. Overview · p. 8
After training in simulation, the proprioceptive controller is deployed directly on physical legged machines, with no fine-tuning.
“After training in simulation, the proprioceptive controller is deployed directly on physical legged machines, with no fine-tuning.”
Quote in the paperSubstantial claim - Findingown data§ Further analysis of emergent behavior · p. 12
The policy attends to the LF leg joint states measured at the foot-trapping event.
“The policy attends to the LF leg joint states measured at the foot-trapping event.”
Quote in the paperNot a substantial claim - Findingown data§ Memory in proprioceptive control · p. 12
The failure rate of limited-memory controllers is particularly high when the hind legs encounter the step.
“As shown in Fig. 5C, the failure rate of limited-memory controllers is particularly high when the hind legs encounter the step.”
Quote in the paperSubstantial claim - Findingreasoning§ Motion synthesis · p. 8
Defining the output in the horizontal frames results in less premature termination at the beginning of the policy training, when the base motion is unstable due to random actions.
“Defining the output in Hi results in less premature termination at the beginning of the policy training, when the base motion is unstable due to random actions.”
Quote in the paperSubstantial claim - Findingown data§ Memory in proprioceptive control · p. 12
Controllers with longer memory adapt hind leg trajectories to ensure higher foot clearances.
“Controllers with longer memory also adapt hind leg trajectories to ensure higher foot clearances.”
Quote in the paperSubstantial claim - Findingown data§ Further analysis of emergent behavior · p. 12
After the successful step-up, the terrain uncertainty remains elevated.
“After the successful step-up, the terrain uncertainty remains elevated (v), indicating an anticipation of generally rough terrain.”
Quote in the paperNot a substantial claim - Findingown data§ Memory in proprioceptive control · p. 12
Memory length does have a strong effect on the controller's ability to traverse a step.
“Memory length does have a strong effect on the controller's ability to traverse a step (Fig. 5B-C).”
Quote in the paperSubstantial claim - Findingown data§ Memory in proprioceptive control · p. 12
The deviation of the TCN-100 controller was 35.5% lower than that of TCN-1 under a 50 N lateral force applied for 5 s during a straight walk.
“The deviation of the TCN-100 controller was 35.5 % lower than that of TCN-1.”
Quote in the paperSubstantial claim - Findingnone given§ Motion synthesis · p. 8
The controller does not use any exteroceptive input (e.g., no haptic sensors, cameras, or depth sensors).
“The controller does not use any exteroceptive input (e.g., no haptic sensors, cameras, or depth sensors).”
Quote in the paperSubstantial claim - Findingown data§ Memory in proprioceptive control · p. 12
Controllers with longer memory are able to handle higher steps.
“Controllers with longer memory are able to handle higher steps.”
Quote in the paperSubstantial claim - Findingown data§ S6. Qualitative evaluation of the adaptive terrain curriculum · p. 16
Undesirable terrains are either too easy or too difficult.
“Undesirable terrains are either too easy or too difficult, as shown in the leftmost and rightmost panels of Fig. S1A.”
Quote in the paperNot a substantial claim - Findingown data§ S7. Reconstruction of the privileged information in different situations · p. 16
While traversing dense vegetation, the decoder detects a force opposite the motion direction.
“While traversing dense vegetation as shown in Fig. S2C, it detects a force opposite the motion direction”
Quote in the paperNot a substantial claim - Findingown data§ S8. Recurrent neural network student policy · p. 16
The GRU-based controller fails to achieve the performance of TCN-100 in step experiments.
“the GRU-based controller fails to achieve the performance of TCN-100 in step experiments”
Quote in the paperNot a substantial claim - Findingown data§ S8. Recurrent neural network student policy · p. 16
The performance of the GRU-based controller is between that of TCN-20 and TCN-100.
“Overall, the performance of the GRU-based controller is between that of TCN-20 and TCN-100.”
Quote in the paperNot a substantial claim - Findingown data§ S8. Recurrent neural network student policy · p. 17
The training time for the TCN is much faster in comparison to the GRU.
“The training time for the TCN is much faster in comparison to the GRU.”
Quote in the paperSubstantial claim - Findingown data§ S7. Reconstruction of the privileged information in different situations · p. 16
The decoder detects downward external force when an unknown 10 kg payload is applied.
“As shown in Fig. S2B, the decoder detects downward external force when an unknown 10 kg payload is applied.”
Quote in the paperNot a substantial claim - Findingown data§ S6. Qualitative evaluation of the adaptive terrain curriculum · p. 16
The particle filter rejects parameters that represent short and steep steps.
“The particle filter rejects parameters that represent short and steep steps (upper-left area).”
Quote in the paperNot a substantial claim- The authors state this as an assumption: Within Nevaluate iterations, the performance of the policy is similar.
- Findingown data§ S8. Recurrent neural network student policy · p. 16
The performance of the GRU-based controller is comparable to TCN-100 in the slope setting.
“The performance is comparable to TCN-100 in the slope setting”
Quote in the paperNot a substantial claim - Findingown data§ S7. Reconstruction of the privileged information in different situations · p. 16
The uncertainty of the elevation estimates is notably high in the natural terrains shown in Fig. S2C and Fig. S2D.
“The uncertainty of the elevation estimates are notably high in the natural terrains shown in Fig. S2C and Fig. S2D”
Quote in the paperSubstantial claim - Findingown data§ S9. Ablation of the latent representation loss for student training · p. 17
The ablated version has lower success rates on steps.
“On the other hand, the ablated version has lower success rates on steps.”
Quote in the paperNot a substantial claim - Premisecitation§ Natural environments · p. 3
Existing terrain estimation pipelines that use cameras or LiDAR fail in environments with snow, water, or dense vegetation.
“Existing terrain estimation pipelines that use cameras or LiDAR [27] fail in environments with snow (Fig. 2A), water (Fig. 2C), or dense vegetation (Fig. 2F).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [27] 27. P. Fankhauser, M. Bloesch, C. Gehring, M. Hutter, R. Siegwart. Robotcentric elevation mapping with uncertainty estimates. Mobile Service Robotics (World Scientific, 2014), 433-440. source
Not checked in the free run.
- Premisecitation§ 1. INTRODUCTION · p. 3
Application of reinforcement learning to legged locomotion has largely been confined to laboratory environments and conditions.
“However, application of RL to legged locomotion has largely been confined to laboratory environments and conditions.”
Quote in the paperSubstantial claim0 of 7 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019). source
Not checked in the free run.
- cites [13] 13. T. Haarnoja, S. Ha, A. Zhou, J. Tan, G. Tucker, S. Levine, Learning to walk via deep reinforcement learning, Robotics: Science and Systems (2019). source
Not checked in the free run.
- cites [14] 14. Z. Xie, P . Clary, J. Dao, P . Morais, J. Hurst, M. van de Panne, Learning locomotion skills for Cassie: Iterative design and sim-to-real, Conference on Robot Learning (2019).
Not checked in the free run.
- cites [16] 16. J. Tan, T. Zhang, E. Coumans, A. Iscen, Y. Bai, D. Hafner, S. Bohez, V. Vanhoucke, Sim-to-real: Learning agile locomotion for quadruped robots, Robotics: Science and Systems (2018). source
Not checked in the free run.
- cites [17] 17. Y. Yang, K. Caluwaerts, A. Iscen, T. Zhang, J. Tan, V. Sindhwani, Data efficient reinforcement learning for legged robots, Conference on Robot Learning (2019).
Not checked in the free run.
- cites [18] 18. S. Ha, P . Xu, Z. Tan, S. Levine, J. Tan, Learning to walk in the real world with minimal human effort, arXiv:2002.08550 (2020).
Not checked in the free run.
- cites [19] 19. X. B. Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, S. Levine, Learning agile robotic locomotion skills by imitating animals, arXiv:2004.00784 (2020). source
Not checked in the free run.
- Interpretationown data§ 1. INTRODUCTION · p. 3
The presented controller is a radically robust controller for blind quadrupedal locomotion on challenging terrain.
“Here we present a radically robust controller for blind quadrupedal locomotion on challenging terrain.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 1. INTRODUCTION · p. 1
Under challenging conditions, the robot must rely crucially on proprioception.
“Under these conditions, the robot must rely crucially on proprioception - the sensing of its own bodily configuration at high temporal resolution.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 1. INTRODUCTION · p. 3
The automated curriculum that synthesizes terrains adaptively based on the controller's performance at different stages of training has proven important in achieving the presented levels of robustness.
“The third idea that has proven important in achieving the presented levels of robustness is an automated curriculum that synthesizes terrains adaptively, based on the controller's performance at different stages of the training process.”
Quote in the paperSubstantial claim - Premisecitation§ 1. INTRODUCTION · p. 1
Conventional controllers for legged locomotion are based on elaborate state machines that explicitly trigger the execution of motion primitives and reflexes.
“Many rely on elaborate state machines that coordinate the execution of motion primitives and reflex controllers [1-5].”
Quote in the paperSubstantial claim0 of 5 sources checked- cites [1] 1. F. Jenelten, J. Hwangbo, F . Tresoldi, C. D. Bellicoso, M. Hutter, Dynamic locomotion on slippery ground, IEEE Robotics and Automation Letters 4170-4176 (2019). source
Not checked in the free run.
- cites [2] 2. G. Bledt, P . M. Wensing, S. Ingersoll, S. Kim, Contact model fusion for
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [3] source
Not checked in the free run.
- cites [4] 4. J. Reher, W. Ma, A. D. Ames, Dynamic walking with compliance on a Cassie bipedal robot, European Control Conference , 2589-2595 (IEEE, 2019). source
Not checked in the free run.
- cites [5] 5. Y. Gong, R. Hartley, X. Da, A. Hereid, O. Harib, J. Huang, J. W. Grizzle, Feedback control of a Cassie bipedal robot: Walking, standing, and riding a Segway, American Control Conference , 4559-4566 (IEEE, 2019). source
Not checked in the free run.
- Interpretationnone given§ 1. INTRODUCTION · p. 3
The result is a legged locomotion controller that is far more robust than any counterparts in existing literature.
“The result is a legged locomotion controller that is far more robust than any counterparts in existing literature.”
Quote in the paperNot a substantial claim - Premisenone given§ 1. INTRODUCTION · p. 1
To date, no published work has demonstrated dynamic locomotion in diverse, challenging natural environments as shown in Fig. 1.
“To date, no published work has demonstrated dynamic locomotion in diverse, challenging natural environments as shown in Fig. 1.”
Quote in the paperNot a substantial claim0 of 5 sources checked- cites [1] 1. F. Jenelten, J. Hwangbo, F . Tresoldi, C. D. Bellicoso, M. Hutter, Dynamic locomotion on slippery ground, IEEE Robotics and Automation Letters 4170-4176 (2019). source
Not checked in the free run.
- cites [2] 2. G. Bledt, P . M. Wensing, S. Ingersoll, S. Kim, Contact model fusion for
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [3] source
Not checked in the free run.
- cites [4] 4. J. Reher, W. Ma, A. D. Ames, Dynamic walking with compliance on a Cassie bipedal robot, European Control Conference , 2589-2595 (IEEE, 2019). source
Not checked in the free run.
- cites [5] 5. Y. Gong, R. Hartley, X. Da, A. Hereid, O. Harib, J. Huang, J. W. Grizzle, Feedback control of a Cassie bipedal robot: Walking, standing, and riding a Segway, American Control Conference , 4559-4566 (IEEE, 2019). source
Not checked in the free run.
- Interpretationreasoning§ Adaptive terrain curriculum · p. 12
The adaptive curriculum modulates the difficulty of sampled terrains so as to maximize the didactic benefit of each episode.
“The adaptive curriculum modulates the difficulty of sampled terrains so as to maximize the didactic benefit of each episode.”
Quote in the paperSubstantial claim- The authors state this as an assumption: Terrain parameters with similar traversability are close in Euclidean distance in parameter space.
- The authors state this as an assumption: A policy trained over the terrains generated by cTs in some area of C will learn to interpolate to nearby terrain parameters.
- The authors state this as an assumption: The evolution of each terrain parameter is a Markov process (only relies on the current value and randomly sampled noise).
- Premisecitation§ Teacher policy · p. 8
Including joint position errors and velocities measured at -0.01 s and -0.02 s in the state allows the policy to exploit the actuator dynamics.
“Joint position errors and velocities measured at -0.01 s and -0.02 s are contained in ot , which is the same as the input to the learned model of the joint-level PD controller. This information allows the policy to exploit the actuator dynamics [12].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019). source
Not checked in the free run.
- Premisecitation§ 3. DISCUSSION · p. 6
Present-day technology is severely limited in its ability to simulate compliant contact, slippage, and deformable and crumbling terrain; phenomena such as mud, snow, thick vegetation, and gushing water are beyond the capabilities of robotics simulation frameworks.
“Present-day technology is severely limited in its ability to simulate compliant contact, slippage, and deformable and crumbling terrain. As a result, phenomena such as mud, snow, thick vegetation, gushing water, and many others are beyond the capabilities of robotics simulation frameworks [30-32].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [30] 30. J. Hwangbo, J. Lee, M. Hutter, Per-contact iteration method for solving contact dynamics, IEEE Robotics and Automation Letters 895-902 (2018). source
Not checked in the free run.
- cites [31] 31. E. Coumans, others, Bullet physics library, Open source: bulletphysics. org (2013).
Not checked in the free run.
- cites [32] 32. R. Smith, others, Open dynamics engine, Open source: ode.org (2005).
Not checked in the free run.
- Interpretationreasoning§ Adaptive terrain curriculum · p. 9
Our method realizes a training curriculum that gradually modifies a distribution over environmental parameters such that the policy can continuously improve locomotion skills and generalize to new environments.
“Our method likewise realizes a training curriculum that gradually modifies a distribution over environmental parameters such that the policy can continuously improve locomotion skills and generalize to new environments.”
Quote in the paperSubstantial claim- The authors state this as an assumption: Terrain parameters with similar traversability are close in Euclidean distance in parameter space.
- The authors state this as an assumption: A policy trained over the terrains generated by cTs in some area of C will learn to interpolate to nearby terrain parameters.
- The authors state this as an assumption: The evolution of each terrain parameter is a Markov process (only relies on the current value and randomly sampled noise).
- Interpretationown data§ Further analysis of emergent behavior · p. 12
The elevated terrain uncertainty after step-up indicates an anticipation of generally rough terrain.
“After the successful step-up, the terrain uncertainty remains elevated (v), indicating an anticipation of generally rough terrain.”
Quote in the paperNot a substantial claim - Interpretationown data§ Further analysis of emergent behavior · p. 12
The ability to reconstruct explicit environmental information from the encoding of the proprioceptive history is a strong indicator that the TCN policy learns to build an internal representation of the environment and uses it for decision making.
“The ability to reconstruct explicit environmental information from the encoding of the proprioceptive history is a strong indicator that the TCNpolicy learns to build an internal representation of the environment and uses it for decision making.”
Quote in the paperNot a substantial claim - Premisecitation§ A. Overview · p. 8
Using a learned dynamics model of the robot's joint PD controller in simulation facilitates the transfer of policies from simulation to reality.
“The simulation uses a learned dynamics model of the robot's joint PD controller [12]. This facilitates the transfer of policies from simulation to reality.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019). source
Not checked in the free run.
- Interpretationown data§ Indoor experiments · p. 6
The presented controller is much more robust to model mismatch.
“We conclude that the presented controller is much more robust to model mismatch.”
Quote in the paperSubstantial claim - Interpretationown data§ 3. DISCUSSION · p. 6
Our work demonstrates that simulating the astonishing variety of the physical world may not be necessary.
“Our work demonstrates that simulating the astonishing variety of the physical world may not be necessary.”
Quote in the paperSubstantial claim - Interpretationreasoning§ Adaptive terrain curriculum · p. 12
Uniform sampling is more likely to draw terrains that cannot be successfully traversed by the policy being trained, causing the policy to fail early and receive less training signal.
“This is because uniform sampling is more likely to draw terrains that cannot be successfully traversed by the policy being trained. On these terrains, the policy fails early and receives less training signal as a result.”
Quote in the paperSubstantial claim - Interpretationown data§ S7. Reconstruction of the privileged information in different situations · p. 16
The TCN policy encodes the roughness of the terrain.
“which indicates that the TCN policy encodes the roughness of the terrain.”
Quote in the paperNot a substantial claim - Interpretationown data§ S8. Recurrent neural network student policy · p. 17
The chief advantage of the TCN is in training efficiency.
“The chief advantage of the TCN is in training efficiency.”
Quote in the paperNot a substantial claim - Interpretationown data§ S7. Reconstruction of the privileged information in different situations · p. 16
The detected force opposite the motion direction makes the policy counteract and push through the vegetation.
“which makes the policy to counteract and push through the vegetation.”
Quote in the paperNot a substantial claim
- May combine two claims. 38
- The quote only partly states this claim. 2
- Classified as probably not a substantial claim. 15
May combine two claims.
Findingown data§ Natural environments · p. 5The baseline could traverse flat and unobstructed patches but failed frequently upon encountering loose branches, thick vegetation, and mud; our controller never failed in these experiments.
“The baseline could traverse flat and unobstructed patches, but failed frequently upon encountering loose branches, thick vegetation, and mud, as shown in Movie S1. Our controller never failed in these experiments.”
Quote in the paperSubstantial claimThe quote only partly states this claim.
Findingown data§ Indoor experiments · p. 5The presented controller can stably locomote over shifting support surfaces in an indoor environment populated by loose debris.
“The presented controller can stably locomote over shifting support surfaces.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ Natural environments · p. 5The presented controller is more energy-efficient, with a lower cost of transport than the baseline.
“As shown in Table 1, the presented controller is more energy-efficient, with a lower COT than the baseline.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ 1. INTRODUCTION · p. 3The controller is consistently effective in zero-shot generalization settings; it remains robust when tested in conditions that have never been encountered during training.
“Remarkably, the controller is consistently effective in zero-shot generalization settings. That is, it remains robust when tested in conditions that have never been encountered during training.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ 1. INTRODUCTION · p. 3When deployed on physical quadrupeds, the controller successfully handles deformable terrain such as mud, moss, and snow, dynamic footholds such as a rolling board or debris, and overground impediments such as thick vegetation, rubble, and gushing water.
“Yet when deployed on physical quadrupeds, the controller successfully handles deformable terrain (mud, moss, snow), dynamic footholds (stepping on a rolling board in a cluttered indoor environment, or debris in the field), and overground impediments such as thick vegetation, rubble, and gushing water.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ 1. INTRODUCTION · p. 3Training a rough-terrain locomotion policy directly via reinforcement learning was not successful: the supervisory signal was sparse and the network failed to learn locomotion within reasonable time budgets.
“We have found that training a rough-terrain locomotion policy directly via reinforcement learning was not successful: the supervisory signal was sparse and the presented network failed to learn locomotion within reasonable time budgets.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ DARPA Subterranean Challenge · p. 5The presented controller drove two ANYmal-B robots in four missions of 60 minutes and exhibited a zero failure rate throughout the DARPA Subterranean Challenge Urban Circuit.
“The presented controller drove two ANYmal-B robots in four missions of 60 minutes. The controller exhibited a zero failure rate throughout the competition.”
Quote in the paperSubstantial claimMay combine two claims.
Findingreasoning§ 1. INTRODUCTION · p. 3The privileged learning protocol is enabled by simulation, but the resulting proprioceptive policy is deployed on physical machines.
“This privileged learning protocol is enabled by simulation, but the resulting proprioceptive policy is not confined to simulation and is deployed on physical machines.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ Indoor experiments · p. 6The presented controller adapts to slippery terrain and successfully locomotes in the commanded direction, whereas the baseline quickly loses balance, aggressively swings the legs, and falls.
“The baseline quickly loses balance, aggressively swings the legs, and falls. In contrast, the presented controller adapts to the slippery terrain and successfully locomotes in the commanded direction.”
Quote in the paperSubstantial claimMay combine two claims. The quote only partly states this claim.
Findingown data§ Indoor experiments · p. 6With a 10 kg payload, which is 22.7% of the total weight of the robot and was never simulated during training, the presented controller can still traverse steps up to 13.4 cm despite the model mismatch.
“As shown in Fig. 3E, the presented controller can still traverse steps up to 13.4 cm despite the model mismatch.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ Privileged training · p. 12The baseline fails to learn to balance and locomote, as indicated by mean episode length during training.
“Fig. 5G shows the mean episode length during training, which indicates that the baseline fails to learn to balance and locomote.”
Quote in the paperSubstantial claimMay combine two claims.
Findingnone given§ 3. DISCUSSION · p. 6Our training environment features only rigid terrain, with no compliance or overground obstructions such as vegetation; nevertheless, controllers trained in this environment successfully meet the diversity of field conditions encountered at deployment.
“Our training environment features only rigid terrain, with no compliance or overground obstructions such as vegetation. Nevertheless, controllers trained in this environment successfully meet the diversity of field conditions encountered at deployment.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Findingown data§ Motion synthesis · p. 8The controller input does not contain any handcrafted features such as foot contact states or estimated terrain geometry.
“The input also does not contain any handcrafted features such as foot contact states or estimated terrain geometry.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given§ Adaptive terrain curriculum · p. 9The threshold 0.2 m/s is about one third of the maximum speed of our robot.
“The threshold is a hyperparameter; 0.2 m/s is about one third of the maximum speed of our robot.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingnone given§ Indoor experiments · p. 5Such reflexes were not specified in any way during training; they developed adaptively.
“Such reflexes were not specified in any way during training: they developed adaptively.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ Further analysis of emergent behavior · p. 12The decoder network can detect foot contacts with horizontal and vertical surfaces while successfully identifying frontal collision as such.
“Additionally, the decoder network can detect foot contacts with horizontal and vertical surfaces while successfully identifying frontal collision as such, as indicated by the estimated terrain normal vector (i + iii).”
Quote in the paperNot a substantial claimMay combine two claims.
Findingcitation§ 3. DISCUSSION · p. 6The presented controller only exhibits the trot gait, which is narrower than the range of gait patterns discovered by quadrupeds in nature.
“First, the presented controller only exhibits the trot gait. This is narrower than the range of gate patterns discovered by quadrupeds in nature [33].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [33] 33. R. M. Alexander, Principles of Animal Locomotion (Princeton University Press, 2003).
Not checked in the free run.
May combine two claims.
Findingown data§ Further analysis of emergent behavior · p. 12When the LF foot collides with the step, the estimated elevation in front of the front legs increases and its uncertainty grows.
“When the LF foot collides with the step, the estimated elevation in front of the front legs increases and its uncertainty grows (i+ii).”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ Indoor experiments · p. 5The presented controller outperforms the baseline in both stepping up and down.
“The presented controller outperforms the baseline in both stepping up and down.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ S7. Reconstruction of the privileged information in different situations · p. 16External disturbance and terrain information can be reconstructed from the TCN embedding.
“The external disturbance and terrain information can also be reconstructed from the TCN embedding.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ S7. Reconstruction of the privileged information in different situations · p. 16The estimated friction coefficient between the feet and the terrain decreases as soon as the first foot starts slipping, remains low throughout the traversal, and increases about 2 s after the robot returns to normal ground.
“The estimate decreases as soon as the first foot starts slipping (i), remains low throughout the traversal (ii) and increases about 2 s after the robot returns to normal ground (iii).”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ S6. Qualitative evaluation of the adaptive terrain curriculum · p. 16The particle filter fits the latent distribution of desirable terrains, which has a bow shape in the frequency-amplitude marginal.
“Fig. S1B-C shows that the particle filter fits the latent distribution of desirable terrains, which has a bow shape in the frequency-amplitude marginal (middle).”
Quote in the paperNot a substantial claim- The authors state this as an assumption: Within Nevaluate iterations, the performance of the policy is similar.
May combine two claims.
Findingown data§ S9. Ablation of the latent representation loss for student training · p. 17The performance is comparable in the uniform slope setting and under external disturbances.
“The performance is comparable in the uniform slope setting and under external disturbances.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ S6. Qualitative evaluation of the adaptive terrain curriculum · p. 16The curriculum initially focuses on wide and shallow steps, and then broadens the distribution to include narrower steps.
“The curriculum initially focuses on wide and shallow steps (middle panels, particularly Iter. 50-60), and then broadens the distribution to include narrower steps (rightmost panels).”
Quote in the paperNot a substantial claim- The authors state this as an assumption: Within Nevaluate iterations, the performance of the policy is similar.
May combine two claims.
Premisecitation§ 1. INTRODUCTION · p. 1Under challenging natural environments, existing published controllers manifest frequent foot slippage, loss of balance, and ultimately catastrophic failure.
“Under such conditions, existing published controllers manifest frequent foot slippage, loss of balance, and ultimately catastrophic failure.”
Quote in the paperSubstantial claim0 of 5 sources checked- cites [1] 1. F. Jenelten, J. Hwangbo, F . Tresoldi, C. D. Bellicoso, M. Hutter, Dynamic locomotion on slippery ground, IEEE Robotics and Automation Letters 4170-4176 (2019).
Not checked in the free run.
- cites [2] 2. G. Bledt, P . M. Wensing, S. Ingersoll, S. Kim, Contact model fusion for
Not checked in the free run.
- cites [3]
Not checked in the free run.
- cites [4] 4. J. Reher, W. Ma, A. D. Ames, Dynamic walking with compliance on a Cassie bipedal robot, European Control Conference , 2589-2595 (IEEE, 2019).
Not checked in the free run.
- cites [5] 5. Y. Gong, R. Hartley, X. Da, A. Hereid, O. Harib, J. Huang, J. W. Grizzle, Feedback control of a Cassie bipedal robot: Walking, standing, and riding a Segway, American Control Conference , 4559-4566 (IEEE, 2019).
Not checked in the free run.
May combine two claims.
Study design and conductnone given§ 1. INTRODUCTION · p. 3In all deployments, robots of the same generation were driven by exactly the same controller under all conditions, and no tuning was required to adapt to different environments.
“In all deployments, robots of the same generation were driven by exactly the same controller under all conditions. No tuning was required to adapt to different environments.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 1. INTRODUCTION · p. 3Realistic modeling of the physical system, including the actuators, is a practice for successful transfer of legged locomotion controllers from simulation to physical machines.
“One is realistic modeling of the physical system, including the actuators [12].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019).
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 1. INTRODUCTION · p. 3Randomization of physical parameters that vary between simulation and reality is a practice for successful transfer, making the controller robust to a range of conditions without precise a priori modeling.
“Another is randomization of physical parameters that vary between simulation and reality, such that the controller becomes robust to a range of conditions that cover those that arise in physical deployment, without the necessity to precisely model these conditions a priori [21].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [21] 21. X. B. Peng, M. Andrychowicz, W. Zaremba, P . Abbeel, Sim-to-real transfer of robotic control with dynamics randomization, IEEE International Conference on Robotics and Automation (ICRA) (IEEE, 2018). source
Not checked in the free run.
May combine two claims.
Interpretationcitation§ Indoor experiments · p. 5This level of robustness is beyond the reach of prior controllers for ANYmal robots and is comparable to the state of the art.
“This level of robustness is beyond the reach of prior controllers for ANYmal robots [1, 26] and is comparable to the state of the art [2, 29].”
Quote in the paperSubstantial claim0 of 4 sources checked- cites [1] 1. F. Jenelten, J. Hwangbo, F . Tresoldi, C. D. Bellicoso, M. Hutter, Dynamic locomotion on slippery ground, IEEE Robotics and Automation Letters 4170-4176 (2019).
Not checked in the free run.
- cites [26]
Not checked in the free run.
- cites [2] 2. G. Bledt, P . M. Wensing, S. Ingersoll, S. Kim, Contact model fusion for
Not checked in the free run.
- cites [29] 29. Ghost Robotics, Vision 60: Latest blind-mode stress testing of V60 legged robot, www.youtube.com/watch?v=tQsLauQWp8M (2019).
Not checked in the free run.
May combine two claims.
Interpretationreasoning§ 1. INTRODUCTION · p. 3The training conditions grow increasingly more challenging, yielding an omnidirectional controller that combines agility with unprecedented resilience.
“The training conditions grow increasingly more challenging, yielding an omnidirectional controller that combines agility with unprecedented resilience.”
Quote in the paperSubstantial claimMay combine two claims.
Premisereasoning§ 1. INTRODUCTION · p. 1Conventional systems for legged locomotion on rough terrain escalate in complexity as more scenarios are taken into account, are laborious to develop and maintain, and remain vulnerable to corner cases.
“Overall, conventional systems for legged locomotion on rough terrain escalate in complexity as more scenarios are taken into account, have become extremely laborious to develop and maintain, and remain vulnerable to corner cases.”
Quote in the paperSubstantial claim0 of 5 sources checked- cites [1] 1. F. Jenelten, J. Hwangbo, F . Tresoldi, C. D. Bellicoso, M. Hutter, Dynamic locomotion on slippery ground, IEEE Robotics and Automation Letters 4170-4176 (2019).
Not checked in the free run.
- cites [2] 2. G. Bledt, P . M. Wensing, S. Ingersoll, S. Kim, Contact model fusion for
Not checked in the free run.
- cites [3]
Not checked in the free run.
- cites [4] 4. J. Reher, W. Ma, A. D. Ames, Dynamic walking with compliance on a Cassie bipedal robot, European Control Conference , 2589-2595 (IEEE, 2019).
Not checked in the free run.
- cites [5] 5. Y. Gong, R. Hartley, X. Da, A. Hereid, O. Harib, J. Huang, J. W. Grizzle, Feedback control of a Cassie bipedal robot: Walking, standing, and riding a Segway, American Control Conference , 4559-4566 (IEEE, 2019).
Not checked in the free run.
May combine two claims.
Interpretationreasoning§ Natural environments · p. 5The quantitative evaluation in Table 1 understates the difference between the two controllers because it only measures speed and energetic efficiency of the baseline when it successfully locomotes, and baseline catastrophic failures are not factored into these measurements.
“The quantitative evaluation reported in Table 1 understates the difference between the two controllers because it only measures speed and energetic efficiency of the baseline when it successfully locomotes. The baseline's catastrophic failures are not factored into these measurements: when the baseline fails, it is reset by a human operator in a more stable configuration.”
Quote in the paperNot a substantial claimMay combine two claims.
Premisecitation§ 1. INTRODUCTION · p. 1Exteroceptive sensors such as cameras and LiDAR cannot reliably measure physical characteristics such as friction and compliance, are impeded by obstructions such as vegetation, snow, and water, and may not have the coverage and temporal resolution to capture changes induced by the robot itself.
“Exteroceptive sensors such as cameras and LiDAR cannot reliably measure physical characteristics such as friction and compliance, are impeded by obstructions such as vegetation, snow, and water, and may not have the coverage and temporal resolution to capture changes induced by the robot itself, such as the crumbling of loose ground under the robot's feet.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [27] 27. P. Fankhauser, M. Bloesch, C. Gehring, M. Hutter, R. Siegwart. Robotcentric elevation mapping with uncertainty estimates. Mobile Service Robotics (World Scientific, 2014), 433-440.
Not checked in the free run.
May combine two claims.
Interpretationreasoning§ Natural environments · p. 3Our controller does not rely on exteroception and is immune to failure of exteroceptive terrain estimation in snow, water, or dense vegetation.
“Our controller does not rely on exteroception and is immune to such failure.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Premisecitation§ 1. INTRODUCTION · p. 1Model-free reinforcement learning has recently emerged as an alternative approach in the development of legged locomotion skills.
“Model-free reinforcement learning (RL) has recently emerged as an alternative approach in the development of legged locomotion skills [12-14].”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019).
Not checked in the free run.
- cites [13] 13. T. Haarnoja, S. Ha, A. Zhou, J. Tan, G. Tucker, S. Levine, Learning to walk via deep reinforcement learning, Robotics: Science and Systems (2019).
Not checked in the free run.
- cites [14] 14. Z. Xie, P . Clary, J. Dao, P . Morais, J. Hurst, M. van de Panne, Learning locomotion skills for Cassie: Iterative design and sim-to-real, Conference on Robot Learning (2019).
Not checked in the free run.
Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.
Interpretationreasoning§ 3. DISCUSSION · p. 6The presented results expedite the deployment of legged machines in environments that are beyond the reach of wheeled and tracked robots and are dangerous or inaccessible to humans, while the presented methodology opens new frontiers for training complex robotic systems in simulation and deploying them in the full richness and complexity of the physical world.
“More broadly, the presented results expedite the deployment of legged machines in environments that are beyond the reach of wheeled and tracked robots and are dangerous or inaccessible to humans, while the presented methodology opens new frontiers for training complex robotic systems in simulation and deploying them in the full richness and complexity of the physical world.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 3. DISCUSSION · p. 6The presented controller relies solely on proprioception, which is a significant advantage because it makes few assumptions on the sensor suite and is not susceptible to failure when exteroception breaks down.
“Second, the presented controller relies solely on proprioception. This is a significant advantage in that the controller makes few assumptions on the sensor suite and is not susceptible to failure when exteroception breaks down.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationown data§ Further analysis of emergent behavior · p. 12The policy has direct access to the measurements at the moment of foot-trapping, and leverages this in the following swing phase.
“The policy has direct access to the measurements at the moment of foot-trapping, and leverages this in the following swing phase.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ Adaptive terrain curriculum · p. 9Florensa et al. similarly choose achievable yet difficult goals for RL agents.
“Florensa et al. [39] similarly choose achievable yet difficult goals for RL agents.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [39] 39. C. Florensa, D. Held, X. Geng, P . Abbeel, Automatic goal generation for reinforcement learning agents, International Conference on Machine Learning , 1514-1523 (2018).
Not checked in the free run.
May combine two claims.
Interpretationreasoning§ 3. DISCUSSION · p. 6Blind locomotion is inherently limited: if the machine is commanded to walk off a cliff, it will; even in less extreme conditions, the robot's gait is fairly conservative because it must feel out the environment with its body as it locomotes.
“Nevertheless, blind locomotion is inherently limited. If the machine is commanded to walk off a cliff, it will. Even in less extreme conditions, the robot's gait is fairly conservative since it must by necessity feel out the environment with its body as it locomotes.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ Motion synthesis · p. 8Previous works commonly leveraged predefined foot contact schedules.
“Previous works commonly leveraged predefined foot contact schedules [2, 26, 35].”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [2] 2. G. Bledt, P . M. Wensing, S. Ingersoll, S. Kim, Contact model fusion for
Not checked in the free run.
- cites [26]
Not checked in the free run.
- cites [35] 35. V. Barasuol, J. Buchli, C. Semini, M. Frigerio, E. R. De Pieri, D. G. Caldwell, A reactive controller framework for quadrupedal locomotion on challenging terrain, 2013 IEEE International Conference on Robotics and Automation , 2554-2561 (IEEE, 2013).
Not checked in the free run.
May combine two claims.
Interpretationreasoning§ A. Overview · p. 6In our formulation, only the direction of the target velocity is given to the controller, unlike many existing works that focus on tracking the target velocity of the base, because the feasible range of target speeds is often unclear on challenging terrain.
“In our formulation, unlike many existing works [12, 14, 16] that focus on tracking the target velocity of the base ( B IB vT ), only the direction ( B IB ˆ vT ) is given to the controller. The reason is that the feasible range of target speeds is often unclear on challenging terrain.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019).
Not checked in the free run.
- cites [14] 14. Z. Xie, P . Clary, J. Dao, P . Morais, J. Hurst, M. van de Panne, Learning locomotion skills for Cassie: Iterative design and sim-to-real, Conference on Robot Learning (2019).
Not checked in the free run.
- cites [16] 16. J. Tan, T. Zhang, E. Coumans, A. Iscen, Y. Bai, D. Hafner, S. Bohez, V. Vanhoucke, Sim-to-real: Learning agile locomotion for quadruped robots, Robotics: Science and Systems (2018).
Not checked in the free run.
May combine two claims.
Premisecitation§ Motion synthesis · p. 8Defining the target foot positions in the horizontal frames of the feet reduces the effect of base attitude on the foot motions and consequently stabilizes training.
“This kinematic trick reduces the effect of base attitude on the foot motions [35] and consequently stabilizes training.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [35] 35. V. Barasuol, J. Buchli, C. Semini, M. Frigerio, E. R. De Pieri, D. G. Caldwell, A reactive controller framework for quadrupedal locomotion on challenging terrain, 2013 IEEE International Conference on Robotics and Automation , 2554-2561 (IEEE, 2013).
Not checked in the free run.
Classified as probably not a substantial claim.
Premisecitation§ S2. Implementation details · p. 15Randomization of dynamic properties improves the robustness of the policy.
“As several studies have shown that randomization of dynamic properties improves the robustness of the policy [12, 16]”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [12] 12. J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots, Science Robotics p. eaau5872 (2019).
Not checked in the free run.
- cites [16] 16. J. Tan, T. Zhang, E. Coumans, A. Iscen, Y. Bai, D. Hafner, S. Bohez, V. Vanhoucke, Sim-to-real: Learning agile locomotion for quadruped robots, Robotics: Science and Systems (2018).
Not checked in the free run.
May combine two claims.
Premisecitation§ Motion synthesis · p. 8Using analytic inverse kinematics maximizes computational efficiency and allows reuse of existing position control actuator models for the sim-to-real transfer.
“The main reason for using analytic IK is to maximize computational efficiency and to reuse existing position control actuator models [6, 15] for the sim-to-real transfer.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [6] 6. J. Hwangbo, C. D. Bellicoso, P . Fankhauser, M. Huttery, Probabilistic foot contact estimation by fusing information from dynamics and differential/forward kinematics, 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 3872-3878 (IEEE, 2016).
Not checked in the free run.
- cites [15] 15. J. Lee, J. Hwangbo, M. Hutter, Robust recovery controller for a quadrupedal robot using deep reinforcement learning, arXiv:1901.07517 (2019).
Not checked in the free run.
Probably not a substantial claim: may not pass S2 (checkable).
Interpretationreasoning§ 3. DISCUSSION · p. 6The presented results substantially advance the published state of the art in legged robotics.
“The presented results substantially advance the published state of the art in legged robotics.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ S4. Reward function for teacher policy training · p. 16The maximum speed reachable on flat terrain with the existing controller is 0.6 m/s.
“The velocity threshold is defined as 0.6 m/s which is the maximum speed reachable on the flat terrain with the existing controller [26].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [26]
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisecitation§ Adaptive terrain curriculum · p. 9The POET method employs minimal criteria and aims to choose environmental parameters that are neither too challenging nor trivial for the agents, realized by selecting task parameters that yield mid-range rewards.
“The method employs minimal criteria (MC) [24, 40] and aims to choose environmental parameters that are neither too challenging nor trivial for the agents: this is realized by selecting task parameters that yield mid-range rewards.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [24] 24. J. C. Brant, K. O. Stanley, Minimal criterion coevolution: a new approach to open-ended search, Genetic and Evolutionary Computation Conference , 67-74 (2017).
Not checked in the free run.
- cites [40] 40. J. Lehman, K. O. Stanley, Revising the evolutionary computation abstraction: minimal criteria novelty search, Genetic and Evolutionary Computation Conference , 103-110 (2010).
Not checked in the free run.
May combine two claims.
Premisecitation§ Student policy · p. 9The TCN architecture affords transparent control over the input history length, can accommodate long histories, and is known to be robust to hyperparameter settings.
“We use the TCN architecture because it affords transparent control over the input history length, can accommodate long histories, and is known to be robust to hyperparameter settings [22].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [22] 22. S. Bai, J. Z. Kolter, V. Koltun, An empirical evaluation of generic convolutional and recurrent networks for sequence modeling, arXiv:1803.01271 (2018).
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Premisenone given§ Teacher policy · p. 8Using <cos(f), sin(f)> instead of f to encode the leg phase is a smooth and unique representation for the angle.
“To encode the leg phase, we use 〈 cos ( f ) , sin ( f ) 〉 instead of f , which is a smooth and unique representation for the angle.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ Adaptive terrain curriculum · p. 9The POET approach generates diverse parameterized terrains for a 2D bipedal agent.
“The paired open-ended trailblazer (POET) approach [25] generates diverse parameterized terrains for a 2D bipedal agent.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [25] 25. R. Wang, J. Lehman, J. Clune, K. O. Stanley, Paired open-ended trailblazer (poet): Endlessly generating increasingly complex and diverse learning environments and their solutions, arXiv:1901.01753 (2019).
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone given§ S3. Foot trajectory generator · p. 15With a set of reasonably tuned f0, h and fi,0, a quadruped can stably step in place.
“With a set of reasonably tuned f 0 , h and f i ,0 , a quadruped can stably step in place.”
Quote in the paperNot a substantial claim
The paper relies on this without support.
Legged robots have the potential to traverse any terrain that their animal counterparts can.
“Legged robots have the potential to traverse any terrain that their animal counterparts can.”
The paper relies on this without support.
Joint encoders and an inertial measurement unit are the most durable and reliable sensors on legged machines.
“The controller uses only proprioceptive measurements from joint encoders and an inertial measurement unit (IMU), which are the most durable and reliable sensors on legged machines.”
The authors state this as an assumption.
The environment is fully observable to the teacher.
“Assuming the environment is fully observable to the teacher, we formulate locomotion control as an MDP and use an off-the-shelf RL method [36] to solve it.”
The authors state this as an assumption.
Terrain parameters with similar traversability are close in Euclidean distance in parameter space.
“1. Terrain parameters with similar Tr ( · , p ) are close in Euclidean distance in parameter space.”
The authors state this as an assumption.
A policy trained over the terrains generated by cTs in some area of C will learn to interpolate to nearby terrain parameters.
“2. A policy trained over the terrains generated by c T s in some area of C will learn to interpolate to nearby terrain parameters.”
The authors state this as an assumption.
The evolution of each terrain parameter is a Markov process (only relies on the current value and randomly sampled noise).
“It satisfies the third assumption (Markov process) because the evolution of each parameter only relies on the current value and randomly sampled noise.”
The authors state this as an assumption.
Within Nevaluate iterations, the performance of the policy is similar.
“We assume that within Nevaluate iterations, the performance of the policy is similar.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
This opens interesting research opportunities for the adaptation of lab-generated simulation to real-world conditions, similar to what has been observed in image synthesis for self driving and robotics.
“This opens interesting research opportunities for the adaptation of lab-generated simulation to real-world conditions, similar to what has been observed in the field of image synthesis for self driving143 and robotics144 in the last decade.”
From Perspectives in machine learning for wildlife conservation (2022) · cites this paper as [144]
Cites
Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling
Chung, J., Gulcehre, C., Cho, K., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1412.3555 · Cited by 11,008 · see citing works
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
Simonyan, K., Vedaldi, A., & Zisserman, A. (2013) · arXiv · doi:10.48550/arxiv.1312.6034 · Cited by 5,370 · see citing works
An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
Bai, S., Kolter, J. Z., & Koltun, V. (2018) · arXiv · doi:10.48550/arxiv.1803.01271 · Cited by 4,655 · see citing works
Self-citationEfficient Bipedal Robots Based on Passive-Dynamic Walkers
Collins, S. H., Ruina, A. L., Tedrake, R., & Wisse, M. (2005) · Science · doi:10.1126/science.1107799 · Cited by 1,878 · see citing works
Learning agile and dynamic motor skills for legged robots
Hwangbo, J., Lee, J., Dosovitskiy, A., Bellicoso, D., Tsounis, V., Koltun, V., & Hutter, M. (2019) · Science Robotics · doi:10.1126/scirobotics.aau5872 · Cited by 1,482 · see citing works
Self-citationSim-to-Real Transfer of Robotic Control with Dynamics Randomization
Peng, X. B., Andrychowicz, M., Zaremba, W., & Abbeel, P. (2018) · 2018 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra.2018.8460528 · Cited by 1,309 · cited 1× · see citing works
“Another is randomization of physical parameters that vary between simulation and reality, such that the controller becomes robust to a range of conditions that cover those that arise in physical deployment, without the necessity to precisely model these conditions a priori [21].”
Principles of Animal Locomotion
Alexander, R. M. (2002) · doi:10.1515/9781400849512 · Cited by 1,151 · see citing works
ANYmal - a highly mobile and dynamic quadrupedal robot
Hutter, M., Gehring, C., Jud, D., Lauber, A., Bellicoso, C. D., Tsounis, V., Hwangbo, J., Bodie, K., Fankhauser, P., Bloesch, M., et al. (2016) · 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2016.7758092 · Cited by 969 · see citing works
Self-citationSim-to-Real: Learning Agile Locomotion For Quadruped Robots
Tan, J., Zhang, T., Coumans, E., Iscen, A., Bai, Y., Hafner, D., Bohez, S., & Vanhoucke, V. (2018) · Robotics: Science and Systems XIV · doi:10.15607/rss.2018.xiv.010 · Cited by 735 · see citing works
Learning Agile Robotic Locomotion Skills by Imitating AnimalsPreprint
Peng, X.-B., Coumans, E., Zhang, T.-N., Lee, T., Tan, J., & Levine, S. (2020) · Robotics: Science and Systems Conference · doi:10.48550/arxiv.2004.00784 · Cited by 692 · see citing works
Solving Rubik's Cube with a Robot Hand
OpenAI, Akkaya, I., Andrychowicz, M., Chociej, M., Litwin, M., McGrew, B., Petron, A., Paino, A., Plappert, M., Powell, G., et al. (2019) · arXiv · doi:10.48550/arxiv.1910.07113 · Cited by 602 · see citing works
Learning to Walk Via Deep Reinforcement Learning
Haarnoja, T., Ha, S., Zhou, A., Tan, J., Tucker, G., & Levine, S. (2019) · Robotics: Science and Systems XV · doi:10.15607/rss.2019.xv.011 · Cited by 430 · see citing works
Learning Agile Robotic Locomotion Skills by Imitating Animals
Bin Peng, X., Coumans, E., Zhang, T., Lee, T.-W., Tan, J., & Levine, S. (2020) · Robotics: Science and Systems XVI · doi:10.15607/rss.2020.xvi.064 · Cited by 357 · see citing works
Teacher–Student Curriculum Learning
Matiisen, T., Oliver, A., Cohen, T., & Schulman, J. (2019) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2019.2934906 · Cited by 351 · see citing works
State Estimation for Legged Robots: Consistent Fusion of Leg Kinematics and IMU
Bloesch, M., Hutter, M., Hoepflinger, M. A., Leutenegger, S., Gehring, C., Remy, C. D., & Siegwart, R. (2013) · The MIT Press eBooks · doi:10.7551/mitpress/9816.003.0008 · Cited by 277 · see citing works
Self-citationPer-Contact Iteration Method for Solving Contact Dynamics
Hwangbo, J., Lee, J., & Hutter, M. (2018) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2018.2792536 · Cited by 243 · see citing works
Self-citationFeedback Control of a Cassie Bipedal Robot: Walking, Standing, and Riding a Segway
Gong, Y., Hartley, R., Da, X., Hereid, A., Harib, O., Huang, J.-K., & Grizzle, J. (2019) · 2019 American Control Conference (ACC) · doi:10.23919/acc.2019.8814833 · Cited by 221 · cited 1× · see citing works
“Many rely on elaborate state machines that coordinate the execution of motion primitives and reflex controllers [1–5].”
Learning symmetric and low-energy locomotion
Yu, W., Turk, G., & Liu, C. K. (2018) · ACM Transactions on Graphics · doi:10.1145/3197517.3201397 · Cited by 208 · see citing works
Dynamic Locomotion Through Online Nonlinear Motion Optimization for Quadrupedal Robots
Bellicoso, C. D., Jenelten, F., Gehring, C., & Hutter, M. (2018) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2018.2794620 · Cited by 206 · see citing works
Self-citationContact Model Fusion for Event-Based Locomotion in Unstructured Terrains
Bledt, G., Wensing, P. M., Ingersoll, S., & Kim, S. (2018) · 2018 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra.2018.8460904 · Cited by 197 · see citing works
ROBOT-CENTRIC ELEVATION MAPPING WITH UNCERTAINTY ESTIMATES
FANKHAUSER, P., BLOESCH, M., GEHRING, C., HUTTER, M., & SIEGWART, R. (2014) · Mobile Service Robotics · doi:10.1142/9789814623353_0051 · Cited by 192 · see citing works
Self-citationRevising the evolutionary computation abstraction
Lehman, J., & Stanley, K. O. (2010) · Genetic and Evolutionary Computation Conference (GECCO) · doi:10.1145/1830483.1830503 · Cited by 182 · see citing works
Wang, R., Lehman, J., Clune, J., & Stanley, K. O. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1901.01753 · Cited by 166 · see citing works
A reactive controller framework for quadrupedal locomotion on challenging terrain
Barasuol, V., Buchli, J., Semini, C., Frigerio, M., Pieri, E. R. D., & Caldwell, D. G. (2013) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2013.6630926 · Cited by 164 · see citing works
State estimation for legged robots on unstable and slippery terrain
Bloesch, M., Gehring, C., Fankhauser, P., Hutter, M., Hoepflinger, M. A., & Siegwart, R. (2013) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2013.6697236 · Cited by 136 · see citing works
Self-citationLearning to Walk in the Real World with Minimal Human EffortPreprint
Ha, S., Xu, P., Tan, Z., Levine, S., & Tan, J. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2002.08550 · Cited by 126 · see citing works
Probabilistic Contact Estimation and Impact Detection for State Estimation of Quadruped Robots
Camurri, M., Fallon, M. F., Bazeille, S., Radulescu, A., Barasuol, V., Caldwell, D. G., & Semini, C. (2017) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2017.2652491 · Cited by 120 · see citing works
Learning Ground Traversability From Simulations
Chavéz-García, R. O., Guzzi, J., Gambardella, L. M., & Giusti, A. (2018) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2018.2801794 · Cited by 107 · see citing works
Dynamic Locomotion on Slippery Ground
Jenelten, F., Hwangbo, J., Tresoldi, F., Bellicoso, C. D., & Hutter, M. (2019) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2019.2931284 · Cited by 84 · see citing works
Self-citationDynamic Walking with Compliance on a Cassie Bipedal Robot
Reher, J., Ma, W.-L., & Ames, A. D. (2019) · 2019 18th European Control Conference (ECC) · doi:10.23919/ecc.2019.8796090 · Cited by 81 · see citing works
Dynamic trotting on slopes for quadrupedal robots
Gehring, C., Bellicoso, C. D., Coros, S., Bloesch, M., Fankhauser, P., Hutter, M., & Siegwart, R. (2015) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2015.7354099 · Cited by 70 · see citing works
Self-citationHwangbo, J., Bellicoso, C. D., Fankhauser, P., & Hutter, M. (2016) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2016.7759570 · Cited by 66 · see citing works
Self-citationLegged Robot State-Estimation Through Combined Forward Kinematic and Preintegrated Contact Factors
Hartley, R., Mangelson, J., Gan, L., Ghaffari, M., Walls, J. M., Eustice, R. M., & Grizzle, J. W. (2018) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2018.8460748 · Cited by 54 · see citing works
Brant, J. C., & Stanley, K. O. (2017) · Genetic and Evolutionary Computation Conference (GECCO) · doi:10.1145/3071178.3071186 · Cited by 47 · see citing works
Slip Detection and Recovery for Quadruped Robots
Focchi, M., Barasuol, V., Frigerio, M., Caldwell, D. G., & Semini, C. (2017) · Springer proceedings in advanced robotics · doi:10.1007/978-3-319-60916-4_11 · Cited by 42 · see citing works
Focchi, M., Orsolino, R., Camurri, M., Barasuol, V., Mastalli, C., Caldwell, D. G., & Semini, C. (2020) · Springer Tracts in Advanced Robotics · doi:10.1007/978-3-030-22327-4_9 · Cited by 28 · see citing works
Robust Recovery Controller for a Quadrupedal Robot using Deep Reinforcement LearningPreprint
Lee, J., Hwangbo, J., & Hutter, M. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1901.07517 · Cited by 26 · see citing works
Self-citation
Flagged first, then by citations
Research works printed but not linked: 8
- Z. Xie P. Clary J. Dao P. Morais J. Hurst M. van de Panne Learning locomotion skills for Cassie: Iterative design and sim-to-real in Conference on Robot Learning (PMLR 2019).
- Y. Yang K. Caluwaerts A. Iscen T. Zhang J. Tan V. Sindhwani Data efficient reinforcement learning for legged robots in Conference on Robot Learning (PMLR 2019).
- D. Chen B. Zhou V. Koltun P. Krähenbühl Learning by cheating in Conference on Robot Learning (2019).
- A. Iscen K. Caluwaerts J. Tan T. Zhang E. Coumans V. Sindhwani V. Vanhoucke Policies modulating trajectory generators in Conference on Robot Learning (PMLR 2018) pp. 916–926.
- J. Schulman S. Levine P. Abbeel M. Jordan P. Moritz Trust region policy optimization in International Conference on Machine Learning (PMLR 2015) pp. 1889–1897.
- S. Ross G. Gordon D. Bagnell A reduction of imitation learning and structured prediction to no-regret online learning in International Conference on Artificial Intelligence and Statistics (AISTATS 2011) pp. 627–635.
- C. Florensa D. Held X. Geng P. Abbeel Automatic goal generation for reinforcement learning agents in International Conference on Machine Learning (PMLR 2018) pp. 1514–1523.
- A. Kendall Y. Gal What uncertainties do we need in bayesian deep learning for computer vision? in Advances in Neural Information Processing Systems 30 (Neural Information Processing Systems Foundation 2017) pp. 5574–5584.
Cited by
Sorted by citations
Learning robust perceptive locomotion for quadrupedal robots in the wild
Miki, T., Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., & Hutter, M. (2022) · Science Robotics · doi:10.1126/scirobotics.abk2822 · Cited by 847 · see citing works
Self-citationPerspectives in machine learning for wildlife conservation
Tuia, D., Kellenberger, B., Beery, S., Costelloe, B. R., Zuffi, S., Risse, B., Mathis, A., Mathis, M. W., van Langevelde, F., Burghardt, T., et al. (2022) · Nature Communications · doi:10.1038/s41467-022-27980-y · Cited by 740 · see citing works
How to train your robot with deep reinforcement learning: Lessons we have learned
Ibarz, J., Tan, J., Finn, C., Kalakrishnan, M., Pastor, P., & Levine, S. (2021) · The International Journal of Robotics Research · doi:10.1177/0278364920987859 · Cited by 570 · see citing works
Neural-Fly enables rapid learning for agile flight in strong winds
O’Connell, M., Shi, G., Shi, X., Azizzadenesheli, K., Anandkumar, A., Yue, Y., & Chung, S. (2022) · Science Robotics · doi:10.1126/scirobotics.abm6597 · Cited by 297 · see citing works
Perceptive Locomotion Through Nonlinear Model-Predictive Control
Grandia, R., Jenelten, F., Yang, S., Farshidian, F., & Hutter, M. (2023) · IEEE Transactions on Robotics · doi:10.1109/tro.2023.3275384 · Cited by 281 · see citing works
Self-citationReal-world humanoid locomotion with reinforcement learning
Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., & Sreenath, K. (2024) · Science Robotics · doi:10.1126/scirobotics.adi9579 · Cited by 252 · see citing works
Reaching the limit in autonomous racing: Optimal control versus reinforcement learning
Song, Y., Romero, Á., Müller, M., Koltun, V., & Scaramuzza, D. (2023) · Science Robotics · doi:10.1126/scirobotics.adg1462 · Cited by 241 · see citing works
Self-citationReinforcement Learning for Robust Parameterized Locomotion Control of Bipedal Robots
Li, Z., Cheng, X., Peng, X. B., Abbeel, P., Levine, S., Berseth, G., & Sreenath, K. (2021) · 2021 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra48506.2021.9560769 · Cited by 219 · see citing works
Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes
Tang, C., Abbatematteo, B., Hu, J., Chandra, R., Martín-Martín, R., & Stone, P. (2024) · Annual Review of Control Robotics and Autonomous Systems · doi:10.1146/annurev-control-030323-022510 · Cited by 217 · see citing works
Optimization-Based Control for Dynamic Legged Robots
Wensing, P. M., Posa, M., Hu, Y. H., Escande, A., Mansard, N., & Prete, A. D. (2023) · IEEE Transactions on Robotics · doi:10.1109/tro.2023.3324580 · Cited by 216 · see citing works
Autonomous Drone Racing with Deep Reinforcement Learning
Song, Y., Steinweg, M., Kaufmann, E., & Scaramuzza, D. (2021) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros51168.2021.9636053 · Cited by 213 · see citing works
Ji, G., Mun, J., Kim, H., & Hwangbo, J. (2022) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2022.3151396 · Cited by 208 · see citing works
Self-citationBlind Bipedal Stair Traversal via Sim-to-Real Reinforcement Learning
Siekmann, J., Green, K., Warila, J., Fern, A., & Hurst, J. (2021) · Robotics: Science and Systems XVII · doi:10.15607/rss.2021.xvii.061 · Cited by 193 · see citing works
Learning agile soccer skills for a bipedal robot with deep reinforcement learning
Haarnoja, T., Moran, B., Lever, G., Huang, S. H., Tirumala, D., Humplik, J., Wulfmeier, M., Tunyasuvunakool, S., Siegel, N., Hafner, R., et al. (2024) · Science Robotics · doi:10.1126/scirobotics.adi8022 · Cited by 191 · see citing works
Learning quadrupedal locomotion on deformable terrain
Choi, S., Ji, G., Park, J., Kim, H., Mun, J., Lee, J. H., & Hwangbo, J. (2023) · Science Robotics · doi:10.1126/scirobotics.ade2256 · Cited by 184 · see citing works
Self-citationExtreme Parkour with Legged Robots
Cheng, X., Shi, K., Agarwal, A., & Pathak, D. (2024) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra57147.2024.10610200 · Cited by 183 · see citing works
Nahrendra, I. M. A., Yu, B., & Myung, H. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10161144 · Cited by 177 · see citing works
Experiment-free exoskeleton assistance via learning in simulation
Luo, S., Jiang, M., Zhang, S., Zhu, J., Yu, S., Silva, I. D., Wang, T., Rouse, E. J., Zhou, B., Yuk, H., et al. (2024) · Nature · doi:10.1038/s41586-024-07382-4 · Cited by 175 · see citing works
Game changers in science and technology - now and beyond
Betz, U. A. K., Arora, L., Assal, R. A., de Azevedo, H. F. Z., Baldwin, J., Becker, M. S., Bostock, S., Cheng, V. W. T., Egle, T., Ferrari, N., et al. (2023) · Technological Forecasting and Social Change · doi:10.1016/j.techfore.2023.122588 · Cited by 146 · see citing works
RLOC: Terrain-Aware Legged Locomotion Using Reinforcement Learning and Optimal Control
Gangapurwala, S., Geisert, M., Orsolino, R., Fallon, M. F., & Havoutis, I. (2022) · IEEE Transactions on Robotics · doi:10.1109/tro.2022.3172469 · Cited by 145 · see citing works
Learning robust autonomous navigation and locomotion for wheeled-legged robots
Lee, J., Bjelonic, M., Reske, A., Wellhausen, L., Miki, T., & Hutter, M. (2024) · Science Robotics · doi:10.1126/scirobotics.adi9641 · Cited by 143 · see citing works
Self-citationFoundation models in smart agriculture: Basics, opportunities, and challenges
Li, J., Xu, M., Xiang, L., Chen, D., Zhuang, W., Yin, X., & Li, Z. (2024) · Computers and Electronics in Agriculture · doi:10.1016/j.compag.2024.109032 · Cited by 141 · see citing works
Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control
Li, Z., Peng, X. B., Abbeel, P., Levine, S., Berseth, G., & Sreenath, K. (2024) · The International Journal of Robotics Research · doi:10.1177/02783649241285161 · Cited by 137 · see citing works
Rapid locomotion via reinforcement learning
Margolis, G. B., Yang, G., Paigwar, K., Chen, T., & Agrawal, P. (2024) · The International Journal of Robotics Research · doi:10.1177/02783649231224053 · Cited by 136 · see citing works
Guastella, D. C., & Muscato, G. (2020) · Sensors · doi:10.3390/s21010073 · Cited by 133 · see citing works
Deep reinforcement learning with shallow controllers: An experimental application to PID tuning
Lawrence, N. P., Forbes, M. G., Loewen, P. D., McClement, D. G., Backström, J. U., & Gopaluni, R. B. (2022) · Control Engineering Practice · doi:10.1016/j.conengprac.2021.105046 · Cited by 133 · see citing works
Machine intelligence for chemical reaction space
Schwaller, P., Vaucher, A. C., Laplaza, R., Bunne, C., Krause, A., Corminbœuf, C., & Laino, T. (2022) · Wiley Interdisciplinary Reviews Computational Molecular Science · doi:10.1002/wcms.1604 · Cited by 131 · see citing works
Rapid Locomotion via Reinforcement Learning
Margolis, G. B., Yang, G., Paigwar, K., Chen, T., & Agrawal, P. (2022) · Robotics: Science and Systems · doi:10.15607/rss.2022.xviii.022 · Cited by 131 · see citing works
Super-Human Performance in Gran Turismo Sport Using Deep Reinforcement Learning
Fuchs, F., Song, Y., Kaufmann, E., Scaramuzza, D., & Durr, P. (2021) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2021.3064284 · Cited by 130 · see citing works
CPG-RL: Learning Central Pattern Generators for Quadruped Locomotion
Bellegarda, G., & Ijspeert, A. (2022) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2022.3218167 · Cited by 125 · see citing works
Scientific exploration of challenging planetary analog environments with a team of legged robots
Arm, P., Waibel, G., Preisig, J., Tuna, T., Zhou, R., Bickel, V., Ligeza, G., Miki, T., Kehl, F., Kolvenbach, H., et al. (2023) · Science Robotics · doi:10.1126/scirobotics.ade9548 · Cited by 124 · see citing works
Self-citationNeuromorphic computing hardware and neural architectures for robotics
Sandamirskaya, Y., Kaboli, M., Conradt, J., & Celikel, T. (2022) · Science Robotics · doi:10.1126/scirobotics.abl8419 · Cited by 122 · see citing works
Policy Search for Model Predictive Control With Application to Agile Drone Flight
Song, Y., & Scaramuzza, D. (2022) · IEEE Transactions on Robotics · doi:10.1109/tro.2022.3141602 · Cited by 121 · see citing works
Farley, A., Wang, J., & Marshall, J. A. (2022) · Simulation Modelling Practice and Theory · doi:10.1016/j.simpat.2022.102629 · Cited by 121 · see citing works
Real-world embodied AI through a morphologically adaptive quadruped robot
Nygaard, T. F., Martin, C. P., Tørresen, J., Glette, K., & Howard, D. (2021) · Nature Machine Intelligence · doi:10.1038/s42256-021-00320-3 · Cited by 120 · see citing works
Jenelten, F., He, J., Farshidian, F., & Hutter, M. (2024) · Science Robotics · doi:10.1126/scirobotics.adh5401 · Cited by 119 · see citing works
Self-citationTranzatto, M., Mascarich, F., Bernreiter, L., Godinho, C., Camurri, M., Khattak, S., Dang, T., Reijgwart, V., Loeje, J., Wisth, D., et al. (2022) · Field Robotics · doi:10.55417/fr.2022011 · Cited by 118 · see citing works
Self-citationLyapunov-stable neural-network control
Dai, H., Landry, B., Yang, L., Pavone, M., & Tedrake, R. (2021) · Robotics: Science and Systems · doi:10.15607/rss.2021.xvii.063 · Cited by 117 · see citing works
Learning to drive from a world on rails
Chen, D., Koltun, V., & Krähenbühl, P. (2021) · IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.01530 · Cited by 114 · see citing works
Self-citationBiConMP: A Nonlinear Model Predictive Control Framework for Whole Body Motion Planning
Meduri, A., Shah, P., Viereck, J., Khadiv, M., Havoutis, I., & Righetti, L. (2023) · IEEE Transactions on Robotics · doi:10.1109/tro.2022.3228390 · Cited by 107 · see citing works
Learning Robust and Agile Legged Locomotion Using Adversarial Motion Priors
Wu, J., Xin, G., Qi, C., & Xue, Y. (2023) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2023.3290509 · Cited by 107 · see citing works
Fast Contact-Implicit Model Predictive Control
Cleac’h, S. L., Howell, T. A., Yang, S., Lee, C.-Y., Zhang, J. Z., Bishop, A. L., Schwager, M., & Manchester, Z. R. (2024) · IEEE Transactions on Robotics · doi:10.1109/tro.2024.3351554 · Cited by 104 · see citing works
DeXtreme: Transfer of Agile In-hand Manipulation from Simulation to Reality
Handa, A., Allshire, A., Makoviychuk, V., Petrenko, A., Singh, R., Liu, J., Makoviichuk, D., Wyk, K. V., Zhurkevich, A., Sundaralingam, B., et al. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10160216 · Cited by 103 · see citing works
Self-reconfigurable multilegged robot swarms collectively accomplish challenging terradynamic tasks
Ozkan-Aydin, Y., & Goldman, D. I. (2021) · Science Robotics · doi:10.1126/scirobotics.abf1628 · Cited by 102 · see citing works
Visual dexterity: In-hand reorientation of novel and complex object shapes
Chen, T., Tippur, M. H., Wu, S., Kumar, V., Adelson, E. H., & Agrawal, P. (2023) · Science Robotics · doi:10.1126/scirobotics.adc9244 · Cited by 101 · see citing works
Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World
Smith, L., Kew, J. C., Peng, X. B., Ha, S., Tan, J., & Levine, S. (2022) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra46639.2022.9812166 · Cited by 100 · see citing works
Adversarial Motion Priors Make Good Substitutes for Complex Reward Functions
Escontrela, A., Peng, X. B., Yu, W., Zhang, T., Işçen, A., Goldberg, K. Y., & Abbeel, P. (2022) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros47612.2022.9981973 · Cited by 100 · see citing works
Embedded shape morphing for morphologically adaptive robots
Sun, J., Lerner, E., Tighe, B., Middlemist, C., & Zhao, J. (2023) · Nature Communications · doi:10.1038/s41467-023-41708-6 · Cited by 99 · see citing works
Automated Reinforcement Learning (AutoRL): A Survey and Open Problems
Parker-Holder, J., Rajan, R., Song, X., Biedenkapp, A., Miao, Y., Eimer, T., Zhang, B., Nguyen, V., Calandra, R., Faust, A., et al. (2022) · Journal of Artificial Intelligence Research · doi:10.1613/jair.1.13596 · Cited by 98 · see citing works
Dynamic robotic tracking of underwater targets using reinforcement learning
Masmitja, I., Martin, M., O’Reilly, T., Kieft, B., Palomeras, N., Navarro, J., & Katija, K. (2023) · Science Robotics · doi:10.1126/scirobotics.ade7811 · Cited by 96 · see citing works
Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning
Luo, J., Xu, C., Wu, J., & Levine, S. (2025) · Science Robotics · doi:10.1126/scirobotics.ads5033 · Cited by 94 · see citing works
The neuromechanics of animal locomotion: From biology to robotics and back
Ramdya, P., & Ijspeert, A. J. (2023) · Science Robotics · doi:10.1126/scirobotics.adg0279 · Cited by 92 · see citing works
Advanced Skills by Learning Locomotion and Local Navigation End-to-End
Rudin, N., Hoeller, D., Bjelonic, M., & Hutter, M. (2022) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros47612.2022.9981198 · Cited by 92 · see citing works
Self-citationA Survey of Sim-to-Real Transfer Techniques Applied to Reinforcement Learning for Bioinspired Robots
Zhu, W. D., Guo, X., Owaki, D., Kutsuzawa, K., & Hayashibe, M. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2021.3112718 · Cited by 92 · see citing works
Development of a Small-Sized Quadruped Robotic Rat Capable of Multimodal Motions
Shi, Q., Gao, J., Wang, S., Quan, X., Jia, G., Huang, Q., & Fukuda, T. (2022) · IEEE Transactions on Robotics · doi:10.1109/tro.2022.3159188 · Cited by 89 · see citing works
From motor control to team play in simulated humanoid football
Liu, S., Lever, G., Wang, Z., Merel, J., Eslami, S. M. A., Hennes, D., Czarnecki, W. M., Tassa, Y., Omidshafiei, S., Abdolmaleki, A., et al. (2022) · Science Robotics · doi:10.1126/scirobotics.abo0235 · Cited by 88 · see citing works
Advanced Skills through Multiple Adversarial Motion Priors in Reinforcement Learning
Vollenweider, E., Bjelonic, M., Klemm, V., Rudin, N., Lee, J., & Hutter, M. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10160751 · Cited by 88 · see citing works
Self-citationMa, Y., Farshidian, F., Miki, T., Lee, J., & Hutter, M. (2022) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2022.3143567 · Cited by 87 · see citing works
Self-citationRobust quadruped jumping via deep reinforcement learning
Bellegarda, G., Nguyen, C., & Nguyen, Q. (2024) · Robotics and Autonomous Systems · doi:10.1016/j.robot.2024.104799 · Cited by 87 · see citing works
A survey of the development of biomimetic intelligence and robotics
Wang, J., Chen, W., Xiao, X., Xu, Y., Li, C., Jia, X., & Meng, M. Q. ‐. (2021) · Biomimetic Intelligence and Robotics · doi:10.1016/j.birob.2021.100001 · Cited by 84 · see citing works
Learning Free Gait Transition for Quadruped Robots Via Phase-Guided Controller
Shao, Y., Jin, Y., Liu, X., He, W., Wang, H., & Yang, W. (2021) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2021.3136645 · Cited by 82 · see citing works
Autonomous Overtaking in Gran Turismo Sport Using Curriculum Reinforcement Learning
Song, Y., Lin, H., Kaufmann, E., Dürr, P., & Scaramuzza, D. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9561049 · Cited by 81 · see citing works
Parallel Learning: Overview and Perspective for Computational Learning Across Syn2Real and Sim2Real
Miao, Q., Lv, Y., Huang, M., Wang, X., & Wang, F. (2023) · IEEE/CAA Journal of Automatica Sinica · doi:10.1109/jas.2023.123375 · Cited by 80 · see citing works
A study on quadruped mobile robots
Taheri, H., & Mozayani, N. (2023) · Mechanism and Machine Theory · doi:10.1016/j.mechmachtheory.2023.105448 · Cited by 80 · see citing works
Versatile multicontact planning and control for legged loco-manipulation
Sleiman, J.-P., Farshidian, F., & Hutter, M. (2023) · Science Robotics · doi:10.1126/scirobotics.adg5014 · Cited by 80 · see citing works
Self-citationDynamics Randomization Revisited: A Case Study for Quadrupedal Locomotion
Xie, Z., Da, X., van de Panne, M., Babich, B., & Garg, A. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9560837 · Cited by 78 · see citing works
What is an artificial muscle? A comparison of soft actuators to biological muscles
Higueras-Ruiz, D. R., Nishikawa, K. C., Feigenbaum, H. P., & Shafer, M. W. (2021) · Bioinspiration & Biomimetics · doi:10.1088/1748-3190/ac3adf · Cited by 78 · see citing works
DayDreamer: World Models for Physical Robot Learning
Wu, P., Escontrela, A., Hafner, D., Goldberg, K., & Abbeel, P. (2022) · arXiv · doi:10.48550/arxiv.2206.14176 · Cited by 76 · see citing works
“4 Related Work Existing work on robot learning commonly leverages large amounts of simulated experience before deploying to the real world (Rusu et al., 2016; Peng et al., 2018; OpenAI et al., 2018; Lee et al., 2020; Irpan et al., 2020; Kumar et al., 2021; Siekmann et al., 2021; Escontrela et al., 2022), leverage flee…”
BackgroundMethodInfluentialNot Only Rewards but Also Constraints: Applications on Legged Robot Locomotion
KIM, Y. H., Oh, H., Lee, J., Choi, J., Ji, G., Jung, M., Youm, D., & Hwangbo, J. (2024) · IEEE Transactions on Robotics · doi:10.1109/tro.2024.3400935 · Cited by 76 · see citing works
Self-citationHeuristics Integrated Deep Reinforcement Learning for Online 3D Bin Packing
Yang, S., Song, S., Chu, S., Song, R., Cheng, J., Li, Y., & Zhang, W. (2023) · IEEE Transactions on Automation Science and Engineering · doi:10.1109/tase.2023.3235742 · Cited by 75 · see citing works
Physics-Informed Machine Learning for Modeling and Control of Dynamical Systems
Nghiem, T. X., Drgoňa, J., Jones, C. N., Nagy, Z., Schwan, R., Dey, B., Chakrabarty, A., Cairano, S. D., Paulson, J. A., Carron, A., et al. (2023) · Proceedings of the ... American Control Conference/Proceedings of the American Control Conference · doi:10.23919/acc55779.2023.10155901 · Cited by 75 · see citing works
Robust Feedback Motion Policy Design Using Reinforcement Learning on a 3D Digit Bipedal Robot
Castillo, G. A., Weng, B., Zhang, W., & Hereid, A. (2021) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros51168.2021.9636467 · Cited by 74 · see citing works
Adapting Rapid Motor Adaptation for Bipedal Robots
Kumar, A., Li, Z., Zeng, J., Pathak, D., Sreenath, K., & Malik, J. (2022) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros47612.2022.9981091 · Cited by 73 · see citing works
Robust High-Speed Running for Quadruped Robots via Deep Reinforcement Learning
Bellegarda, G., Chen, Y., Liu, Z., & Nguyen, Q. (2022) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros47612.2022.9982132 · Cited by 71 · see citing works
Actor-Critic Model Predictive Control
Romero, Á., Song, Y., & Scaramuzza, D. (2024) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra57147.2024.10610381 · Cited by 71 · see citing works
High-speed quadrupedal locomotion by imitation-relaxation reinforcement learning
Jin, Y., Liu, X., Shao, Y., Wang, H., & Yang, W. (2022) · Nature Machine Intelligence · doi:10.1038/s42256-022-00576-3 · Cited by 69 · see citing works
Robust and Versatile Bipedal Jumping Control through Reinforcement Learning
Li, Z., Peng, X. B., Abbeel, P., Levine, S., Berseth, G., & Sreenath, K. (2023) · Robotics: Science and Systems · doi:10.15607/rss.2023.xix.052 · Cited by 69 · see citing works
Electrohydraulic musculoskeletal robotic leg for agile, adaptive, yet energy-efficient locomotion
Buchner, T. J. K., Fukushima, T., Kazemipour, A., Gravert, S.-D., Prairie, M., Romanescu, P., Arm, P., Zhang, Y., Wang, X., Zhang, S. L., et al. (2024) · Nature Communications · doi:10.1038/s41467-024-51568-3 · Cited by 68 · see citing works
Chu, S., Lin, M., Li, D., Lin, R., & Xiao, S. (2024) · Ocean Engineering · doi:10.1016/j.oceaneng.2024.120139 · Cited by 67 · see citing works
What Matters in Learning from Offline Human Demonstrations for Robot Manipulation
Mandlekar, A., Xu, D., Wong, J., Nasiriany, S., Wang, C., Kulkarni, R., Fei-Fei, L., Savarese, S., Zhu, Y., & Martín-Martín, R. (2021) · arXiv · doi:10.48550/arxiv.2108.03298 · Cited by 66 · see citing works
Mosaic: in-memory computing and routing for small-world spike-based neuromorphic systems
Dalgaty, T., Moro, F., Demirağ, Y., Prà, A. D., Indiveri, G., Vianello, E., & Payvand, M. (2024) · Nature Communications · doi:10.1038/s41467-023-44365-x · Cited by 66 · see citing works
Multistage Cable Routing Through Hierarchical Imitation Learning
Luo, J., Xu, C., Geng, X., Feng, G., Fang, K., Tan, L., Schaal, S., & Levine, S. (2024) · IEEE Transactions on Robotics · doi:10.1109/tro.2024.3353075 · Cited by 66 · see citing works
Han, L., Zhu, Q., Sheng, J., Zhang, C., Li, T., Zhang, Y., Zhang, H., Liu, Y., Zhou, C., Zhao, R., et al. (2024) · Nature Machine Intelligence · doi:10.1038/s42256-024-00861-3 · Cited by 66 · see citing works
Learning Minimum-Time Flight in Cluttered Environments
Pěnička, R., Song, Y., Kaufmann, E., & Scaramuzza, D. (2022) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2022.3181755 · Cited by 65 · see citing works
Minimizing Energy Consumption Leads to the Emergence of Gaits in Legged RobotsPreprint
Fu, Z., Kumar, A., Malik, J., & Pathak, D. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2111.01674 · Cited by 64 · see citing works
Learning-based legged locomotion: State of the art and future perspectives
Ha, S., Lee, J., van de Panne, M., Xie, Z., Yu, W., & Khadiv, M. (2025) · The International Journal of Robotics Research · doi:10.1177/02783649241312698 · Cited by 63 · see citing works
Self-citationLearning Visible Connectivity Dynamics for Cloth Smoothing
Lin, X., Wang, Y., Huang, Z., & Held, D. (2021) · arXiv · doi:10.48550/arxiv.2105.10389 · Cited by 62 · see citing works
Recent Progress in Legged Robots Locomotion Control
Carpentier, J., & Wieber, P.-B. (2021) · Current Robotics Reports · doi:10.1007/s43154-021-00059-0 · Cited by 61 · see citing works
Shi, H., Zhou, B., Zeng, H., Wang, F., Dong, Y., Li, J., Wang, K., Tian, H., & Meng, M. Q. ‐. (2022) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2022.3145495 · Cited by 61 · see citing works
Fast and Efficient Locomotion via Learned Gait TransitionsPreprint
Yang, Y., Zhang, T., Coumans, E., Tan, J., & Boots, B. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2104.04644 · Cited by 60 · see citing works
Measuring and modeling the motor system with machine learning
Hausmann, S. B., Vargas, A. M., Mathis, A., & Mathis, M. W. (2021) · Current Opinion in Neurobiology · doi:10.1016/j.conb.2021.04.004 · Cited by 59 · see citing works
Rough Terrain Navigation for Legged Robots using Reachability Planning and Template Learning
Wellhausen, L., & Hutter, M. (2021) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros51168.2021.9636358 · Cited by 59 · see citing works
Self-citationA Secure Robot Learning Framework for Cyber Attack Scheduling and Countermeasure
Wu, C., Yao, W., Luo, W., Pan, W., Sun, G., Xie, H., & Wu, L. (2023) · IEEE Transactions on Robotics · doi:10.1109/tro.2023.3275875 · Cited by 59 · see citing works
Wing-strain-based flight control of flapping-wing drones through reinforcement learning
Kim, T., Hong, I., Im, S., Rho, S., Kim, M., Roh, Y., Kim, C., Park, J., Lim, D., Lee, D., et al. (2024) · Nature Machine Intelligence · doi:10.1038/s42256-024-00893-9 · Cited by 57 · see citing works
Self-citationA Survey of Traversability Estimation for Mobile Robots
Sevastopoulos, C., & Konstantopoulos, S. (2022) · IEEE Access · doi:10.1109/access.2022.3202545 · Cited by 55 · see citing works
GLiDE: Generalizable Quadrupedal Locomotion in Diverse Environments with a Centroidal Model
Xie, Z., Da, X., Babich, B., Garg, A., & van de Panne, M. (2022) · Springer proceedings in advanced robotics · doi:10.1007/978-3-031-21090-7_31 · Cited by 55 · see citing works
Circus ANYmal: A Quadruped Learning Dexterous Manipulation with Its Limbs
Shi, F., Homberger, T., Lee, J., Miki, T., Zhao, M., Farshidian, F., Okada, K., Inaba, M., & Hutter, M. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9561926 · Cited by 55 · see citing works
Self-citationRMA: Rapid Motor Adaptation for Legged RobotsPreprint
Kumar, A., Fu, Z., Pathak, D., & Malik, J. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2107.04034 · Cited by 54 · see citing works
Self-supervised Learning of LiDAR Odometry for Robotic Applications
Nubert, J., Khattak, S., & Hutter, M. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9561063 · Cited by 53 · see citing works
Self-citationLegs as Manipulator: Pushing Quadrupedal Agility Beyond Locomotion
Cheng, X., Kumar, A., & Pathak, D. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10161470 · Cited by 53 · see citing works
- Authors
- Lee, J., Hwangbo, J., Wellhausen, L., Koltun, V., & Hutter, M.
- Affiliations
- ETH Zurich (CH)
- Korea Advanced Institute of Science and Technology (KR)
- Intel (United States) (US)
- Corresponding author
- Lee, Joonho
- Venue
- Science Robotics
- Published
- 2020-10-21
- Volume
- 5
- Issue
- 47
- ISSN
- 2470-9476 (ISSN-L)
- OpenAlex
- W3093922502
- Semantic Scholar
- eadbe2e4f9de47dd357589cf59e3d1f0199e5075
- Cited by
- 1,156
- References
- 29
- Citations by year
- 2026: 184 · 2025: 298 · 2024: 244 · 2023: 209 · 2022: 140 · 2021: 79 · 2020: 2
- FWCI
- 33.76
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- Science Robotics (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- H2020 European Research Council (852044)
- Horizon 2020 (780883)
- National Centre of Competence in Research Robotics
- Intel Network on Intelligent Systems
- Topics
- Physical Sciences › Engineering › Biomedical Engineering › Robotic Locomotion and Control
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Physical Sciences › Engineering › Biomedical Engineering › Prosthetics and Rehabilitation Robotics
- Keywords
- quadrupedal locomotion, legged robots, proprioceptive feedback, reinforcement learning, sim-to-real transfer, zero-shot generalization, deformable terrain, neural network policy
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
