Learning from Demonstrations Through the Use of Non-rigid Registration
AI snapshot
Written from the paper’s full text.
TL;DR
The paper presents trajectory transfer, a method that adapts a demonstrated end-effector trajectory to a new scene by non-rigidly registering the training point cloud to the test point cloud and warping the trajectory accordingly. It validates the approach mainly by enabling a PR2 robot to tie five knot types from demonstrations, and also shows qualitative results on other manipulation tasks.
Key findings
- With one demonstration per knot, success under random rope perturbations depended on perturbation size: at \(d_{pert}=3\) cm, success was 5/5 for overhand, figure-eight, and square, 3/5 for double-overhand, and 1/5 for clove hitch; at \(d_{pert}=10\) cm, success was 4/5, 1/5, 3/5, 3/5, and 0/5 respectively.
- The PR2 autonomously tied five knot types: overhand, figure-eight, double-overhand, square knot, and clove hitch around a pole.
- For the overhand knot, the authors collected 148 trajectory segments, including 36 demonstrations of the standard 3-step sequence and 40 additional segments starting from failure states; the nearest-neighbor policy was qualitatively able to tie knots from a larger set of initial configurations and recover from failure states.
- The same algorithm and code were applied to folding a T-shirt, picking up a plate with a two-arm motion, and opening a bottle, with qualitative success reported.
Methods
- Trajectory transfer: register the training scene point cloud to the test scene point cloud using a modified TPS-RPM non-rigid registration algorithm, then apply the resulting transformation \(f\) to demonstrated end-effector positions and use the Jacobian \(Jf(p)\) to transform orientations.
- Convert the transferred end-effector trajectory into a joint trajectory by solving an optimization problem that trades off smoothness, end-effector pose error, collision avoidance, and joint limits, using the trajectory optimization method from Schulman et al. and initializing from the demonstration.
- For multi-step tasks, repeatedly choose the nearest demonstration by registration cost, apply trajectory transfer, and execute the resulting trajectory; experiments used a PR2 with kinesthetic teaching and RGBD point clouds, with rope points extracted by color filtering.
Limitations
- The paper states that the cost-function invariance assumption in Equation (7) is strong and can fail even for rigid transformations, for example when absolute end-effector orientation matters.
- The paper states that dynamics covariance holds only approximately in knot tying and was checked qualitatively by comparing final rope states at training and test time.
- The paper reports only qualitative results for the many-demonstration overhand knot policy and for the other manipulation tasks; no quantitative success rates are given for those settings. The single-demonstration results also show low success for clove hitch and for larger perturbations.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 4 cited sources checked · 0 supported · 0 flagged
- Findingown data§ Learning from Demonstrations Through the Use of Non-Rigid Registration · p. 1
A PR2 robot autonomously tied five different types of knots in rope: overhand, figure-eight, double-overhand, square knot, and clove hitch (around a pole).
“As our main experimental validation, we enable a PR2 robot to autonomously tie five different types of knots in rope.”
- Also stated in: § 5 Experiments with rope · p. 9 · plain
“We enabled the PR2 to autonomously tie five different types of knot: overhand, figure-eight, double-overhand, square knot, and clove hitch (around a pole).”
Quote in the paperSubstantial claim - Also stated in: § 5 Experiments with rope · p. 9 · plain
- Findingnone given§ 1 Introduction · p. 2
The nearest-neighbor policy recovers from failures, provided that the demonstrator has provided examples of the failure states along with the corrective motions that get out of them.
“This policy recovers from failures, assuming that the demonstrator has provided examples of the failure states along with the corrective motions that get out of them.”
- Also stated in: § 4 Generalizing from many examples: nearest neighbor method · p. 9 · plain
“This policy enables the robot to recover from errors, provided that the demonstrator has provided a demonstration starting from the failure state.”
Quote in the paperNot a substantial claim- The authors state this as an assumption: The demonstrator has provided examples of the failure states along with the corrective motions that get out of them.
- Also stated in: § 4 Generalizing from many examples: nearest neighbor method · p. 9 · plain
- Findingown data§ 5 Experiments with rope · p. 10
The nearest-neighbor policy is able to choose corrective movements to recover from the failure states.
“and it is also able to choose corrective movements to recover from the failure states.”
Quote in the paperSubstantial claim - Findingown data§ 9.1 Thin plate splines · p. 15
The regularization r(B) was necessary because sometimes the transformation is underdetermined in certain dimensions.
“We found that the regularization r ( B ) was necessary because sometimes the transformation is underdetermined in certain dimensions.”
Quote in the paperSubstantial claim - Findingown data§ 7 Conclusion · p. 12
The method successfully picked up a plate when the training round plate was registered to a rectangular plate at test time.
“Bottom row, left: demonstration and autonomous execution of plate pickup. Round plate from training was registered to rectangular plate at test time.”
Quote in the paperSubstantial claim
- Findingown data§ 5 Experiments with rope · p. 10
The nearest-neighbor policy is able to successfully tie knots from a much larger set of initial configurations than was possible with a single demonstration.
“we find that the nearest-neighbor policy from Section 4 is able to successfully tie knots from a much larger set of initial configurations than was possible with a single demonstration”
Quote in the paperSubstantial claim - Findingown data§ 5 Experiments with rope · p. 10
A common failure mode was that the rope would end up in a previously unseen crossing state, often due to a small difference in its position.
“Two of the common failure modes were (1) the rope would end up in a previously unseen crossing state, often due to a small difference in its position; (2) the robot would grab the wrong piece of rope, or two pieces instead of one.”
Quote in the paperNot a substantial claim - Findingown data§ 3.2 Trajectory transfer procedure · p. 6
In our experiments, initializing the joint trajectory optimization with the joint trajectory from the demonstration led to finding good locally-optimal trajectories.
“In our experiments, this initialization strategy led to finding good locally-optimal trajectories.”
Quote in the paperSubstantial claim - Findingown data§ 7 Conclusion · p. 12
The method successfully executed T-shirt folding on a small T-shirt based on a single demonstration on a large T-shirt.
“executing shirt folding procedure autonomously on small T-shirt, based on single demonstration on large T-shirt.”
Quote in the paperSubstantial claim - Findingown data§ 5 Experiments with rope · p. 10
For knot tying from a single demonstration, the failure rate increases at higher perturbation distance d_pert, where the test situation is further away from the training situation.
“As expected, the failure rate increases at higher d pert, where the test situation is further away from the training situation.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 1 Introduction · p. 2
Trajectory transfer enables the robot to generalize from a single demonstration to a much larger part of the state space.
“Trajectory transfer enables the robot to generalize from a single demonstration to a much larger part of the state space.”
Quote in the paperNot a substantial claim - Interpretationreasoning§ 1 Introduction · p. 2
The nearest-neighbor procedure, using registration cost to choose the nearest demonstration and then applying trajectory transfer, can be used as a complete policy that enables the robot to complete a multi-step task, eliminating the need to program a custom state machine.
“This simple procedure can be used as a complete policy that enables the robot to complete a multi-step task, eliminating the need to program a custom state machine.”
Quote in the paperSubstantial claim - Premisecitation§ 2 Related work · p. 3
The approach of Calinon and Billard assumes access to a featurization of the environment, since regression requires an input vector of fixed dimensionality.
“The approach of Calinon and Billard assumes access to a featurization of the environment, since regression requires an input vector of fixed dimensionality.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [8] Calinon, S., Guenter, F., Billard, A.: On learning, representing, and generalizing a task in a humanoid robot. IEEE Trans. Syst. Man Cybern. Part B: Cybern. 37(2), 286–298 (2007) source
- cites [9] Calinon, S., D’halluin, F., Caldwell, D., Billard, A.: Handling of multiple constraints and motion alternatives in a robot programming by demonstration framework. In: 9th IEEE-RAS International Conference on Humanoid Robots, 2009. Humanoids 2009, pp. 582–588, IEEE, New Jersey (2009) source
- Interpretationreasoning§ 3.3.1 Cost function invariance · p. 7
Trajectory transfer can be justified by an invariance assumption relating it to cost function learning.
“Our trajectory transfer procedure can be justified by an invariance assumption, which relates it to cost function learning.”
Quote in the paperSubstantial claim- The authors state this as an assumption: There exists a class of smooth transformations f such that L(STATE, TRAJ) = L(f(STATE), f(TRAJ)).
- Premisecitation§ 3.1.1 Registration with known correspondences · p. 4
Thin plate splines extrapolate well on spatial data.
“It has been observed in other fields (e.g., [17]) that thin plate splines extrapolate well on spatial data; extrapolation is important for our application.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [17] Carr, J.C., Beatson, R.K., Cherrie, J.B., Mitchell, T.J., Fright, W.R., McCallum, B.C., Evans, T.R.: Reconstruction and representation of 3D objects with radial basis functions. In: Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pp. 67–76. ACM, New York (2001) source
- Interpretationnone given§ 7 Conclusion · p. 12
The non-rigid registration metric can be used to find the nearest demonstration when multiple demonstrations have been provided.
“The non-rigid registration metric can also be used to find the nearest demonstration, when multiple demonstrations have been provided.”
Quote in the paperNot a substantial claim - Interpretationnone given§ 7 Conclusion · p. 12
This simple scheme enables our algorithm to successfully perform challenging multi-step manipulation tasks.
“This simple scheme enables our algorithm to successfully perform challenging multi-step manipulation tasks.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 2 Related work · p. 3
The approach of Calinon and Billard is not applicable to knot tying, where there is no fixed-length vector of landmarks that can be extracted from every rope configuration.
“Their approach is not applicable to knot tying, where there is no fixed-length vector of landmarks that can be extracted from every rope configuration.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 3.3.2 Dynamics invariance · p. 8
Trajectory transfer works in scenarios where the dynamics of the system are approximately invariant, more properly covariant, under sufficiently smooth coordinate transformations.
“An alternative perspective is that trajectory transfer works in scenarios where the dynamics of the system are approximately invariant-more properly, covariant -under sufficiently smooth coordinate transformations.”
Quote in the paperSubstantial claim- The authors state this as an assumption: The dynamics of the system are covariant under the transformation f (Equation 10 holds).
- May combine two claims. 5
- The quote only partly states this claim. 1
- Classified as probably not a substantial claim. 4
May combine two claims.
Findingown data§ 5 Experiments with rope · p. 10A common failure mode was that the robot would grab the wrong piece of rope, or two pieces instead of one.
“Two of the common failure modes were (1) the rope would end up in a previously unseen crossing state, often due to a small difference in its position; (2) the robot would grab the wrong piece of rope, or two pieces instead of one.”
Quote in the paperNot a substantial claimThe quote only partly states this claim.
Findingown data§ 7 Conclusion · p. 12The method successfully opened a bottle whose size and shape differed from the one used for training.
“Bottom row, right: demonstration and autonomous execution of bottle opening. The bottle used at test time had a different size and shape from the one used for training.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ 3.3.2 Dynamics invariance · p. 8Assumption 4, that the dynamics are covariant under f, holds approximately in the knot tying domain; this was checked qualitatively by comparing the final states of the rope at training and test time.
“Assumption 4 holds approximately in this domain; we have checked this qualitatively by comparing the final states of the rope at training and test time.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown data§ 3.2 Trajectory transfer procedure · p. 6Inverse kinematics alone is not sufficient for converting the transferred end-effector trajectory into a joint trajectory because there is often no continuous and collision-free trajectory that achieves the desired end-effector trajectory.
“We found that this approach was not sufficient because there is often no continuous and collision-free trajectory that achieves the desired end-effector trajectory, so we need to compromise.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationnone given§ 9.1 Thin plate splines · p. 15With the quadratic regularization term r(B)=trace((B-I)^T D (B-I)), the optimization problem still can be solved analytically as a least squares problem.
“With this quadratic regularization term, the optimization problem still can be solved analytically as a least squares problem.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationreasoning§ 3.3.1 Cost function invariance · p. 7There exist situations where the cost function invariance assumption (Equation 7) does not hold, even for a rigid transformation f, for example when absolute orientation of the robot's end-effector matters.
“That said, one can imagine situations where Equation (7) does not hold, even for a rigid transformation f -for example, when absolute orientation of the robot's end-effector matters.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 3.1.1 Registration with known correspondences · p. 4Equation (1) can be analytically solved efficiently.
“As shown in [18], Equation (1) can be analytically solved efficiently; details are provided in Appendix 9.1.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [18] Wahba, G.: Spline models for observational data, Society for Industrial Mathematics, vol. 59 (1990) source
May combine two claims.
Interpretationown data§ 5 Experiments with rope · p. 10When starting from a state that is very close to a demonstration, the success rate is high, at least for the easier knots.
“The above results suggest that when starting from a state that is very close to a demonstration, the success rate is high, at least for the easier knots.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Interpretationreasoning§ 3.3.2 Dynamics invariance · p. 8In the knot tying domain, the goal set is defined topologically, and a transformation f will preserve it, provided that it is a homeomorphism.
“In the knot tying domain, the goal set is defined topologically, and a transformation f will preserve it, provided that it is a homeomorphism (i.e., a continuous function whose inverse is continuous), justifying assumption 3.”
Quote in the paperSubstantial claimMay combine two claims.
Interpretationreasoning§ 2 Related work · p. 3Our approach operates directly on point clouds and is suitable for knot tying, and it can be applied to a new task without developing a new vision system.
“In contrast, our approach operates directly on point clouds-the outputs of our sensor hardware-so it is suitable for this task, and it can be applied to a new task without developing a new vision system.”
Quote in the paperSubstantial claim
The authors state this as an assumption.
The demonstrator has provided examples of the failure states along with the corrective motions that get out of them.
“This policy recovers from failures, assuming that the demonstrator has provided examples of the failure states along with the corrective motions that get out of them.”
The authors state this as an assumption.
There exists a class of smooth transformations f such that L(STATE, TRAJ) = L(f(STATE), f(TRAJ)).
“Our trajectory transfer procedure can be justified by assuming that there is a class of smooth transformations f with the following property:”
The authors state this as an assumption.
The dynamics of the system are covariant under the transformation f (Equation 10 holds).
“4. Equation (10) holds, i.e., the dynamics are covariant under f .”
The authors state this as an assumption.
The goal set is preserved by f (f(G)=G).
“3. f ( G ) = G , i.e., the goal set is preserved by f”
The authors state this as an assumption.
f transforms STATE_train into STATE_test (f(STATE_train)=STATE_test).
“1. f ( STATE t train ) = STATE t test , i.e., f transforms STATE train into STATEtest.”
The authors state this as an assumption.
The demonstration ends up in the goal state (STATE_{t+1}^{train} ∈ G).
“2. STATE t + 1 train ∈ G , i.e., the demonstration ends up in the goal state.”
The paper relies on this without support.
For each segment, the end-effector trajectory depends on the geometry of the rope in some way that must be inferred by the algorithm.
“For each of these segments, the end-effector trajectory ought to depend on the geometry of the rope in some way that must be inferred by the algorithm.”
The paper relies on this without support.
The procedure for tying a given type of knot can be broken into several segments based on grab and release events.
“The procedure for tying a given type of knot can be broken into several segments based on grab and release events.”
Cites
Spline Models for Observational Data
Wahba, G. (1990) · doi:10.2307/2938687 · Cited by 7,398
ROS: an open-source Robot Operating System
Quigley, M. (2009) · Medical Entomology and Zoology · Cited by 7,121 · see citing works
Shape matching and object recognition using shape contexts
Belongie, S., Malik, J., & Puzicha, J. (2002) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/34.993558 · Cited by 6,306 · see citing works
Bradski, G. R. (2000) · Medical Entomology and Zoology · Cited by 5,994 · see citing works
Spline Models for Observational Data.
Müller, H., & Wahba, G. (1991) · Journal of the American Statistical Association · doi:10.2307/2290434 · Cited by 5,016 · see citing works
3D is here: Point Cloud Library (PCL)
Rusu, R. B., & Cousins, S. (2011) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2011.5980567 · Cited by 4,980 · see citing works
Spline Models for Observational Data
Wahba, G. (1990) · doi:10.1137/1.9781611970128 · Cited by 4,184 · see citing works
Abraham, R., & Young, H. D. (1968) · American Journal of Physics · doi:10.1119/1.1974504 · Cited by 3,129 · see citing works
Image quilting for texture synthesis and transfer
Efros, A. A., & Freeman, W. T. (2001) · ACM SIGGRAPH Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH) · doi:10.1145/383259.383296 · Cited by 2,426 · see citing works
2008 IEEE/RSJ International Conference on Intelligent Robots and Systems
(2008) · doi:10.1109/iros13647.2008 · Cited by 2,203 · see citing works
Hill, D. L. G., Batchelor, P. G., Holden, M., & Hawkes, D. J. (2001) · Physics in Medicine and Biology · doi:10.1088/0031-9155/46/3/201 · Cited by 2,093 · see citing works
Reconstruction and representation of 3D objects with radial basis functions
Carr, J. C., Beatson, R. K., Cherrie, J. B., Mitchell, T. J., Fright, W. R., McCallum, B. C., & Evans, T. R. (2001) · Proceedings of the 28th annual conference on Computer graphics and interactive techniques · doi:10.1145/383259.383266 · Cited by 1,852 · cited 1× · see citing works
“, [17]) that thin plate splines extrapolate well on spatial data; extrapolation is”
BackgroundA new point matching algorithm for non-rigid registration
Chui, H., & Rangarajan, A. (2003) · Computer Vision and Image Understanding · doi:10.1016/s1077-3142(03)00009-2 · Cited by 1,560 · see citing works
On Learning, Representing, and Generalizing a Task in a Humanoid Robot
Calinon, S., Guenter, F., & Billard, A. (2007) · IEEE Transactions on Systems, Man and Cybernetics, Part B (Cybernetics) · doi:10.1109/tsmcb.2006.886952 · Cited by 1,090 · see citing works
Robot Programming by Demonstration
Billard, A., Calinon, S., Dillmann, R., & Schaal, S. (2008) · Springer eBooks · doi:10.1007/978-3-540-30301-5_60 · Cited by 1,078 · see citing works
Ensemble of exemplar-SVMs for object detection and beyond
Malisiewicz, T., Gupta, A., & Efros, A. A. (2011) · 2011 International Conference on Computer Vision · doi:10.1109/iccv.2011.6126229 · Cited by 865 · see citing works
A new algorithm for non-rigid point matching
Chui, H., & Rangarajan, A. (2002) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2000.854733 · Cited by 443 · see citing works
Finding Locally Optimal, Collision-Free Trajectories with Sequential Convex Optimization
Schulman, J., Ho, J., Lee, A., Awwal, I., Bradlow, H., & Abbeel, P. (2013) · Robotics: Science and Systems · doi:10.15607/rss.2013.ix.031 · Cited by 442 · see citing works
Self-citationOpenRAVE: A Planning Architecture for Autonomous Robotics
Diankov, R., & Kuffner, J. (2008) · Cited by 381 · see citing works
Nonparametric scene parsing: Label transfer via dense scene alignment
Liu, C., Yuen, J., & Torralba, A. B. (2009) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2009.5206536 · Cited by 313 · see citing works
International Symposium on Robotics Research
(2013) · The International Journal of Robotics Research · doi:10.1177/0278364913487641 · Cited by 259 · see citing works
Example-Based 3D Scan Completion
Pauly, M. V., Mitra, N. J., Giesen, J., Groß, M. H., & Guibas, L. (2005) · doi:10.2312/sgp/sgp05/023-032 · Cited by 233 · see citing works
Knotting/unknotting manipulation of deformable linear objects
Wakamatsu, H., Arai, E., & Hirai, S. (2006) · The International Journal of Robotics Research · doi:10.1177/0278364906064819 · Cited by 132 · see citing works
Motion Planning for Robotic Manipulation of Deformable Linear Objects
Saha, M., Isto, P., & Latombe, J. (2008) · Springer eBooks · doi:10.1007/978-3-540-77457-0_3 · Cited by 101 · see citing works
Hand eye coordination in rope handling
INABA, M., & INOUE, H. (1985) · Journal of the Robotics Society of Japan · doi:10.7210/jrsj.3.538 · Cited by 95 · see citing works
Demonstration-Guided Motion Planning
Gu, Y., & Alterovitz, R. (2016) · Springer tracts in advanced robotics · doi:10.1007/978-3-319-29363-9_17 · Cited by 84 · see citing works
Motion planning for robotic manipulation of deformable linear objects
Saha, M., & Isto, P. (2006) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/robot.2006.1642074 · Cited by 81 · see citing works
Hand-eye coordination in rope handling
Inoue, H. (1984) · Medical Entomology and Zoology · Cited by 77 · see citing works
Calinon, S., D'halluin, F., Caldwell, D. G., & Billard, A. G. (2009) · 2009 9th IEEE-RAS International Conference on Humanoid Robots · doi:10.1109/ichr.2009.5379592 · Cited by 73 · see citing works
(2003) · doi:10.1109/robot.2003.1242193 · Cited by 73 · see citing works
One-handed knotting of a flexible rope with a high-speed multifingered hand having tactile sensors
Yamakawa, Y., Namiki, A., Ishikawa, M., & Shimojo, M. (2007) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2007.4399379 · Cited by 61 · see citing works
Bell, M. (2010) · doi:10.1349/ddlp.367 · Cited by 35 · see citing works
Nonparametric scene parsing: Label transfer via dense scene alignment
Liu, C., Yuen, J., & Torralba, A. B. (2009) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvprw.2009.5206536 · Cited by 9 · see citing works
Knotting/raveling manipulation of linear objects
Wakamatsu, H., TSUMAYA, A., SHIRASE, K., Arai, E., & Hirai, S. (2004) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2003.1249642 · Cited by 5 · see citing works
Flagged first, then by citations
Cited by
Sorted by citations
An Introduction to Deep Reinforcement Learning
François-Lavet, V., Henderson, P., Islam, R., Bellemare, M. G., & Pineau, J. (2018) · Foundations and Trends® in Machine Learning · doi:10.1561/2200000071 · Cited by 1,353 · see citing works
“Thus, transfer learning may also be of crucial importance (Schulman et al., 2016; Liu et al., 2017).”
Recent Advances in Robot Learning from Demonstration
Ravichandar, H. C., Polydoros, A., Chernova, S., & Billard, A. (2019) · Annual Review of Control Robotics and Autonomous Systems · doi:10.1146/annurev-control-100819-063206 · Cited by 762 · see citing works
Deep Imitation Learning for Complex Manipulation Tasks from Virtual Reality Teleoperation
Zhang, T., McCarthy, Z., Jow, O., Lee, D., Chen, X., Goldberg, K., & Abbeel, P. (2018) · 2018 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra.2018.8461249 · Cited by 609 · see citing works
“It has been applied successfully to a wide range of domains in robotics, for example to autonomous driving [4], [5], [6], autonomous helicopter flight [7], gesturing [8], and manipulation [9], [10].”
Self-citationAn algorithmic perspective on imitation learning
Osa, T., Pajarinen, J., Neumann, G., Bagnell, J. A., Abbeel, P., & Peters, J. (2018) · Foundations and Trends in Robotics · doi:10.1561/2300000053 · Cited by 405 · see citing works
“Way points with non-rigid registration [Schulman et al., 2013] A point cloud of the given scene Generalization based on point clouds of a given scene Stochasticity of demonstrations is not considered”
BackgroundMethodInfluentialSelf-citationAn Algorithmic Perspective on Imitation Learning
Osa, T., Pajarinen, J., Neumann, G. M., Bagnell, J. A., Abbeel, P., & Peters, J. (2018) · now publishers, Inc. eBooks · doi:10.1561/9781680834116 · Cited by 387 · see citing works
Self-citationModeling, learning, perception, and control methods for deformable object manipulation
Yin, H., Varava, A., & Kragic, D. (2021) · Science Robotics · doi:10.1126/scirobotics.abd8803 · Cited by 299 · see citing works
“Subsequent research generalizes motions for configurations that are very different from demonstrations (13, 108) and connect deformable shape registration to adaptation of demonstrated trajectories.”
BackgroundKassahun, Y., Yu, B., Tibebu, A. T., Stoyanov, D., Giannarou, S., Metzen, J. H., & Vander Poorten, E. (2016) · International Journal of Computer Assisted Radiology and Surgery · doi:10.1007/s11548-015-1305-z · Cited by 238 · see citing works
Cable manipulation with a tactile-reactive gripper
She, Y., Wang, S., Dong, S., Sunil, N., Rodríguez, A., & Adelson, E. H. (2021) · The International Journal of Robotics Research · doi:10.1177/02783649211027233 · Cited by 267 · see citing works
Neural Descriptor Fields: SE(3)-Equivariant Object Representations for Manipulation
Simeonov, A., Du, Y., Tagliasacchi, A., Tenenbaum, J. B., Rodríguez, A., Agrawal, P., & Sitzmann, V. (2022) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra46639.2022.9812146 · Cited by 148 · see citing works
“Template-matching with coarse 3D primitives [15, 23, 42] or non-rigid registration [36] can”
BackgroundMethodLearning force-based manipulation of deformable objects from multiple demonstrations
Lee, A. X., Lu, H., Gupta, A., Levine, S., & Abbeel, P. (2015) · 2015 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra.2015.7138997 · Cited by 140 · see citing works
“[1] for warping the demonstrated trajectories using the TPS warping function.”
BackgroundMethodResultInfluentialSelf-citationSeita, D., Florence, P., Tompson, J., Coumans, E., Sindhwani, V., Goldberg, K., & Zeng, A. (2021) · 2021 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra48506.2021.9561391 · Cited by 130 · see citing works
“In recent years, learning-based approaches for manipulation have grown in popularity to improve the generalization of knot-tying and also to manipulate rope towards general target configurations [52], [46], [49], [57], [64], [61].”
BackgroundToward next-generation learned robot manipulation
Cui, J., & Trinkle, J. (2021) · Science Robotics · doi:10.1126/scirobotics.abd9461 · Cited by 116 · see citing works
“Some successful examples are door opening (13), knot tying (14), and picking up of daily objects (15–17).”
BackgroundDemo2Vec: Reasoning Object Affordances from Online Videos
Fang, K., Wu, T.-L., Yang, D., Savarese, S., & Lim, J. J. (2018) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr.2018.00228 · Cited by 105 · see citing works
“To teach a robot about how humans manipulate and interact with objects, previous methods learn this knowledge from simulated agent-object interactions [33, 23, 29], demonstrations observed by the robot camera in the robot workspace [22, 17] or demonstrations observed from a third-person viewpoint [13, 14].”
BackgroundLearning Rope Manipulation Policies Using Dense Object Descriptors Trained on Synthetic Depth Data
Sundaresan, P., Grannen, J., Thananjeyan, B., Balakrishna, A., Laskey, M., Stone, K., Gonzalez, J. E., & Goldberg, K. (2020) · 2020 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra40945.2020.9197121 · Cited by 107 · see citing works
“There is also extensive literature on deformable manipulation [18, 25, 27, 36, 40].”
BackgroundOnline Trajectory Planning and Force Control for Automation of Surgical Tasks
Osa, T., Sugita, N., & Mitsuishi, M. (2017) · IEEE Transactions on Automation Science and Engineering · doi:10.1109/tase.2017.2676018 · Cited by 101 · see citing works
“[9], [15] developed a scheme to generalize demonstrated trajectories for new situations [9], [15], [16].”
Background(2021) · Journal of Machine Learning Research · Cited by 224 · see citing works
“, the tip of a tool [99]), or use nonrigid registration or geometry warping to determine correspondences of entire sets of points across task instances [338, 156, 361, 315, 12].”
BackgroundRobot Programming by Demonstration with Interactive Action Visualizations
Alexandrova, S., Çakmak, M., Hsiao, K., & Takayama, L. A. (2014) · Robotics: Science and Systems · doi:10.15607/rss.2014.x.048 · Cited by 88 · see citing works
“include invariance of certain state variables [10, 11], allowed variance of certain state variables [5, 6], characteristics of the task in different parts of the state space [15, 8], relativeness to objects in the environment [12], segmentations of the task [12], task constraints [14] or partial-ordering of task steps…”
BackgroundDeep Transfer Learning of Pick Points on Fabric for Robot Bed-Making
Seita, D., Jamali, N., Laskey, M., Tanwani, A. K., Berenstein, R., Baskaran, P., Iba, S., Canny, J., & Goldberg, K. Y. (2022) · Springer proceedings in advanced robotics · doi:10.1007/978-3-030-95459-8_17 · Cited by 88 · see citing works
“ng to design and not generally available o-the-shelf for robotics tasks. Additional fabric manipulation techniques include trajectory transfer and learning from image-based wrinkles. Schulman et al. [27] propose to adapt a demonstrated trajectory (including pick points) from training time to the geometry at test time…”
BackgroundTransition State Clustering: Unsupervised Surgical Trajectory Segmentation for Robot Learning
Krishnan, S., Garg, A., Patil, S., Lea, C., Hager, G. D., Abbeel, P., & Goldberg, K. Y. (2017) · Springer proceedings in advanced robotics · doi:10.1007/978-3-319-60916-4_6 · Cited by 72 · see citing works
Self-citationMandlekar, A., Ramos, F. T., Boots, B., Savarese, S., Fei-Fei, L., Garg, A., & Fox, D. (2020) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra40945.2020.9196935 · Cited by 79 · see citing works
“The loss function for the RNN is a simple Behavioral Cloning loss L θ ( a t : t + T , s t : t + T .”
MethodInfluentialA Framework for Manipulating Deformable Linear Objects by Coherent Point Drift
Tang, T., Wang, C., & Tomizuka, M. (2018) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2018.2852770 · Cited by 75 · see citing works
Artificial Intelligence and Digital Microscopy Applications in Diagnostic Hematopathology
Achi, H. E., & Khoury, J. D. (2020) · Cancers · doi:10.3390/cancers12040797 · Cited by 75 · see citing works
“trained robots to tie knots using human demonstration via trajectory tra sferring [5].”
MethodChi, W., Liu, J., Rafii-Tari, H., Riga, C., Bicknell, C. D., & Yang, G. (2018) · International Journal of Computer Assisted Radiology and Surgery · doi:10.1007/s11548-018-1743-5 · Cited by 69 · see citing works
“Recent studies have looked into generalizing learned demonstrations to previously unseen initial conditions [7], as well as an adaptive trajectory planning to deal with dynamic changes in the environment [8].”
BackgroundMethodYou Only Demonstrate Once: Category-Level Manipulation from Single Visual Demonstration
Wen, B., Lian, W., Bekris, K., & Schaal, S. (2022) · Robotics: Science and Systems XVIII · doi:10.15607/rss.2022.xviii.044 · Cited by 70 · see citing works
“dynamic models, can be further grouped into two types: policy learning [15]–[17] and trajectory learning [18]–[20].”
BackgroundManipulating deformable objects by interleaving prediction, planning, and control
McConachie, D., Dobson, A., Ruan, M., & Berenson, D. (2020) · The International Journal of Robotics Research · doi:10.1177/0278364920918299 · Cited by 61 · see citing works
“…on learning from demonstration avoid planning and deformable object modelling challenges entirely by using offline demonstrations to teach the robot specific manipulation tasks (Huang et al. 2015; Schulman et al. 2016); however, when a new task is attempted a new training set needs to be generated.”
BackgroundLearning Dense Visual Correspondences in Simulation to Smooth and Fold Real Fabrics
Ganapathi, A., Sundaresan, P., Thananjeyan, B., Balakrishna, A., Seita, D., Grannen, J., Hwang, M., Hoque, R., Gonzalez, J. E., Jamali, N., et al. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9561980 · Cited by 69 · see citing works
“While there is significant prior work on geometric [1, 23, 35, 46] and learning based approaches [36, 37, 47] for fabric manipulation, these approaches often involve designing or learning task-specific manipulation policies, making it difficult to efficiently reuse information across tasks.”
BackgroundAutomated planning for robotic layup of composite prepreg
Malhan, R. K., Shembekar, A. V., Kabir, A. M., Bhatt, P. M., Shah, B. C., Zanio, S., Nutt, S., & Gupta, S. K. (2020) · Robotics and Computer-Integrated Manufacturing · doi:10.1016/j.rcim.2020.102020 · Cited by 58 · see citing works
Movement primitives via optimization
Dragan, A. D., Muelling, K., Bagnell, J. A., & Srinivasa, S. S. (2015) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2015.7139510 · Cited by 52 · see citing works
“Among several tools for addressing this problem [1, 4–6], a commonly used one is a Dynamic Movement Primitive (DMP) [7, 8].”
MethodAn online method for tight-tolerance insertion tasks for string and rope
Wang, W., Berenson, D., & Balkcom, D. (2015) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2015.7139532 · Cited by 51 · see citing works
“We use a quasi-static model, so that the the equilibrium configuration of the flexible object is completely determined by the configuration(s) of the grippers (represented by a vector q) together with internal spring and external gravity forces.”
MethodDeep Imitation Learning for Complex Manipulation Tasks from Virtual Reality Teleoperation
Zhang, T., McCarthy, Z., Jow, O., Lee, D., Chen, X., Goldberg, K., & Abbeel, P. (2017) · arXiv · doi:10.48550/arxiv.1710.04615 · Cited by 618 · see citing works
“It has been applied successfully to a wide range of domains in robotics, for example to autonomous driving [4], [5], [6], autonomous helicopter flight [7], gesturing [8], and manipulation [9], [10].”
Self-citationOnline Trajectory Planning in Dynamic Environments for Surgical Task Automation
Osa, T., Sugita, N., & Mitsuishi, M. (2014) · Robotics: Science and Systems · doi:10.15607/rss.2014.x.011 · Cited by 48 · see citing works
Mandlekar, A., Booher, J., Spero, M., Tung, A., Gupta, A., Zhu, Y., Garg, A., Savarese, S., & Fei-Fei, L. (2019) · arXiv · doi:10.48550/arxiv.1911.04052 · Cited by 48 · see citing works
Cloth Manipulation Using Random-Forest-Based Imitation Learning
Jia, B., Pan, Z., Hu, Z., Pan, J., & Manocha, D. (2019) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2019.2897370 · Cited by 42 · see citing works
“Recently, [31, 32, 33] proposed using non-rigid registration to transfer human demonstrations of cloth manipulations to real robots and [34] required an adaptive cloth simulator to predict the future state of a cloth.”
BackgroundMotion Planning for Dual-Arm Manipulation of Elastic Rods
Sintov, A., Macenski, S., Borum, A. D., & Bretl, T. W. (2020) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2020.3011352 · Cited by 45 · see citing works
“surgical suturing [2], knot tying [3], and automated handling of flexible material such as sheet metal parts [4].”
BackgroundLet’s Take a Walk on Superpixels Graphs: Deformable Linear Objects Segmentation and Model Estimation
Gregorio, D. D., Palli, G., & Stefano, L. D. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-20890-5_42 · Cited by 42 · see citing works
Robust Deformation Model Approximation for Robotic Cable Manipulation
Jin, S., Wang, C., & Tomizuka, M. (2019) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros40897.2019.8968157 · Cited by 42 · see citing works
“[10] proposed to teach robots to manipulate deformable objects from human demonstrations.”
BackgroundIterative Residual Policy for Goal-Conditioned Dynamic Manipulation of Deformable Objects
Chi, C., Burchfiel, B., Cousineau, E. A., Feng, S., & Song, S. (2022) · Robotics: Science and Systems · doi:10.15607/rss.2022.xviii.016 · Cited by 43 · see citing works
Mandlekar, A., Booher, J., Spero, M., Tung, A. T. Y., Gupta, A., Zhu, Y., Garg, A., Savarese, S., & Fei-Fei, L. (2019) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros40897.2019.8968114 · Cited by 47 · see citing works
Huang, S. H., Pan, J., Mulcaire, G., & Abbeel, P. (2015) · 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2015.7353475 · Cited by 39 · see citing works
Self-citationRobots of the Lost Arc: Self-Supervised Learning to Dynamically Manipulate Fixed-Endpoint Cables
Zhang, H., Ichnowski, J., Seita, D., Wang, J., Huang, H., & Goldberg, K. Y. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9561630 · Cited by 40 · see citing works
“Various approaches have been proposed for quasistatic manipulation of cables, such as using demonstrations [16], [17], [18], self-supervised learning [19], [20], model-free reinforcement learning [21], [22], contrastive learning [23], dense object descriptors [24], generative modeling [25], and state estimation [26].”
MethodLee, A. X., Gupta, A., Lu, H. H., Levine, S., & Abbeel, P. (2015) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2015.7354120 · Cited by 34 · see citing works
“Thin Plate Splines As in the previous work that uses non-rigid registration for trajectory transfer [1], we use the thin plate spline (TPS) [18], [19] parametrization for the transformation T , which can be obtained by regularizing the second-order derivatives of the transformation.”
BackgroundMethodInfluentialSelf-citationOn the manipulation of articulated objects in human-robot cooperation scenarios
Capitanelli, A., Marco Maratea, A., Mastrogiovanni, F., & Vallati, M. (2018) · Archive of Doctoral Theses and Digital Collections (University of Calabria) · Cited by 34 · see citing works
Lee, A. X., Huang, S. H., Hadfield-Menell, D., Tzeng, E., & Abbeel, P. (2014) · 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems · doi:10.1109/iros.2014.6943185 · Cited by 32 · see citing works
“’s approach to learning from demonstrations [1] calculates the new trajectory by first using a modified thin plate spline robust point matching (TPS-RPM) algorithm [19] to calculate a non-rigid scene registration that maps points from the demonstration scene to the new scene.”
BackgroundMethodResultInfluentialSelf-citationOsa, T., Harada, K., Sugita, N., & Mitsuishi, M. (2014) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2014.6907819 · Cited by 34 · see citing works
“proposed a related method [5], [8] that allows the generalization of demonstrated trajectories to new situations by computing a mapping from a demonstration situation onto a new situation.”
MethodGeneralizing pouring actions between objects using warped parameters
Brandi, S., Kroemer, O., & Peters, J. (2014) · IEEE-RAS International Conference on Humanoid Robots (Humanoids) · doi:10.1109/humanoids.2014.7041426 · Cited by 34 · see citing works
“[7] computed warping functions for the entire scene.”
BackgroundLearning legged swimming gaits from experience
Meger, D., Higuera, J. C. G., Xu, A., et al. (2015) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2015.7139509 · Cited by 35 · see citing works
“This choice is inspired by authors such as Xu et al. [15], [16] who have recently shown that human guidance, in the forms of demonstrations or direct participation, are effective aids for learning practical robotic tasks.”
Surgical GAN: Towards real-time path planning for passive flexible tools in endovascular surgeries
Zhao, Y. H., Wang, Y., Zhang, J., Liu, X., Li, Y., Guo, S., Xu, Y., & Hong, S. (2022) · Neurocomputing · doi:10.1016/j.neucom.2022.05.044 · Cited by 33 · see citing works
Estimating Model Utility for Deformable Object Manipulation Using Multiarmed Bandit Methods
McConachie, D., & Berenson, D. (2018) · IEEE Transactions on Automation Science and Engineering · doi:10.1109/tase.2018.2822669 · Cited by 29 · see citing works
“Approaches like [14] and [15] bypass this challenge by using off-line demonstrations to teach the robot specific manipulation tasks; however, when a new task is attempted, a new training set needs to be generated.”
BackgroundBandit-Based Model Selection for Deformable Object Manipulation
McConachie, D., & Berenson, D. (2020) · Springer proceedings in advanced robotics · doi:10.1007/978-3-030-43089-4_45 · Cited by 25 · see citing works
“Approaches like [6,7] bypass this challenge by using offline demonstrations to teach the robot specific manipulation tasks; however, when a new task is attempted a new training set needs to be generated.”
BackgroundLearning Visual Affordance Grounding From Demonstration Videos
Luo, H., Zhai, W., Zhang, J., Cao, Y. H., & Tao, D. (2023) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2023.3298638 · Cited by 29 · see citing works
“Schulman et al. [29] proposed a method based on nonrigid trajectory transfer for adapting the demonstrated trajectory from the training geometry to the test geometry, enabling the robot to tie different types of knots in the rope automatically.”
MethodRobotic manipulation of deformable objects by tangent space mapping and non-rigid registration
Tang, T., Liu, C., Chen, W., & Tomizuka, M. (2016) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2016.7759418 · Cited by 25 · see citing works
“[1] proposed a non-rigid registration method to teach the robot to manipulate the deformable objects by human demonstration.”
MethodDynamically Consistent Online Adaptation of Fast Motions for Robotic Manipulators
Pekarovskiy, A., Nierhoff, T., Hirche, S., & BUSS, M. J. (2017) · IEEE Transactions on Robotics · doi:10.1109/tro.2017.2765666 · Cited by 24 · see citing works
“A nonrigid registration [23] method is applied to the tasks with variable scene geometry and initial conditions.”
MethodKnot-tying with flying machines for aerial construction
Augugliaro, F., Zarfati, E., Mirjan, A., et al. (2015) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2015.7354218 · Cited by 24 · see citing works
“While rope manipulation and knot-tying have been extensively studied in the context of dexterous manipulation (see, for example, [19] and reference therein), this work approaches knot-tying by taking into account the specific characteristics of flying machines.”
BackgroundUsing dVRK teleoperation to facilitate deep learning of automation tasks for an industrial robot
Liang, J., Mahler, J., Laskey, M., Li, P., & Goldberg, K. Y. (2017) · IEEE International Conference on Automation Science and Engineering (CASE) · doi:10.1109/coase.2017.8256067 · Cited by 22 · see citing works
“Schulman et al. [39] used Thin Plate Splines to warp kinesthetically demonstrated trajectories of rope tying with a PR2 to solve novel test cases.”
MethodLearning generalizable surface cleaning actions from demonstration
Elliott, S., Xu, Z., & Çakmak, M. (2017) · IEEE RO-MAN · doi:10.1109/roman.2017.8172424 · Cited by 23 · see citing works
“dressed challenges in manipulating objects [20], [3], learning high-level task structures [16], [24], and learning from nonexpert demonstrators [8], [26].”
BackgroundA Mobile Manipulation System for One-Shot Teaching of Complex Tasks in Homes
Bajracharya, M., Borders, J., Helmick, D., Kollar, T., Laskey, M., Leichty, J., Ma, J., Nagarajan, U., Ochiai, A., Petersen, J., et al. (2020) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra40945.2020.9196677 · Cited by 24 · see citing works
Face valuing: Training user interfaces with facial expressions and reinforcement learningPreprint
Veeriah, V., Pilarski, P. M., & Sutton, R. S. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1606.02807 · Cited by 21 · see citing works
“There are numerous works exploring learning from demonstration (e.g., Koenig & Mataric, 2012; Schulman et al., 2013; Alizadeh et al., 2014).”
BackgroundOne-shot robot programming by demonstration using an online oriented particles simulation
Groth, C., & Henrich, D. (2014) · IEEE International Conference on Robotics and Biomimetics (ROBIO) · doi:10.1109/robio.2014.7090323 · Cited by 20 · see citing works
“The approach in [18] uses a so called non-rigid registration to transfer”
MethodSakr, M., Freeman, M., der Loos, H. F. M. V., & Croft, E. A. (2020) · IEEE RO-MAN · doi:10.1109/ro-man47096.2020.9223430 · Cited by 21 · see citing works
“Others have used kinesthetic teaching in more complex tasks such as ball paddling [11], box flipping using chopsticks [12] and knot tying [13].”
MethodOnline, Interactive User Guidance for High-dimensional, Constrained Motion Planning
Islam, F., Salzman, O., & Likhachev, M. A. (2018) · International Joint Conference on Artificial Intelligence · doi:10.24963/ijcai.2018/683 · Cited by 19 · see citing works
“While there has been much work on planning using human demonstration (see, e.g., [2], [14], [25], [28], [35]), there has been far less research incorporating guidance as an interactive part of the planning loop.”
BackgroundGanapathi, A., Sundaresan, P., Thananjeyan, B., Balakrishna, A., Seita, D., Grannen, J., Hwang, M., Hoque, R., Gonzalez, J. E., Jamali, N., et al. (2020) · Cited by 19 · see citing works
Real-Time Non-Visual Shape Estimation and Robotic Dual-Arm Manipulation Control of an Elastic Wire
Mishani, I., & Sintov, A. (2021) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2021.3128707 · Cited by 19 · see citing works
“Other applications include knot tying [5], surgical suturing [6], hot wire carving recently demonstrated by a robotic system [7] or manipulation of cables using autonomous aerial vehicles [8].”
BackgroundLocal Neural Descriptor Fields: Locally Conditioned Object Representations for Manipulation
Chun, E., Du, Y., Simeonov, A., et al. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10160423 · Cited by 18 · see citing works
“When object models are known, pose estimation may be used for manipulation [22, 30, 31].”
MethodTrajectory-Based Skill Learning Using Generalized Cylinders
Ahmadzadeh, S. R., & Chernova, S. (2018) · Frontiers in Robotics and AI · doi:10.3389/frobt.2018.00132 · Cited by 18 · see citing works
Robotic Compliant Object Prying Using Diffusion Policy Guided by Vision and Force Observations
Kang, J. H., Joshi, S., Huang, R., & Gupta, S. K. (2025) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2025.3553689 · Cited by 19 · see citing works
“It has demonstrated remarkable potential across various real-world robotic manipulation tasks [9], [10], [11], [12], [13], [14], [15].”
Iterative residual policy: For goal-conditioned dynamic manipulation of deformable objects
Chi, C., Burchfiel, B., Cousineau, E. A., Feng, S., & Song, S. (2023) · The International Journal of Robotics Research · doi:10.1177/02783649231201201 · Cited by 17 · see citing works
“Goal-conditioned manipulation is particularly difficult for deformable objects due to their high degree-of-freedom and under-actuation [1, 21, 18, 22, 25, 26, 9, 24].”
BackgroundInterleaving Planning and Control for Deformable Object Manipulation
McConachie, D., Ruan, M., & Berenson, D. (2019) · Springer proceedings in advanced robotics · doi:10.1007/978-3-030-28619-4_68 · Cited by 15 · see citing works
Learning Rope Manipulation Policies Using Dense Object Descriptors Trained on Synthetic Depth DataPreprint
Sundaresan, P., Grannen, J., Thananjeyan, B., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2003.01835 · Cited by 16 · see citing works
“There is also extensive literature on deformable manipulation [18, 25, 27, 36, 40].”
BackgroundRobotic Actuation and Control of a Catheter for Structural Intervention Cardiology
Zhang, X., Palumbo, M. C., Perico, F., Magro, M., Fortuna, A., Magni, T., Votta, E., Segato, A., & Momi, E. D. (2022) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros47612.2022.9981676 · Cited by 17 · see citing works
“This enabled the robot to perform the task on its own, even under different conditions [12].”
MethodFeedback motion planning for liquid pouring using supervised learning
Pan, Z., & Manocha, D. (2017) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2017.8202300 · Cited by 15 · see citing works
“The extension of conventional motion planning algorithms to the manipulation of non-rigid objects has been addressed in the context of virtual suturing [21], cloth folding [22], and surgical simulation [23].”
BackgroundLearning Dense Visual Correspondences in Simulation to Smooth and Fold Real FabricsPreprint
Ganapathi, A., Sundaresan, P., Thananjeyan, B., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2003.12698 · Cited by 15 · see citing works
“[11, 12] show impressive human-robot fabric manipulation under the assumption that fabric has already been grasped, and Schulman et al. [33] demonstrate deformable object manipulation while requiring task-specific kinesthetic demonstrations.”
Multi-Stage Suture Detection for Robot Assisted Anastomosis Based on Deep Learning
Hu, Y., Gu, Y., Yang, J., & Yang, G. (2018) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2018.8461131 · Cited by 15 · see citing works
“perception to obtain 3D points cloud of the thread, probabilistic model [4], deformable registration [5] and graph model [6] have been applied to these recognition processes.”
MethodLee, A. X., Goldstein, M. A., Barratt, S. T., & Abbeel, P. (2015) · 2015 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra.2015.7139289 · Cited by 12 · see citing works
“Abstract— Recent work [1], [2], [3] has shown promising results in learning from demonstrations for the manipulation of deformable objects.”
BackgroundInfluentialSelf-citationSimultaneous End-User Programming of Goals and Actions for Robotic Shelf Organization
Liang, Y., Pellier, D., Fiorino, H., et al. (2018) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2018.8593518 · Cited by 12 · see citing works
“A majority of the work focuses on directly learning a policy or modeling higher level actions from lower level control signals [11], [12], [13], [14], while some explore learning task goals or task structure represented in various ways [15], [16], [17], [18], [19].”
BackgroundA novel cable-grasping planner for manipulator based on the operation surface
Wu, W. H., Zhu, Y., Zheng, X., & Guo, Y. H. (2021) · Robotics and Computer-Integrated Manufacturing · doi:10.1016/j.rcim.2021.102252 · Cited by 11 · see citing works
Generalized Cylinders for Learning, Reproduction, Generalization, and Refinement of Robot Skills
Ahmadzadeh, S. R., Rana, M. A., & Chernova, S. (2017) · doi:10.15607/rss.2017.xiii.021 · Cited by 12 · see citing works
“[25] demonstrated the usefulness of nonrigid registration in LfD by employing it for autonomous knot tying.”
BackgroundAutomated Planning Techniques for Robot Manipulation Tasks Involving Articulated Objects
Capitanelli, A., Maratea, M., Mastrogiovanni, F., et al. (2017) · Lecture notes in computer science · doi:10.1007/978-3-319-70169-1_36 · Cited by 11 · see citing works
Seita, D., Florence, P., Tompson, J., et al. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2012.03385 · Cited by 11 · see citing works
“In recent years, learning-based approaches for manipulation have grown in popularity to improve the generalization of knot-tying and also to manipulate rope towards general target configurations [52], [46], [49], [57], [64], [61].”
BackgroundNIFT: Neural Interaction Field and Template for Object Manipulation
Huang, Z., Xu, J., Dai, S., et al. (2023) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48891.2023.10160666 · Cited by 10 · see citing works
“For example, pose estimation [28, 35, 39] is the key for known objects, while primitive-based templatematching [13, 27, 34] can be more robust to shape and pose changes.”
BackgroundCross-Scene Suture Thread Parsing for Robot Assisted Anastomosis based on Joint Feature Learning
Gu, Y., Hu, Y., Zhang, L., Yang, J., & Yang, G. (2018) · IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) · doi:10.1109/iros.2018.8593622 · Cited by 9 · see citing works
“The typical examples include rope detection [10], [11], [12], [13] and thread detection [4], [1], [2], [3], [14].”
BackgroundMethodEfficient programming of manipulation tasks by demonstration and adaptation
Elliott, S., Toris, R., & Çakmak, M. (2017) · IEEE RO-MAN · doi:10.1109/roman.2017.8172448 · Cited by 9 · see citing works
Computational Modeling: Human Dynamic Model
Liu, L., Cooper, J. L., & Ballard, D. H. (2021) · Frontiers in Neurorobotics · doi:10.3389/fnbot.2021.723428 · Cited by 7 · see citing works
Shape-Based Transfer of Generic Skills
Thompson, S., Kaelbling, L. P., & Lozano‐Pérez, T. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9560894 · Cited by 10 · see citing works
“Schulman et al. [5] and Lee et al. [6] describe a method for adapting task demonstrations across objects of varying geometry by performing non-rigid registration between a known scene and a novel one.”
MethodAutomated Planning Encodings for the Manipulation of Articulated Objects in 3D with Gravity
Bertolucci, R., Capitanelli, A., Maratea, M., et al. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-35166-3_10 · Cited by 9 · see citing works
Learning From Observation of Tabletop Knotting Using a Simple Task Model
Takizawa, M., Yao, Z., Onda, H., Kudoh, S., & Suehiro, T. (2019) · IEEE/SICE International Symposium on System Integration (SII) · doi:10.1109/sii.2019.8700429 · Cited by 9 · see citing works
“of a person, who moved to teach a trajectory that fitted the situation [15].”
BackgroundManipulation of Articulated Objects Using Dual-arm Robots via Answer Set Programming
Bertolucci, R., Capitanelli, A., Dodaro, C., et al. (2020) · Theory and Practice of Logic Programming · doi:10.1017/s1471068420000459 · Cited by 8 · see citing works
A Simulation-Based Grasp Planner for Enabling Robotic Grasping during Composite Sheet Layup
Manyar, O. M., Desai, J., Deogaonkar, N., Joesph, R. J., Malhan, R. K., McNulty, Z., Wang, B., Barbič, J., & Gupta, S. K. (2021) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra48506.2021.9560939 · Cited by 8 · see citing works
Control flow for robust one-shot robot programming using entity-based resources
Orendt, E. M., & Henrich, D. (2017) · International Conference on Advanced Robotics (ICAR) · doi:10.1109/icar.2017.8023498 · Cited by 7 · see citing works
“Although there is just a small number of One-Shot PbD approaches, but they all match in the benefit of intuitive use [6], [5], [8], [9], [10].”
BackgroundOnline Grasp Plan Refinement for Reducing Defects During Robotic Layup of Composite Prepreg Sheets
Malhan, R. K., Joseph, R. J., Shembekar, A. V., Kabir, A. M., Bhatt, P. M., & Gupta, S. K. (2020) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra40945.2020.9196876 · Cited by 8 · see citing works
Optimal Cosserat-based deformation control for robotic manipulation of linear objects
Artinian, A., Quentin, H., Amar, F. B., et al. (2023) · IEEE/ASME International Conference on Advanced Intelligent Mechatronics · doi:10.1109/aim46323.2023.10196216 · Cited by 6 · see citing works
“These approaches have the advantage of not relying on a specific model but they are often hard to generalize from a small amount of training data [13], [14].”
BackgroundTowards Occlusion-Aware Pose Estimation of Surgical Suturing Threads
Gu, Y., Yang, J., & Yang, G. (2022) · IEEE Transactions on Biomedical Engineering · doi:10.1109/tbme.2022.3198402 · Cited by 6 · see citing works
“whether a predefined model is used, they can be categorized into explicit methods [3], [4], [5], [6], [7], [8], [9], [10], [11] and implicit methods [12], [13], [14], [15], [16], [17], [18], [19],”
MethodLearning-based Feedback Controller for Deformable Object ManipulationPreprint
Jia, B., Hu, Z., Pan, Z., et al. (2018) · arXiv (Cornell University) · doi:10.48550/arxiv.1806.09618 · Cited by 6 · see citing works
“accomplish the task. The approaches in the final group do not explicitly model the deformation parameters of the object. Instead, they use vision or learning methods to accomplish tasks directly [1], [10], [14], [22]–[26]. These methods focus on high-level policies but cannot achieve accurate operations; their success…”
BackgroundInfluentialRobots of the Lost Arc: Learning to Dynamically Manipulate Fixed-Endpoint Ropes and Cables.Preprint
Zhang, H., Ichnowski, J., Seita, D., Wang, J., & Goldberg, K. Y. (2020) · arXiv (Cornell University) · Cited by 4 · see citing works
Wang, Z., & Qureshi, A. H. (2025) · IEEE Robotics and Automation Letters · doi:10.1109/lra.2025.3539910 · Cited by 5 · see citing works
“In recent years, learning-based methods [9], [10], [11], [12], [13], [14], [15], [16], [17] have grown rapidly and demonstrated impressive performance and generalization abilities.”
Perception of Deformable Objects and Compliant Manipulation for Service Robots
Stückler, J., & Behnke, S. (2015) · Soft Robotics · doi:10.1007/978-3-662-44506-8_7 · Cited by 5 · see citing works
“In contrast to the method in [8,9], we do not assume the estimated displacement field to be valid at any pose on the motion trajectory.”
BackgroundMethodBeyond lowest-warping cost action selection in trajectory transfer
Hadfield-Menell, D., Lee, A. X., Finn, C., et al. (2015) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2015.7139644 · Cited by 5 · see citing works
“Recent approaches leverage TPS to perform trajectory transfer [2].”
BackgroundMethodInfluentialSelf-citationDesign and implementation of a multi-DOF hand and basic motions for rope-knotting tasks
Takizawa, M., Kudoh, S., & Suehiro, T. (2022) · Advanced Robotics · doi:10.1080/01691864.2022.2070866 · Cited by 5 · see citing works
Movement primitives with multiple phase parameters
Ewerton, M., Maeda, G. J., Neumann, G. M., et al. (2016) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra.2016.7487134 · Cited by 5 · see citing works
“Learning from human demonstrations has been playing an increasingly important role in robotics, especially since robots are getting better at adapting demonstrated movements to new situations and demands [1], [2], [3].”
BackgroundTrajectory Adaptation of Robot Arms for Head-Pose Dependent Assistive Tasks.
Rahmatizadeh, R., Abolghasemi, P., Bölöni, L. L., Jabalameli, A., & Behal, A. (2016) · Journal of International Crisis and Risk Communication Research · Cited by 5 · see citing works
Visualizing Crocheting Overlay
Cho, S., Nguyen, L. L., Wen, Y., et al. (2022) · Creativity and Cognition · doi:10.1145/3527927.3533024 · Cited by 5 · see citing works
“, understanding the ends of the rope and its trajectory [18]).”
Background
- Authors
- Schulman, J., Ho, J., Lee, C., & Abbeel, P.
- Affiliations
- University of California, Berkeley (US)
- Corresponding author
- Schulman, John
- Venue
- Springer Tracts in Advanced Robotics
- Publisher
- Springer International Publishing
- Published
- 2016
- Pages
- 339-354
- ISSN
- 1610-742X (ISSN-L)
- 1610-7438
- OpenAlex
- W2472053385
- Cited by
- 133
- References
- 31
- Citations by year
- 2026: 1 · 2025: 4 · 2024: 4 · 2023: 9 · 2022: 14 · 2021: 15 · 2020: 18 · 2019: 12 · 2018: 18 · 2017: 13 · 2016: 6 · 2015: 13 · 2014: 6
- FWCI
- 38.92
- Type
- book chapter
- Language
- en
- Open access
- closed
- Locations
- Springer tracts in advanced robotics (publishedVersion)
- Topics
- Physical Sciences › Engineering › Control and Systems Engineering › Robot Manipulation and Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Physical Sciences › Engineering › Biomedical Engineering › Soft Robotics and Applications
- Keywords
- imitation learning, non-rigid registration, robotic manipulation, multi-step tasks, PR2 robot, knot tying, rope manipulation
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 2026-10-06.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-06.
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
