Data-Driven Robotic Manipulation of Cloth-like Deformable Objects: The Present, Challenges and Future Prospects

May combine two claims.

Only Lee et al. (2021) collect real-world data; most learning-based methods operate in simulated environments, where some apply domain randomisation to achieve Sim2Real.

“Whilst only Lee et al. (2021) [229] collect real-world data, most learning-based methods operate in simulated environments, where some apply domain randomisation [81] to achieve Sim2Real [38,227,233,237].”

• cites [229] Lee, R., Ward, D., Dasagi, V., Cosgun, A., Leitner, J., and Corke, P. (2021, January 8–11). Learning arbitrary-goal fabric folding with one hour of real robot experience. Proceedings of the Conference on Robot Learning (CoRL), London, UK. sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [38] Hoque, 2022, Visuospatial foresight for physical sequential fabric manipulation, Auton. Robot. sourceNot checked — no full text of the source
• cites [227] Ha, H., and Song, S. (2022, January 15–18). Flingbot: The unreasonable effectiveness of dynamic manipulation for cloth unfolding. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourceDisputed

  Split into 2 atoms — 1 confirmed, 0 partially supported, 1 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured
• cites [233] Yan, W., Vangipuram, A., Abbeel, P., and Pinto, L. (2021, January 8–11). Learning predictive representations for deformable objects using contrastive estimation. Proceedings of the Conference on Robot Learning (CoRL), London, UK. sourceSupported
• cites [237] Weng, T., Bajracharya, S.M., Wang, Y., Agrawal, K., and Held, D. (2022, January 15–18). Fabricflownet: Bimanual cloth manipulation with a flow-based policy. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Robots still struggle to solve high-level tasks such as locomotion, autonomous driving and dexterous manipulation, even though low-level controllers and motion planning algorithms are robust enough to drive joints and end-effectors to desired configurations.

“Even though low-level controllers and motion planning algorithms are robust enough to drive the joints and end-effectors to reach their desired configuration, robots still struggle to solve high-level tasks, such as locomotion [3], autonomous driving [4] and dexterous manipulation [5].”

• cites [3] Li, Z., Cheng, X., Peng, X.B., Abbeel, P., Levine, S., Berseth, G., and Sreenath, K. (June, January 20). Reinforcement learning for robust parameterized locomotion control of bipedal robots. Proceedings of the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China. sourceNot checked — no full text of the source
• cites [4] Kiran, 2021, Deep reinforcement learning for autonomous driving: A survey, IEEE Trans. Intell. Transp. Syst. sourceSupported
• cites [5] Rajeswaran, A., Kumar, V., Gupta, A., Vezzani, G., Schulman, J., Todorov, E., and Levine, S. (2018, January 26–30). Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations. Proceedings of the Robotics: Science and Systems (RSS), Pittsburgh, PA, USA. sourcePartly supported

  Split into 2 atoms — 1 confirmed, 0 partially supported, 0 disputed, 0 misattributed, 1 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Representation learning aims to learn better latent state estimation to overcome data efficiency and mode collapse, leading to robust and interpretable control systems.

“Representation learning (Section 5.2) [64] aims to learn better latent state estimation to overcome both issues, leading to robust and interpretable control systems.”

• cites [64] Stooke, A., Lee, K., Abbeel, P., and Laskin, M. (2021, January 18–24). Decoupling representation learning from reinforcement learning. Proceedings of the International Conference on Machine Learning (ICML), online. sourcePartly supported

  The Introduction (p. 1) describes learning visual features independently of rewards to improve RL learning performance, and the Conclusion (p. 8) reports maintained or improved control performance. These passages support the representation-learning and performance portions of the claim, but not the assertions about overcoming mode collapse or producing robust and interpretable control systems. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Neural networks are convenient and effective at combining multi-modal sensory inputs, as well as integrating passive and interactive perception.

“NNs are also convenient and effective at combining multi-modal sensory inputs [55], as well as integrating passive and interactive perception [56,57,58].”

• cites [55] Erickson, Z., Clegg, A., Yu, W., Turk, G., Liu, C.K., and Kemp, C.C. (June, January 29). What does the person feel? Learning to infer applied forces during robot-assisted dressing. Proceedings of the 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore. sourceNot checked — no full text of the source
• cites [56] Gao, Y., Hendricks, L.A., Kuchenbecker, K.J., and Darrell, T. (2016, January 16–21). Deep learning for tactile understanding from visual and haptic data. Proceedings of the 2016 IEEE International Conference on Robotics and Automation (ICRA), Stockholm, Sweden. sourceNot checked — no full text of the source
• cites [57] Lee, M.A., Zhu, Y., Srinivasan, K., Shah, P., Savarese, S., Fei-Fei, L., Garg, A., and Bohg, J. (2019, January 20–24). Making sense of vision and touch: Self-supervised learning of multimodal representations for contact-rich tasks. Proceedings of the 2019 International Conference on Robotics and Automation (ICRA), Montreal, QC, Canada. sourceSupported
• cites [58] Sung, J., Lenz, I., and Saxena, A. (June, January 29). Deep multimodal embedding: Manipulating novel objects with point-clouds, language and trajectories. Proceedings of the 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Approaches for modelling dynamics of DOs include mass-spring, particle-based, position-based, constitutive, and neural network systems.

“Approaches for modelling dynamics of DOs include mass-spring, particle-based, position-based, constitutive, and neural network systems [11].”

• cites [11] Guler, 2020, Modeling of deformable objects for robotic manipulation: A tutorial and review, Front. Robot. AI sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

In robotics, skills are often modelled using the Option framework.

“In robotics, skills are often modelled using the Option framework [78].”

• cites [78] Sutton, 1999, Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning, Artif. Intell. sourcePartly supported

  Section 2 (Options; page unavailable) defines options as temporally extended actions and gives robot docking and robot-arm control as examples. This supports modelling robotic skills as options, but the examples do not establish that robotics skills are often modelled this way. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Adversarial inverse reinforcement learning methods attempt to automate reward engineering with demonstration data while preserving similar computational complexity as their RL counterparts.

“Adversarial inverse reinforcement learning (Adversarial IRL) methods [66] also attempt to automate reward engineering with demonstration data while preserving similar computational complexity as their RL counterparts [67,68].”

• cites [66] Orsini, 2021, What matters for adversarial imitation learning?, Adv. Neural Inf. Process. Syst. sourcePartly supported

  The Introduction and Appendix B describe adversarial imitation learning as learning from expert demonstrations when reward functions are difficult to design, supporting the reward-engineering part of the claim. The passages found do not establish that its computational complexity is similar to that of its RL counterparts. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [67] Ho, 2016, Generative adversarial imitation learning, Adv. Neural Inf. Process. Syst. sourcePartly supported

  The Introduction (p. 1) describes learning from expert demonstrations without a reinforcement signal, and Section 5 (p. 6) presents an adversarial imitation method that uses a discriminator as a local cost function. The Discussion (p. 8) says its environment-sample requirements were comparable to those of TRPO, but does not establish similar overall computational complexity to RL methods. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [68] Rafailov, 2021, Visual adversarial imitation learning using variational models, Adv. Neural Inf. Process. Syst. sourcePartly supported

  The Introduction (p. 1) says that learning from expert demonstrations can avoid reward specification and that adversarial imitation learning uses a learned classifier as an RL reward signal. Neither that passage nor the Related Work discussion (p. 9) establishes that these methods preserve computational complexity similar to their RL counterparts. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Meta-learning attempts to dynamically adjust hyper-parameters regarding its experience in the target domain and training status.

“To automate this process, meta-learning attempts to dynamically adjust the hyper-parameters regarding its experience in the target domain and training status [65].”

• cites [65] Pong, V.H., Nair, A.V., Smith, L.M., Huang, C., and Levine, S. (2022, January 17–23). Offline meta-reinforcement learning with online self-supervision. Proceedings of the International Conference on Machine Learning, Baltimore, MA, USA. sourcePartly supported

  The Introduction (p. 1) describes meta-RL policies adapting to new tasks using collected experience. But the paper does not say that meta-learning dynamically adjusts hyperparameters according to target-domain experience or training status; C.3 (pp. 18–19) instead lists hyperparameters used for the experiments. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

IL using real-world demonstration data has become one of the most popular and robust control algorithms in research and industry.

“In this era, IL using real-world demonstration data has become one of the most popular and robust control algorithms in research and industry [59,60].”

• cites [59] Pan, Y., Cheng, C.A., Saigol, K., Lee, K., Yan, X., Theodorou, E.A., and Boots, B. (2017). Agile Autonomous Driving using End-to-End Deep Imitation Learning. arXiv. sourcePartly supported

  The Introduction and Data Collection sections describe imitation learning from expert demonstrations collected on a real-world driving track. They support its use as a control approach in this research study, but do not establish that it is among the most popular or robust algorithms across research and industry. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [60] Codevilla, F., Müller, M., López, A., Koltun, V., and Dosovitskiy, A. (2018, January 21–25). End-to-end driving via conditional imitation learning. Proceedings of the 2018 International Conference on Robotics and Automation (ICRA), Brisbane, Australia. sourcePartly supported

  The Introduction (p. 1) describes imitation learning from human driving demonstrations as a promising approach receiving renewed interest, and Related Work (p. 2) lists applications across several tasks. The paper does not establish that it is among the most popular control algorithms in research and industry; Training Data Distribution (p. 4) also cautions that policies trained only on natural demonstrations are typically unstable, rather than broadly robust. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

The first challenge in DO manipulation and simulation is to select an appropriate representation for the target object, where the major criteria are expressiveness, accuracy and flexibility of the representation.

“The first challenge in DO manipulation and simulation is to select an appropriate representation for the target object, where the major criteria are expressiveness, accuracy and flexibility [11] of the representation.”

• cites [11] Guler, 2020, Modeling of deformable objects for robotic manipulation: A tutorial and review, Front. Robot. AI sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

POMDPs more accurately describe real-world applications of robotic systems and are a major tool for training modern DRL and DIL end-to-end policy controllers.

“POMDPs more accurately describe real-world applications of robotic systems [7] and are a major tool for training modern DRL and DIL end-to-end policy controllers.”

• cites [7] Ibarz, 2021, How to train your robot with deep reinforcement learning: Lessons we have learned, Int. J. Robot. Res. sourcePartly supported

  Section 4.8 (p. 13) says real robotic systems have latency and unobservable intervening states that violate an assumption of the MDP formulation. This supports the cited clause’s concern about modeling real-world robots, but the paper does not say that POMDPs describe them more accurately. The citation is attached to that clause, not to the subsequent claim about training DRL and DIL controllers. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

May combine two claims.

Simulation-based approaches provide a cheap and safe development environment for RMSes, but there are challenges to deploying simulation-trained systems into real-world settings.

“Simulation-based approaches provide a cheap and safe development environment for RMSes, but there are challenges to deploying simulation-trained systems into real-world settings [62].”

• cites [62] Zhao, W., Queralta, J.P., and Westerlund, T. (2020, January 1–4). Sim-to-real transfer in deep reinforcement learning for robotics: A survey. Proceedings of the 2020 IEEE Symposium Series on Computational Intelligence (SSCI), Canberra, Australia. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Gaussian distributions are used in PoMPs and SEDs for better generalisation behaviour.

“Gaussian distributions are used in PoMPs [114] and SEDs [115] for better generalisation behaviour.”

• cites [114] Paraschos, 2013, Probabilistic movement primitives, Adv. Neural Inf. Process. Syst.Supported
• cites [115] Billard, 2011, Learning stable nonlinear dynamical systems with gaussian mixture models, IEEE Trans. Robot. sourceNot checked — no full text of the source

May combine two claims.

Discrete-action value-based methods such as Q-learning have been applied in cloth shaping and knot-untying tasks, while continuous-action methods such as DDPG and SAC have mainly been adopted in cloth-shaping tasks.

“Discrete-action value-based methods Q-learning [123] have been applied in cloth shaping (Section 6.1.2) and knot-untying (Section 6.2.2) tasks, while continuous-action methods, such as DDPG [124] and SAC [125], have mainly been adopted in cloth-shaping tasks (Section 6.1.2).”

• cites [123] Watkins, 1992, Q-learning, Mach. Learn. sourceSupported
• cites [124] Lillicrap, T.P., Hunt, J.J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. (2016, January 2–4). Continuous control with deep reinforcement learning. Proceedings of the International Conference on Learning Representations (ICLR), San Juan, Puerto Rico. sourceSupported
• cites [125] Haarnoja, T., Zhou, A., Abbeel, P., and Levine, S. (2018, January 10–15). Soft actor–critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. Proceedings of the International Conference on Machine Learning (ICML), Stockholm, Sweden.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Q-learning is an off-policy method that solves tabular Markov decision process problems.

“Q-learning [123] is an off-policy method that solves tabular MDP problems.”

• cites [123] Watkins, 1992, Q-learning, Mach. Learn. sourceSupported

May combine two claims.

Deep reinforcement learning is data-hungry and trained through trial and error, requiring human involvement for constantly resetting the environment and preventing unsafe actions.

“First, DRL is data-hungry and trained through trial and error which requires human involvement for constantly resetting the environment and preventing unsafe actions [7].”

• cites [7] Ibarz, 2021, How to train your robot with deep reinforcement learning: Lessons we have learned, Int. J. Robot. Res. sourcePartly supported

  Section 4.7 (p. 13) says deep RL requires substantial data and that, in most robot-learning experiments to date, a human operator has reset the scene and stopped unsafe operation. The claim overstates this as a constant requirement: the same section says robots need to run without human supervision to scale data collection. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Expert-script demonstrators have seen substantial use in cloth-shaping systems and knot-tying systems.

“Such demonstrators have seen substantial use in cloth-shaping systems [21,29,38] and knot-tying systems [92].”

• cites [21] Seita, D., Jamali, N., Laskey, M., Tanwani, A.K., Berenstein, R., Baskaran, P., Iba, S., Canny, J., and Goldberg, K. (2019, January 6–10). Deep transfer learning of pick points on fabric for robot bed-making. Proceedings of the The International Symposium of Robotics Research, Hanoi, Vietnam.Not checked yet

  Not checked in the free run.

• cites [29] Lin, X., Wang, Y., Olkin, J., and Held, D. (2021, January 8–11). Softgym: Benchmarking deep reinforcement learning for deformable object manipulation. Proceedings of the Conference on Robot Learning (CoRL), London, UK. sourceCould not verify

  The citation is attached to the knot-tying portion of the sentence. Our copy describes rope-configuration tasks (4.2 Tasks, p. 4) and mentions demonstrations for cloth manipulation (2 Related Works, p. 2), but neither passage addresses expert-script demonstrators in knot-tying systems. Searches for knot tying, scripted experts, and demonstrators found no supporting passage. (LLM, conf=0.76)

  Experimental — accuracy not yet measured
• cites [38] Hoque, 2022, Visuospatial foresight for physical sequential fabric manipulation, Auton. Robot. sourceNot checked — no full text of the source
• cites [92] Suzuki, T., Ebihara, Y., and Shintani, K. (2005, January 18–21). Dynamic analysis of casting and winding with hyper-flexible manipulator. Proceedings of the ICAR’05, 12th International Conference on Advanced Robotics, Seattle, WA, USA. sourceNot checked — no full text of the source

May combine two claims.

High-dimensional or continuous state settings exacerbate the challenge of exploration and further affect data efficiency.

“(4) high-dimensional/continuous state settings exacerbate the challenge of exploration and further affect data efficiency.”

• cites [7] Ibarz, 2021, How to train your robot with deep reinforcement learning: Lessons we have learned, Int. J. Robot. Res. sourcePartly supported

  Section 4.2 (p. 8) says high-dimensional observations increase training and sample complexity, supporting the data-efficiency concern. Section 4.4 (p. 10) says exploration is particularly difficult with sparse rewards, but does not say that high-dimensional or continuous states exacerbate exploration or that this is what reduces data efficiency. (LLM, conf=0.84)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DAgger by Coaching and Data as Demonstrator have been proposed to improve DAgger's learning performance, but they have not been applied in the cloth-like deformable object domain.

“DAgger by Coaching [102] and Data as Demonstrator (DaD) [103] have been proposed to improve learning performance of DAgger, but they have not been applied in the CDO domain.”

• cites [102] He, 2012, Imitation learning by coaching, Adv. Neural Inf. Process. Syst.Not checked — no full text of the source
• cites [103] Venkatraman, A., Hebert, M., and Bagnell, J.A. (2015, January 25–30). Improving multi-step prediction of learned time series models. Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, TX, USA. sourcePartly supported

  The DATA AS DEMONSTRATOR Algorithm section (p. 2) introduces DaD, and the Conclusion (p. 6) reports improved multi-step prediction. But the reported comparison is with a single-step prediction baseline, not DAgger; the Reduction to imitation learning section (p. 3) instead says DaD draws on DAgger’s theoretical guarantees. The paper does not establish the broader claim that DaD had not been applied in the cloth-like deformable object domain. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

IL suffers from several issues: demonstration data can be suboptimal, contain noise, and have a different domain compared to the application domain.

“The major bottlenecks are demonstration data that can (1) be suboptimal, (2) contain noise and (3) have a different domain (different MDP/POMDP) compared to the application domain [8].”

• cites [8] Osa, 2018, An algorithmic perspective on imitation learning, Found. Trends® Robot. sourcePartly supported

  Section 1.3 describes a mismatch between the distributions of expert demonstrations and learner reproductions, and §1.4 says differences in embodiment and observability can require adaptation. The copy does not establish that this means different MDPs or POMDPs. It gives an example of noisy observations and control inputs, but does not clearly identify noisy or suboptimal demonstration data as major bottlenecks; §1.3 instead says imitation learning often assumes optimal or at least good demonstrations. Page numbers were not available in the indexed passages. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Transporter Net solves vision-based pick-and-place behaviour cloning using neural networks.

“Zeng et al. (2021) [104] propose Transporter Net to solve vision-based pick-and-place BC using NNs.”

• cites [104] Zeng, A., Florence, P., Tompson, J., Welker, S., Chien, J., Attarian, M., Armstrong, T., Krasin, I., Duong, D., and Sindhwani, V. (2021, January 8–11). Transporter networks: Rearranging the visual world for robotic manipulation. Proceedings of the Conference on Robot Learning (CoRL), London, UK. sourceSupported

May combine two claims.

Model-based reinforcement learning is generally more data-efficient than model-free reinforcement learning and generalises better with large and diverse data.

“MBRL is generally more data-efficient than MFRL algorithms and it generalises better with large and diverse data [7].”

• cites [7] Ibarz, 2021, How to train your robot with deep reinforcement learning: Lessons we have learned, Int. J. Robot. Res. sourceSupported

May combine two claims.

Deep reinforcement learning encoders suffer from observation overfitting, so data augmentation can improve data efficiency and generalisation ability of a DRL agent.

“In general, DRL encoders suffer from observation overfitting [160,161], so DA can also be used to improve data efficiency and generalisation ability of a DRL agent [141,142].”

• cites [160] Song, X., Jiang, Y., Tu, S., Du, Y., and Neyshabur, B. (2020, January 26–30). Observational Overfitting in Reinforcement Learning. Proceedings of the International Conference on Learning Representations (ICLR), online. sourceNot checked — no full text of the source
• cites [161] Packer, C., Gao, K., Kos, J., Krähenbühl, P., Koltun, V., and Song, D. (2018). Assessing generalization in deep reinforcement learning. arXiv. sourcePartly supported

  The Introduction (p. 1) says deep RL can overfit to a specific environment and be sensitive to environmental changes. But the paper studies changes in system dynamics (Introduction, p. 2); it does not establish that DRL encoders suffer from observation overfitting. The data-augmentation claim is attached to different citations in the citing sentence, so it is not assessed against this source. (LLM, conf=0.84)

  Experimental — accuracy not yet measured
• cites [141] Laskin, 2020, Reinforcement learning with augmented data, Adv. Neural Inf. Process. Syst. sourceSupported
• cites [142] Hansen, N., and Wang, X. (June, January 30). Generalization in reinforcement learning by soft data augmentation. Proceedings of the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

HER results in vast improvement compared with its base algorithm in many hard environments.

“It was shown experimentally that HER results in vast improvement compared with its base algorithm in many hard environments.”

• cites [182] Andrychowicz, 2017, Hindsight experience replay, Adv. Neural Inf. Process. Syst. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Double Deep Q-learning further reduces bias by choosing the next action using the training Q function and evaluating using the target one.

“Double Deep Q-learning (DDQN) [127] further reduces the bias by choosing the next action using the training Q function and evaluating using the target one.”

• cites [127] Van Hasselt, H., Guez, A., and Silver, D. (2016, January 12–17). Deep reinforcement learning with double q-learning. Proceedings of the AAAI Conference on Artificial Intelligence, Phoenix, AZ, USA. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Twin Delayed DDPG mitigates DDPG overestimation by introducing clipped double Q-learning, delayed policy updates and target policy smoothing.

“Twin Delayed DDPG (TD3) [128] mitigates this by introducing clipped double Q-learning, delayed policy updates and target policy smoothing.”

• cites [128] Fujimoto, S., Hoof, H., and Meger, D. (2018, January 10–15). Addressing function approximation error in actor–critic methods. Proceedings of the International Conference on Machine Learning (ICML), Stockholm, Sweden. sourceSupported

May combine two claims.

Deep reinforcement learning using data augmentation can suffer from unstable training and poor performance.

“It has been shown that DRL using DA can suffer from unstable training and poor performance [141].”

• cites [141] Laskin, 2020, Reinforcement learning with augmented data, Adv. Neural Inf. Process. Syst. sourcePartly supported

  The results table in Section 5.2 (p. 7) shows that some data augmentations produced lower scores than the unaugmented PPO baseline, supporting the possibility of poor performance. The passages found do not establish that data augmentation causes unstable training. (LLM, conf=0.80)

  Experimental — accuracy not yet measured

May combine two claims.

PlaNet has been applied in cloth-shaping literature but has so far shown poor performance, likely due to blurry observation prediction.

“PlaNet has been applied in the cloth-shaping literature (Section 6.1.2), but it has so far shown poor performance, likely due to the blurry observation prediction [38,179].”

• cites [38] Hoque, 2022, Visuospatial foresight for physical sequential fabric manipulation, Auton. Robot. sourceNot checked — no full text of the source
• cites [179] Lin, X., Wang, Y., Huang, Z., and Held, D. (2022, January 15–18). Learning visible connectivity dynamics for cloth smoothing. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourcePartly supported

  Related Work (section 2; page unavailable) says predicted cloth images from pixel-based dynamics models are usually blurry, and Simulation Results (section 4.2; page unavailable) reports that VCD outperforms its baselines. But Experimental Setup (section 4.1; page unavailable) names VSF, CFM, and MVP—not PlaNet—as the baselines. The copy therefore supports the general concern about blurry predictions and weaker baseline performance, but not the PlaNet-specific claim. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

QfBC applies the BC loss to learn demonstration data to solve long-horizon, sparse-reward and goal-conditioned tasks.

“Building upon DDPG [124] and HER [182], Nair et al. (2018) [220] propose a method called Q-filtered Behaviour Cloning (QfBC) that also applies the BC loss to learn demonstration data to solve long-horizon, sparse-reward and goal-conditioned tasks.”

• cites [220] Nair, A., McGrew, B., Andrychowicz, M., Zaremba, W., and Abbeel, P. (2018, January 21–25). Overcoming exploration in reinforcement learning with demonstrations. Proceedings of the 2018 IEEE International Conference on Robotics and Automation (ICRA), Brisbane, Australia. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

In POMDP settings, SLAC improves upon SAC's objective by incorporating latent dynamic learning to learn soft value and policy functions efficiently.

“In POMDP settings, Stochastic Latent Actor-Critic (SLAC) [147] improves upon SAC’s objective by incorporating latent dynamic learning to learn soft value and policy functions efficiently.”

• cites [147] Lee, 2020, Stochastic latent actor–critic: Deep reinforcement learning with a latent variable model, Adv. Neural Inf. Process. Syst. sourceSupported

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based).

DQfD has outperformed its pure RL and IL counterparts on most of the Atari games.

“DQfN has outperformed its pure RL and IL counterparts on most of the Atari games.”

• cites [218] Hester, T., Vecerik, M., Pietquin, O., Lanctot, M., Schaul, T., Piot, B., Horgan, D., Quan, J., Sendonaris, A., and Osband, I. (2018, January 2–7). Deep q-learning from demonstrations. Proceedings of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA. sourcePartly supported

  The Results (p. 5) report that DQfD beat pure RL (PDD DQN) on 41 of 42 games during the first million steps and started ahead of pure imitation learning on 31 games. Those comparisons support an early-performance advantage on most games, but the citing sentence omits that time qualification and implies an overall advantage. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

May combine two claims.

Discretising the drag strategy by putting different orientations and pulling distances into bins simplifies the problem and leads to better training, but is less flexible during run time and harder to generalise to other tasks.

“For example, Lee et al. (2021) [229] discretise the drag strategy by putting different orientations and pulling distances into bins; even though such action space simplifies the problem and leads to better training, it is less flexible during run time and harder to generalise to other tasks.”

• cites [229] Lee, R., Ward, D., Dasagi, V., Cosgun, A., Leitner, J., and Corke, P. (2021, January 8–11). Learning arbitrary-goal fabric folding with one hour of real robot experience. Proceedings of the Conference on Robot Learning (CoRL), London, UK. sourcePartly supported

  Section 4.4 (p. 5) says discretising the action space reduces offline-training difficulties, and Section 5.6 (p. 8) says it reduces learning complexity. But Section 5.6 contradicts the claimed loss of runtime flexibility: the method uses finer angle and distance bins at test time without retraining. The paper does not establish that it is harder to generalise to other tasks. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Learning a latent representation loses the detailed information of the target cloth such as folds and wrinkles.

“Lin et al. (2022) [179] also argue that learning a latent representation loses the detailed information of the target cloth such as folds and wrinkles.”

• cites [179] Lin, X., Wang, Y., Huang, Z., and Held, D. (2022, January 15–18). Learning visible connectivity dynamics for cloth smoothing. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourceSupported

May combine two claims.

Mesh representation of the cloth can bring topological information into planning for grasping and flattening, and plays a significant role in keeping track of the state of the article, especially when there are occlusions in the visual input.

“Mesh representation of the cloth can bring topological information into planning for grasping and flattening, and it plays a significant role in keeping track of the state of the article, especially when there are occlusions in the visual input [179,235,267].”

• cites [179] Lin, X., Wang, Y., Huang, Z., and Held, D. (2022, January 15–18). Learning visible connectivity dynamics for cloth smoothing. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourceSupported
• cites [235] Arnold, S., Tanaka, D., and Yamazaki, K. (2021). Cloth Manipulation Planning on Basis of Mesh Representations with Incomplete Domain Knowledge and Voxel-to-Mesh Estimation. arXiv. sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [267] Huang, Z., Lin, X., and Held, D. (2022). Mesh-based Dynamics with Occlusion Reasoning for Cloth Manipulation. arXiv. sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Demonstration learning algorithms often used in robotics applications are DQfD, DDPGfD and Q-filtered BC.

“Demonstration learning (DemoL) algorithms often used in robotics applications are DQfD [218], DDPGfD [219] and Q-filtered BC [220].”

• cites [218] Hester, T., Vecerik, M., Pietquin, O., Lanctot, M., Schaul, T., Piot, B., Horgan, D., Quan, J., Sendonaris, A., and Osband, I. (2018, January 2–7). Deep q-learning from demonstrations. Proceedings of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA. sourcePartly supported

  The Abstract (p. 1) identifies DQfD as an algorithm that learns from demonstrations, supporting its inclusion in the list. The Discussion (p. 6) describes potential real-world settings, including autonomous vehicles, but the paper evaluates DQfD on Atari games (Experimental Setup, p. 4); it does not establish that DQfD is often used in robotics applications. DDPGfD and Q-filtered BC have separate citations and are not assessed against this source. (LLM, conf=0.86)

  Experimental — accuracy not yet measured
• cites [219] Vecerík, M., Hester, T., Scholz, J., Wang, F., Pietquin, O., Piot, B., Heess, N., Rothörl, T., Lampe, T., and Riedmiller, M.A. (2017). Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards. arXiv.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites [220] Nair, A., McGrew, B., Andrychowicz, M., Zaremba, W., and Abbeel, P. (2018, January 21–25). Overcoming exploration in reinforcement learning with demonstrations. Proceedings of the 2018 IEEE International Conference on Robotics and Automation (ICRA), Brisbane, Australia. sourceSupported

May combine two claims.

POMDPs introduce challenges related to accurate estimation of the environment state, and higher-quality representations can lead to better exploration.

“POMDPs introduce another layer of challenges related to accurate estimation of the environment state—higher-quality representations can lead to better exploration [209,210].”

• cites [209] Jiang, N., Krishnamurthy, A., Agarwal, A., Langford, J., and Schapire, R.E. (2017, January 6–11). Contextual decision processes with low bellman rank are pac-learnable. Proceedings of the International Conference on Machine Learning (ICML), Sydney, Australia. sourcePartly supported

  Section 2.1 (p. 4) describes POMDPs as having hidden states and explains that observation histories or sliding windows can serve as contexts to address partial observability. The Introduction (pp. 1–2) links rich observations and function approximation to the need for systematic exploration, but does not establish that higher-quality state representations themselves lead to better exploration. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [210] Machado, M.C., Bellemare, M.G., and Bowling, M. (2020, January 7–12). Count-based exploration with the successor representation. Proceedings of the AAAI Conference on Artificial Intelligence, New York, NY, USA. sourcePartly supported

  The Conclusion states that better representations can lead to faster exploration. The paper’s Preliminaries discuss difficulty learning accurate models in large state spaces, but the passages do not address POMDPs or the challenge of accurately estimating an environment’s state under partial observability. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Visible Connectivity Dynamics (VCD) applies mesh models on the visible part of the cloth to achieve more precise planning.

“Visible Connectivity Dynamics (VCD) [179] applies such mesh models on the visible part of the cloth to achieve more precise planning.”

• cites [179] Lin, X., Wang, Y., Huang, Z., and Held, D. (2022, January 15–18). Learning visible connectivity dynamics for cloth smoothing. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourceSupported

May combine two claims.

The reward and goal signals of some MSCF tasks, such as triangle folding and square folding, cannot be directly extracted from vision input due to self-occlusion, although they can be divided into smaller subtasks to adapt to vision signals.

“The reward and goal signals of some MSCF tasks, such as triangle folding and square folding, cannot be directly extracted from vision input due to the self-occlusion, although we can divide them into smaller subtasks to adapt to vision signals [237].”

• cites [237] Weng, T., Bajracharya, S.M., Wang, Y., Agrawal, K., and Held, D. (2022, January 15–18). Fabricflownet: Bimanual cloth manipulation with a flow-based policy. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

For cloth-flattening tasks, normalised improvement and normalised coverage are the most reliable metrics that can be automated both in simulation and reality; however, this metric does not explicitly include information on wrinkles on the cloth, which is crucial for evaluation when a trial is nearly successful.

“For cloth-flattening tasks, normalised improvement [179] and normalised coverage [38,179] are the most reliable metrics that can be automated both in simulation and reality; however, this metric does not explicitly include information on wrinkles on the cloth, which is crucial for evaluation when a trial is nearly successful.”

• cites [179] Lin, X., Wang, Y., Huang, Z., and Held, D. (2022, January 15–18). Learning visible connectivity dynamics for cloth smoothing. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured
• cites [38] Hoque, 2022, Visuospatial foresight for physical sequential fabric manipulation, Auton. Robot. sourceNot checked — no full text of the source

May combine two claims.

Robust velocity control is still an unsolved challenge in this domain, although few attempt such low-level velocity control to conduct dynamic manipulation.

“Robust velocity control is still an unsolved challenge in this domain [20,29,226], although few attempt such low-level velocity control to conduct dynamic manipulation [20,29,226,227,228].”

• cites [20] Matas, J., James, S., and Davison, A.J. (2018, January 29–31). Sim-to-real reinforcement learning for deformable object manipulation. Proceedings of the Conference on Robot Learning, Zurich, Switzerland. sourcePartly supported

  Section 5.1 (p. 6) describes a cloth-manipulation system controlled by end-effector and gripper velocities, and the Introduction (p. 1) says relatively little work had applied end-to-end learning to deformable objects. The paper does not establish that robust velocity control remains unsolved or characterize its tasks as dynamic manipulation, so those parts of the claim are unsupported. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [29] Lin, X., Wang, Y., Olkin, J., and Held, D. (2021, January 8–11). Softgym: Benchmarking deep reinforcement learning for deformable object manipulation. Proceedings of the Conference on Robot Learning (CoRL), London, UK. sourcePartly supported

  The Introduction (p. 1) describes deformable-object manipulation as a longstanding challenge. The Action Space section (p. 3) says SoftGym abstracts away low-level manipulation, but neither passage establishes that robust velocity control remains unsolved or that few researchers attempt it for dynamic manipulation. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [226] Jangir, R., Alenyà, G., and Torras, C. (August, January 31). Dynamic cloth manipulation with deep reinforcement learning. Proceedings of the 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France. sourcePartly supported

  The Introduction (p. 1) says most prior deformable-object work considered static manipulation, and State and Action (p. 4) says this study's policy specifies target manipulator velocities for dynamic cloth tasks. These passages support the claim that few approaches attempt this kind of velocity-based dynamic manipulation, but the paper does not state that robust velocity control remains unsolved. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [227] Ha, H., and Song, S. (2022, January 15–18). Flingbot: The unreasonable effectiveness of dynamic manipulation for cloth unfolding. Proceedings of the Conference on Robot Learning (CoRL), Auckland, New Zealand. sourcePartly supported

  Section 4.3 (p. 7) describes FlingBot-S as predicting fling speed for dynamic cloth manipulation, supporting the point that the paper attempts velocity selection. But the paper does not say robust low-level velocity control remains unsolved; Section 6.3 (p. 13) instead reports robust real-world unfolding across a wide range of fling speeds when given a good grasp. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [228] Hietala, J., Blanco-Mulero, D., Alcan, G., and Kyrki, V. (2021). Closing the Sim2Real Gap in Dynamic Cloth Manipulation. arXiv.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim. May combine two claims.

The system needs to detect success and different error states.

“In addition, the system needs to detect success and different error states [298,301,309,315,316,317].”

• cites [298] Yamazaki, K., Oya, R., Nagahama, K., Okada, K., and Inaba, M. (2014, January 13–15). Bottom dressing by a life-sized humanoid robot provided failure detection and recovery functions. Proceedings of the 2014 IEEE/SICE International Symposium on System Integration, Tokyo, Japan. sourceNot checked — no full text of the source
• cites [301] Chance, G., Camilleri, A., Winstone, B., Caleb-Solly, P., and Dogramadzi, S. (2016, January 6–29). An assistive robot to support dressing-strategies for planning and error handling. Proceedings of the 2016 6th IEEE International Conference on Biomedical Robotics and Biomechatronics (BioRob), Singapore. sourceNot checked — no full text of the source
• cites [309] Yamazaki, 2016, Bottom dressing by a dual-arm robot using a clothing state estimation based on dynamic shape changes, Int. J. Adv. Robot. Syst. sourceNot checked — no full text of the source
• cites [315] Yamazaki, K., Oya, R., Nagahama, K., and Inaba, M. (2013, January 15–17). A method of state recognition of dressing clothes based on dynamic state matching. Proceedings of the 2013 IEEE/SICE International Symposium on System Integration, Kobe, Japan. sourceNot checked — no full text of the source
• cites [316] Chance, 2017, A quantitative analysis of dressing dynamics for robotic dressing assistance, Front. Robot. AI sourcePartly supported

  The Methodology system setup section says the sensors are intended to detect dressing errors, and Predictive Models for error Detection distinguishes normal dressing from a snagging error. The paper does not establish detection of successful completion or distinguish among different error states; it tests one simulated error type. (LLM, conf=0.85)

  Experimental — accuracy not yet measured
• cites [317] Yu, W., Kapusta, A., Tan, J., Kemp, C.C., Turk, G., and Liu, C.K. (June, January 29). Haptic simulation for robot-assisted dressing. Proceedings of the 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore. sourceSupported

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

A robust RMS should be able to recognise the failing states and recover from these failures.

“A robust RMS should be able to recognise the failing states and recover from these failures [9,36].”

• cites [9] Maitin-Shepard, J., Cusumano-Towner, M., Lei, J., and Abbeel, P. (2010, January 3–7). Cloth grasp point detection based on multiple-view geometric cues with application to robotic towel folding. Proceedings of the 2010 IEEE International Conference on Robotics and Automation (ICRA), Anchorage, AK, USA. sourceNot checked — no full text of the source
• cites [36] Sundaresan, P., Grannen, J., Thananjeyan, B., Balakrishna, A., Ichnowski, J., Novoseller, E.R., Hwang, M., Laskey, M., Gonzalez, J.E., and Goldberg, K. (2021). Untangling Dense Non-Planar Knots by Learning Manipulation Features and Recovery Policies. arXiv. sourceSupported

May combine two claims.

Many real-world failures in cloth shaping are caused by grasping deficiencies of the system, including inaccurate grasp-point estimation, misgrasping the target point, grasping multiple layers if there is a fold, and rigid damage on the end-effector caused by a hard surface.

“Many real-world failures in cloth shaping are caused by grasping deficiencies of the system, including (1) inaccurate grasp-point estimation, (2) misgrasping the target point on the cloth, (3) grasping multiple layers of the fabric if there is a fold and (4) rigid damage on the end-effector caused by a hard surface.”

Probably not a substantial claim: may not pass S4 (based).

CDOs' many degrees of freedom introduce severe self-occlusion and complex state-action dynamics, which are significant obstacles to perception and manipulation systems.

“CDOs’ many degrees of freedom (DoF) introduce severe self-occlusion and complex state–action dynamics as significant obstacles to perception and manipulation systems.”

May combine two claims.

Classical control methods have been successfully applied to CDO manipulation, but they tend to be restricted to a narrow range of defined configurations.

“Classical control methods have been successfully applied to CDO manipulation [9,10], but they tend to be restricted to a narrow range of defined configurations.”

• cites [9] Maitin-Shepard, J., Cusumano-Towner, M., Lei, J., and Abbeel, P. (2010, January 3–7). Cloth grasp point detection based on multiple-view geometric cues with application to robotic towel folding. Proceedings of the 2010 IEEE International Conference on Robotics and Automation (ICRA), Anchorage, AK, USA. sourceNot checked — no full text of the source
• cites [10] Doumanoglou, A., Kargakos, A., Kim, T.K., and Malassiotis, S. (June, January 31). Autonomous active recognition and unfolding of clothes using random decision forests and probabilistic planning. Proceedings of the 2014 IEEE international conference on robotics and automation (ICRA), Hong Kong, China. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

These challenges exacerbate existing issues of modern robotic control methods such as imitation learning and reinforcement learning.

“These challenges exacerbate existing issues of modern robotic control methods such as imitation learning (IL) and reinforcement learning (RL).”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

End-to-end policy learning suffers from data efficiency and mode collapse.

“End-to-end policy learning suffers from data efficiency and mode collapse (similar but different state leads to the same action).”

Probably not a substantial claim: may not pass S3 (load-bearing).

DRL and DIL often come with a set of hyper-parameters which have to be manually tuned for each domain.

“DRL and DIL often come with a set of hyper-parameters which have to be manually tuned for each domain.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The next challenge is to model the dynamic behaviour of DOs, where computational complexity, physical accuracy and implementation simplicity are the important criteria for assessing the practicality of a dynamic model.

“The next challenge is to model the dynamic behaviour of DOs, where computational complexity, physical accuracy and implementation simplicity are the important criteria for assessing the practicality of a dynamic model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

One of the key features of modern robotics methods is that they formulate the manipulation problem as a Partially Observable Markov Decision Process and leverage Contextual Markov Decision Processes for developing more robust and general skill controllers.

“One of the key features of modern robotics methods is that they formulate the manipulation problem as a Partially Observable Markov Decision Process (Section 3.2) and leverage Contextual Markov Decision Processes (Section 3.2) for developing more robust and general skill controllers (Section 3.3).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Many simulation environments are used to model CDOs in literature, and are seen as essential for the development of DRL and DIL controllers.

“Many simulation environments are used to model CDOs in literature, and are seen as essential for the development of DRL and DIL controllers.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The most common IL method applied in CDO is behaviour cloning, and there are no applications of IRL in the domain.

“The most common IL method applied in CDO is BC, and we are aware of no applications of IRL in the domain.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Early examples of IL methods, such as behaviour cloning and learning from observation, have demonstrated good performance in knot-tying tasks.

“Early examples of IL methods, such as behaviour cloning (BC) (Section 4.1) and learning from observation (LfO) (Section 4), have demonstrated good performance in knot-tying tasks [61].”

• cites [61] Morita, T., Takamatsu, J., Ogawara, K., Kimura, H., and Ikeuchi, K. (2003, January 14–19). Knot planning from observation. Proceedings of the 2003 IEEE International Conference on Robotics and Automation (Cat. No. 03CH37422), Taipei, Taiwan. sourceNot checked — no full text of the source

May combine two claims.

There is a considerable difference in the observation space and some mismatches in the transition functions between simulation and reality.

“There is a considerable difference in the observation space and some mismatches in the transition functions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Goal-conditioned MDP can be regarded as a special subset of CMDP, where a goal only influences the reward function.

“We can regard Goal-conditioned MDP as a special subset of CMDP, where a goal g∈G only influences the reward function.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Speed and quality of simulations have improved tremendously, making it possible to collect millions of data points using simulation for training perception and control systems and to train a non-linear parametric dynamic model.

“Speed and quality of simulations have improved tremendously, making it possible to collect millions of data points using simulation for training the perception and control systems and to train a non-linear parametric dynamic model.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Skill controllers and goal-conditioned RL are built upon the CMDP framework.

“Skill controllers (Section 3.3) and goal-conditioned RL (Section 5.4) are built upon this framework, and its application will be discussed in cloth-shaping tasks (Section 6.1.2).”

May combine two claims.

None of the simulators can model the interaction between the cloth and the gripper; instead, Nvidia Flex and Pybullet anchor the cloth to the gripper.

“None of the simulators can model the interaction between the cloth and the gripper—instead, Nvidia Flex and Pybullet “anchor” the cloth to the gripper (marking certain particles as gripped so that rest of the cloth moves accordingly).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The fundamental difference between IL and RL is that no reward functions are provided in IL, so it helps to eliminate the effort of engineering rewards.

“The fundamental difference between IL and RL is that no reward functions are provided in IL, so it helps to eliminate the effort of engineering rewards that can be difficult in many robotics applications.”

Probably not a substantial claim: may not pass S3 (load-bearing).

With the advance of deep learning and high-performance hardware, almost all parts of an RMS can be practically replaced with neural networks and trained on collected data.

“With the advance of deep learning (DL) [6] and high-performance hardware, almost all parts of an RMS can be practically replaced with neural networks (NNs) and trained on collected data.”

• cites [6] Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning, MIT Press.Not checked yet

  Not checked in the free run.

May combine two claims.

Since the intermediate state representation is automatically learned in an end-to-end manner in modern robotics, small perturbations in the observation space can cause a significant difference in the latent state representation, leading to substantial deviation from the correct policy.

“However, since the intermediate state representation is automatically learned in an end-to-end manner in modern robotics, small perturbations in the observation space can cause a significant difference in the latent state representation; consequently, this can lead to a substantial deviation from the correct policy.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In robotics, context captures across-task variations and can help to formalise robust and multi-task control settings.

“In robotics, context captures the across-task variations [47] and it can help to formalise robust and multi-task control settings.”

• cites [47] Kroemer, 2021, A Review of Robot Learning for Manipulation: Challenges, Representations, and Algorithms, J. Mach. Learn. Res.Not checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

In classical robotics, the mismatch caused by the dynamic model is partially mitigated through system identification that calibrates the simulator using real-world data.

“In classical robotics, the mismatch caused by the dynamic model is partially mitigated through system identification [83] that calibrates the simulator using real-world data.”

• cites [83] Kristinsson, 1992, System identification and control using genetic algorithms, IEEE Trans. Syst. Man Cybern. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

End-to-end policy learning attempts to eliminate the selection of intermediate state representation from the design process by learning a latent representation so that a control algorithm can be applied to multiple domains.

“End-to-end policy learning attempts to eliminate the selection of intermediate state representation from the design process by learning a latent representation so that a control algorithm can be applied to multiple domains.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Solving long-horizon multi-step problems usually involves decomposing a task into component skills, developing skill policies independently, and using a higher-level policy to decide when to use which skill.

“Solving long-horizon multi-step problems usually involves (1) decomposing a task into component skills with regards to the modular structure, (2) developing skill policies independently and (3) using higher-level policy to decide when to use which skill.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

So far, meta-learning methods and Adversarial IRL have not been used in the CDO domain.

“So far, meta-learning methods and Adversarial IRL have not been used in the CDO domain.”

Probably not a substantial claim: may not pass S4 (based).

Neural network parameters must be trained from scratch for different tasks.

“However, neural network parameters must be trained from scratch for different tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Transfer learning is often applied to transfer domain knowledge learned from a certain task to other domains, where it can be applied directly (zero-shot transfer) or after a few trials (few-shot).

“Hence, transfer learning (TL) is often applied to transfer the domain knowledge learned from a certain task to other domains, where it can be applied directly (zero-shot transfer) or after a few trials (few-shot).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Safety is a major concern in robotic applications, especially in real-world trials.

“Safety is another major concern in robotic applications, especially in real-world trials.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

DeformableRaven models a rope and the hem of a sack using a chain of beads, which introduces a significant reality gap.

“DeformableRaven models a rope and the hem of a sack using a chain of beads, which introduces a significant reality gap.”

May combine two claims.

Demonstration data can be collected through kinesthetic teaching, a motion capture system, a teleoperation system, and an expert script policy.

“Demonstration data can be collected through (1) kinesthetic teaching, (2) a motion capture system, (3) a teleoperation system and (4) an expert script policy.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DAgger solves compounding error and can be regarded as a reduction of imitation learning to supervised learning with interaction.

“Ross et al. (2011) [101] propose Dataset Aggregation (DAgger) to solve compounding error, which can be regarded as a reduction of IL to SL with interaction.”

• cites [101] Ross, S., Gordon, G., and Bagnell, D. (2011, January 11–13). A reduction of imitation learning and structured prediction to no-regret online learning. Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, Fort Lauderdale, FL, USA.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep reinforcement learning creates possibilities for multi-task or goal-conditioned robots, but specification of goals and inference of the reward function for such an agent remain hard challenges.

“DRL creates possibilities for creating multi-task/goal-conditioned robots (Section 5.4).”

• Also stated in: § 5. Reinforcement Learning for CDO Manipulation · plain

  “However, the specification of goals and inference of the reward function for such an agent remain a hard challenges.”

May combine two claims.

Seita et al. extend Transporter Net to the goal-conditioned control domain and apply it to cloth-shaping and bag-manipulation tasks in simulation.

“Seita et al. (2021) [34] extend Transporter Net to the goal-conditioned control domain and apply it to cloth-shaping and bag-manipulation tasks in simulation.”

• cites [34] Seita, D., Florence, P., Tompson, J., Coumans, E., Sindhwani, V., Goldberg, K., and Zeng, A. (June, January 30). Learning to rearrange deformable cables, fabrics, and bags with goal-conditioned transporter networks. Proceedings of the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

Deep Q-learning extends Q-learning to solve high-dimensional or continuous state problems in practice, where actions are still discrete.

“Deep Q-learning (DQN) [126] extends Q-learning to solve high-dimensional/continuous state problems in practice, where actions are still discrete.”

• cites [126] Mnih, 2015, Human-level control through deep reinforcement learning, Nature sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Dynamic Movement Primitives generally have limited capability for generalisation and often require more features for different usage.

“However, DMP generally has limited capability for generalisation and often requires more features for different usage.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Exploration and reward shaping are two significant challenges to applying reinforcement learning in practice.

“Exploration and reward shaping are two significant challenges to applying reinforcement learning in practice.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

LfO is often applied when the demonstrator's action space differs from the agent's.

“LfO is often applied when the demonstrator’s action space differs from the agent’s.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Vinh et al. and Kudoh et al. apply trajectory replay methods to achieve knot-tying tasks.

“Vinh et al. (2012) [91] and Kudoh et al. (2015) [105] apply such methods to achieve knot-tying tasks”

• cites [91] Van Vinh, T., Tomizawa, T., Kudoh, S., and Suehiro, T. (2012, January 14–18). A new strategy for making a knot with a general-purpose arm. Proceedings of the 2012 IEEE International Conference on Robotics and Automation (ICRA), St. Paul, MN, USA. sourceNot checked — no full text of the source
• cites [105] Kudoh, S., Gomi, T., Katano, R., Tomizawa, T., and Suehiro, T. (October, January 28). In-air knotting of rope by a dual-arm multi-finger robot. Proceedings of the 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Hamburg, Germany. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Motion capture data often present a correspondence problem between the demonstrator and the learner.

“However, such data often present a correspondence problem between the demonstrator and the learner.”

Probably not a substantial claim: may not pass S4 (based).

There are no applications of MBBC in the CDO domain.

“As there are no applications of MBBC in the CDO domain, we only focus on the details of MFBC algorithms.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Asynchronous control, the delay between action and sensory input, violates the formulation of a Markov decision process, especially for dynamic control settings.

“Last but not least, asynchronous control, i.e., the delay between action and the sensory input, violates the formulation of MDP, especially for dynamic control settings.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Dynamic Movement Primitives work well for learning point-to-point trajectories because they can easily generalise to different start and goal positions.

“DMP works well for learning point-to-point trajectories as it can easily generalise to different start and goal positions.”

May combine two claims.

Trajectory replay methods are only functional under a small state subset of the task domain and often require extra control to reach and detect states of interest.

“These methods are only functional under a small state subset of the task domain, and often require extra control to reach and detect these states of interest.”

May combine two claims.

In a supervised approach to BC, actions in expert trajectories are assumed to be independent and identically distributed, but causal relationships exist between action and states, which contradicts the i.i.d. assumption.

“In a supervised approach to BC, we assume that actions in the expert trajectories are independent and identically distributed (i.i.d.) but causal relationships exist in BC applications between the action and the states, which contradicts the i.i.d. assumption.”

• AssumptionThe authors state this as an assumption: Actions in expert trajectories are independent and identically distributed in a supervised approach to BC.

May combine two claims.

Reward functions need to be engineered from perception, which contradicts unsupervised perception, and reward design is a hard problem in robotics even with an oracle.

“Second, reward function needs to be engineered from perception, and this contradicts the idea of tackling the perception in an unsupervised manner. Moroever, even with an oracle (where the agent has access to the true state), reward design is a hard problem in robotics.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Lee et al. extend TPS-RPM transfer to force control, and Lee et al. replace TPS-RPM with Coherent Point Drift to improve transferring quality.

“Lee et al. (2015a) [118] extend the work to force control, while Lee et al. (2015b) [119] replace TPS-RPM with Coherent Point Drift to improve the transferring quality.”

• cites [118] Lee, A.X., Lu, H., Gupta, A., Levine, S., and Abbeel, P. (2015, January 26–30). Learning force-based manipulation of deformable objects from multiple demonstrations. Proceedings of the 2015 IEEE International Conference on Robotics and Automation (ICRA), Seattle, WA, USA. sourceNot checked — no full text of the source
• cites [119] Lee, A.X., Goldstein, M.A., Barratt, S.T., and Abbeel, P. (2015, January 26–30). A non-rigid point and normal registration algorithm with applications to learning from demonstrations. Proceedings of the 2015 IEEE International Conference on Robotics and Automation (ICRA), Seattle, WA, USA. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Expert-script policies are functional systems that can perform the target task relatively well in real-world or privileged policies that can gain access to the true state of the environment in simulation.

“Finally, expert-script policies are functional systems that can perform the target task relatively well in real-world or privileged policies that can gain access to the true state of the environment in simulation.”

May combine two claims.

Applying deep reinforcement learning with a partially observable Markov decision process helps overcome state estimation in robotics and automates perception through end-to-end training and multi-modal sensory inputs.

“The application of DRL with POMDP also helps to overcome the challenge of state estimation in robotics [120]. It automates the perception in robotics through end-to-end training and combining multi-modal sensory inputs in the perception network [47].”

• cites [120] Sutton, R.S., and Barto, A.G. (2018). Reinforcement Learning: An Introduction, MIT Press.Not checked yet

  Not checked in the free run.

• cites [47] Kroemer, 2021, A Review of Robot Learning for Manipulation: Challenges, Representations, and Algorithms, J. Mach. Learn. Res.Not checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Robotics is usually formulated in a partially observable setting, where estimating the true state of the environment is a hard problem.

“Third, robotics is usually formulated in a partially observable setting, where estimating the true state of the environment is a hard problem.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Schulman et al. propose to transfer a demonstration trajectory using non-rigid registration between the observed scene and a demonstrated scene, using Thin Plate Spline Robust Point Matching on point clouds.

“Schulman et al. (2016) [116], on the other hand, propose to transfer the demonstration trajectory using non-rigid registration between the observed scene and a demonstrated scene. They conduct the registration using the Thin Plate Spline Robust Point Matching (TPS-RPM) approach proposed by Chui and Rangarajan (2003) [117] on the scene represented as point clouds.”

• cites [116] Schulman, J., Ho, J., Lee, C., and Abbeel, P. (2016). Robotics Research, Springer.Not checked yet

  Not checked in the free run.

• cites [117] Chui, 2003, A new point matching algorithm for non-rigid registration, Comput. Vis. Image Underst. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing).

TRPO is conceptually and computationally challenging mainly due to the constraint in Equation (12).

“TRPO is conceptually and computationally challenging mainly due to the constraint in Equation (12).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Clipping in PPO is a regulariser that makes changes in policy less drastic, but it can still allow large step updates to the policy.

“Clipping is a regulariser that makes changes in policy less drastic, but it can still allow large step updates to the policy.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Small parameter space updates in policy-gradient methods can cause big policy space changes and make training unstable.

“Another issue with policy-gradient methods is that small parameter space updates can cause big policy space changes and make training unstable.”

May combine two claims.

Soft Data Augmentation improves sample efficiency and stabilises reinforcement learning optimisation by decoupling data augmentation from policy learning.

“Soft Data Augmentation (SODA) [142] further improves sample efficiency and stabilises the RL optimisation by decoupling data augmentation from policy learning.”

• cites [142] Hansen, N., and Wang, X. (June, January 30). Generalization in reinforcement learning by soft data augmentation. Proceedings of the 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China. sourceNot checked — no full text of the source

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

If an RL agent explores insufficiently, it may not find highly rewarded states, and its policy will be suboptimal.

“If the agent explores insufficiently, it may not find highly rewarded states, and its policy will be suboptimal.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Hafner et al. propose the Recurrent State Space Model, a probabilistic model that predicts latent states of an environment with a latent dynamic model and refines its prediction based on a sequence of observations.

“Hafner et al. (2019) [156] propose the Recurrent State Space Model (RSSM), which is a probabilistic model that predicts latent states of an environment with a latent dynamic model and refines its prediction based on a sequence of observations.”

• cites [156] Hafner, D., Lillicrap, T., Fischer, I., Villegas, R., Ha, D., Lee, H., and Davidson, J. (2019, January 10–15). Learning latent dynamics for planning from pixels. Proceedings of the International Conference on Machine Learning (ICML), Long Beach, CA, USA.Not checked yet

  Not checked in the free run.

May combine two claims.

Representation learning helps reduce the curse of dimensionality in Markov decision process policy learning and plays the role of state estimation in partially observable Markov decision process settings.

“This helps to reduce the curse of dimensionality in MDP policy learning and, most importantly, plays the role of state estimation in POMDP settings.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DQN's replay buffer can produce a biased estimate of the Q-value, and an additional target network mitigates the bias and helps stabilise training.

“Apart from the off-policy strategy used in Q-learning, it also utilises a replay buffer to learn from all the historical transitions to improve learning efficiency; however, it can produce a biased estimate of the Q-value.”

• Also stated in: § 5.1. Model-Free RL · plain

  “To mitigate the bias, it uses an additional target network, updated periodically from the training network with polyak averaging (soft update, where the new target network is the interpolation between the old target network and training network) [6] to estimate the target value, which also helps to stabilise the training.”

• cites [126] Mnih, 2015, Human-level control through deep reinforcement learning, Nature sourceNot checked — no full text of the source
• cites [6] Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning, MIT Press.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Visual dynamic learning becomes more challenging when occlusions occur among objects, and memory-based approaches such as recurrent networks must be adopted in such settings.

“Visual dynamic learning becomes more challenging when there are occlusions occurring among objects; memory-based approaches, such as recurrent networks, must be adopted in such settings.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Policy gradient is an on-policy approach due to the distribution of the expectation in the gradient, which introduces biases in the off-policy policy-gradient formulation.

“Policy gradient is an on-policy approach due to the distribution of the expectation in the gradient, which introduces biases in the off-policy policy-gradient formulation.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Good encoding representation from videos is challenging due to the high dimensionality and variability of the visual space.

“Good encoding representation from videos is also challenging due to the high dimensionality and variability of the visual space.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The main challenge for representation learning in partially observable Markov decision process settings is to obtain accurate state estimation from partial observation.

“The main challenge for representation learning in POMDP settings is to obtain accurate state estimation from partial observation.”

May combine two claims.

Type I model-based reinforcement learning is commonly applied in cloth shaping and assistive dressing, while no application of Type II and Type III model-based reinforcement learning in cloth-like deformable object manipulation is known.

“Type I MBRL is commonly applied in cloth shaping (Section 6.1.2) and assistive dressing (Section 6.3.2), while we are aware of no application of Type II and III MBRL in CDO.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The balance between exploration and exploitation directly influences sampling efficiency, which affects learning efficiency, the upper bound of performance and training stability.

“The balance between exploration and exploitation directly influences the sampling efficiency of an algorithm that, in turn, affects (1) learning efficiency, (2) the upper bound of its performance and (3) the training stability.”

May combine two claims.

In complex settings, simple exploration approaches suffer from large state-action space and from sparse, delayed and deceptive rewards.

“In more complex settings (high-dimensional or continuous MDP/POMDPs), simple exploration approaches suffer from large state–action space and from sparse, delayed and deceptive rewards [207], where the latter case is known as the hard-exploration problem [208].”

• cites [207] Yang, T., Tang, H., Bai, C., Liu, J., Hao, J., Meng, Z., and Liu, P. (2021). Exploration in deep reinforcement learning: A comprehensive survey. arXiv.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The most practical way to bypass sample efficiency problems in complex MDP/POMDP is to use demonstration trajectories.

“The most practical way to bypass sample efficiency problems in complex MPD/POMDP is to use demonstration trajectories.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Cloth shaping is a crucial skill for doing laundry in daily life.

“Cloth shaping is a crucial skill for doing laundry in daily life [251].”

• cites [251] Hamajima, 1998, Planning strategy for task untangling laundry-isolating clothes from a washed mass, Robotics Mechatron. sourceNot checked — no full text of the source

May combine two claims.

Self-occlusion can hide effective key points for manipulation, which makes it hard to extract the state of the cloth and presents challenges for engineering reward functions and automating the goal evaluation for complex CDO manipulation tasks.

“Self-occlusion can hide effective key points for manipulation, which makes it hard to extract the state of the cloth and presents challenges for engineering reward functions and automating the goal evaluation for complex CDO manipulation tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Ecoffet et al. (2019) point out that the two direct consequences of myopic approaches are detachment and derailment.

“Ecoffet et al. (2019) [208] point out that the two direct consequences of myopic approaches are (1) the agent always forgets the highly rewarded region in the state space (detachment) and (2) the agent cannot robustly return to such states if remembered (derailment).”

• cites [208] Ecoffet, A., Huizinga, J., Lehman, J., Stanley, K.O., and Clune, J. (2019). Go-Explore: A New Approach for Hard-Exploration Problems. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

MaxEnt-RL incorporates the Maximum Entropy Principle to learn a distribution of policies whose mean is close to the optimal policy.

“Instead of learning deterministic policy, maximum entropy reinforcement learning (MaxEnt-RL) [215,216] incorporates the Maximum Entropy Principle to learn a distribution of policies whose mean is close to the optimal policy.”

• cites [215] Ziebart, B.D., Maas, A.L., Bagnell, J.A., and Dey, A.K. (2008, January 13–17). Maximum entropy inverse reinforcement learning. Proceedings of the AAAI, Chicago, IL, USA.Not checked yet

  Not checked in the free run.

• cites [216] Levine, S. (2018). Reinforcement learning and control as probabilistic inference: Tutorial and review. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

DDPGfD is proposed for continuous cases and can also deal with sparse rewards.

“DDPG from Demonstrations (DDPGfD) [219] is proposed for continuous cases and can also deal with sparse rewards.”

• cites [219] Vecerík, M., Hester, T., Scholz, J., Wang, F., Pietquin, O., Piot, B., Heess, N., Rothörl, T., Lampe, T., and Riedmiller, M.A. (2017). Leveraging Demonstrations for Deep Reinforcement Learning on Robotics Problems with Sparse Rewards. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Levine (2018) shows that MaxEnt-RL can be derived with control as probabilistic inference with known deterministic dynamics in tabular cases and variational inference without the knowledge of dynamics.

“By introducing binary optimality variables in high-dimensional continuous MDP settings, Levine (2018) [216] shows that MaxEnt-RL can be derived with control as probabilistic inference with known deterministic dynamics in tabular cases and variational inference without the knowledge of dynamics.”

• cites [216] Levine, S. (2018). Reinforcement learning and control as probabilistic inference: Tutorial and review. arXiv.Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Usually, more end-effectors lead to more efficient manipulation.

“Usually, more end-effectors lead to more efficient manipulation.”

Probably not a substantial claim: may not pass S4 (based).

There are aleatoric uncertainties about the cloth's deformation, meaning that the cloth does not always deform the same way under the same action in the real world.

“There are aleatoric uncertainties about the cloth’s deformation, meaning that the cloth does not always deform the same way under the same action in the real world.”

May combine two claims.

The most fundamental challenge in manipulating CDOs is that they have many DoF, which leads to complex dynamics and severe self-occlusion.

“The most fundamental challenge in manipulating CDOs is that they have many DoF, which leads to complex dynamics and severe self-occlusion.”

May combine two claims.

Dynamic manipulation leads to a much quicker feedback cycle and self-correction on failures.

“In contrast, dynamic manipulation leads to a much quicker feedback cycle and self-correction on failures.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The more complex the shape of the CDO, the harder the task will be.

“The more complex the shape of the CDO, the harder the task will be.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The number of end-effectors, their capabilities and the usage of extra tools can drastically change the nature of the task.

“The number of end-effectors, their capabilities and the usage of extra tools can drastically change the nature of the task.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Both knot tying and untying start from a structural representation in a computer simulation using knowledge of knot theory.

“Both KT and KU start from a structural representation in a computer simulation using knowledge of knot theory.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Early knot untying approaches in topological simulation include random perturbation, annealing schedule, energy minimisation with gradient descent, and motion planning.

“The early KU approaches in topological simulation include random perturbation [276,277], annealing schedule [278], energy minimisation with gradient descent [279] and motion planning [280,281].”

• cites [276] Simon, 1994, Energy functions for polygonal knots, J. Knot Theory Its Ramifications sourceNot checked — no full text of the source
• cites [277] Scharein, R.G. (1998). Interactive Topological Drawing. [Ph.D. Thesis, University of British Columbia].Not checked — no full text of the source
• cites [278] Ligocki, 1994, Recognizing knots using simulated annealing, J. Knot Theory Its Ramif. sourceNot checked — no full text of the source
• cites [279] Huang, M., Grzeszczuk, R.P., and Kauffman, L.H. (November, January 27). Untangling knots by stochastic energy optimization. Proceedings of the Seventh Annual IEEE Visualization’96, San Francisco, CA, USA.Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

On the table, a target knot can be tied and untied by the four Reidemeister moves and rope permutation delivered with pick-and-place action primitives, but this process cannot guarantee the tightness of the knot.

“On the table, a target knot can be tied and untied by the four RMs and rope permutation that can be delivered with P&P action primitives, but this process cannot guarantee the tightness of the knot.”

May combine two claims.

Most approaches in the knot tying/untying literature choose to operate at the topological level, so the perception system needs to extract a topological representation from its sensory data.

“Most approaches in the KT/KU literature choose to operate at the topological level, so the perception system needs to extract a topological representation (TR) from its sensory data.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

KT depends on whether the rope is tied on the table or in the air.

“KT depends on whether the rope is tied on the table or in the air.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Learning from observation (LfO) is another line of research in knot tying/untying literature.

“Apart from motion planning as described in Section 6.2.1, LfO is another line of research in KT/KU literature.”