Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

Deep Reactive Policy: Learning Reactive Manipulator Motion Planning for Dynamic Environments

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims DRP, a point-cloud-conditioned transformer policy augmented with a reactive goal shifter, outperforms classical and learned planners on dynamic manipulation tasks in simulation and the real world.

desk verdict Solid learning pipeline and useful benchmark, but the reactive module that drives the headline gains rests on an unvalidated frame-differencing heuristic. read the letter →

arxiv 2509.06953 v1 pith:72SFRXOD submitted 2025-09-08 cs.RO cs.AIcs.CVcs.LGcs.SYeess.SY

Jiahui Yang , Jason Jingzhou Liu , Yulong Li , Youssef Khaky , Kenneth Shaw , Deepak Pathak This is my paper
classification cs.ROcs.AIcs.CVcs.LGcs.SYeess.SY
keywords reactivemotionplanningvisuo-motorpolicypointcloudperceptiontransformerdynamicobstacleavoidancegoalblockingstudent-teacherdistillationRiemannianpolicies
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a learned, closed-loop visuo-motor policy can handle dynamic and partially observable manipulation scenes better than classical planners and prior neural policies. The method, DRP, combines a transformer policy pretrained on 10 million expert trajectories, a student-teacher finetuning stage that improves static obstacle avoidance, and a lightweight reactive module that adjusts the goal when dynamic obstacles approach. If correct, this makes reactive collision-free motion generation practical without full environment models or expensive replanning, which is the key bottleneck for robots in homes and other human environments. The strongest empirical evidence is on floating-obstacle and dynamic-goal-blocking tasks, where DRP reaches 75.5% and 65.25% success respectively, compared with 39.5% and 3.0% for the expert planner used to generate its training data.

What carries the argument

The load-bearing mechanism is the three-stage stack. IMPACT is a transformer encoder-decoder that consumes scene and robot point clouds downsampled by set abstraction, plus encoded joint configuration and goal, and outputs a chunk of delta joint targets. Iterative student-teacher finetuning couples the pretrained IMPACT with a Geometric Fabrics teacher that uses signed distance fields to refine joint targets, repeatedly distilling these refinements back into the point-cloud student. DCP-RMP is the dynamic-reactive component: it extracts moving points by querying a KD-tree built from the previous frame against the current frame, finds the closest dynamic point to the robot surface, constructs

What would settle it

Run DRP in a static scene while vibrating or rapidly repositioning one camera, or while a moving obstacle passes in front of static furniture; if success rate drops sharply relative to the clean static case and the robot collides with falsely flagged static geometry, the frame-differencing assumption is the cause. A cleaner simulation test: replace DCP-RMP's KD-tree mask with the ground-truth dynamic-object mask and compare success rates; if the ground-truth mask restores the reported 75.5% and 65.25% while frame differencing does not, the dynamic-object detector is the load-bearing component.

Watch

Extended reading notes

Core claim

DRP is an end-to-end neural motion policy that maps raw point clouds plus current and target joint positions directly to joint action chunks. Its core, IMPACT, is an action-chunking transformer pretrained on 10 million cuRobo-generated trajectories, then finetuned by iterative student-teacher distillation with a Geometric Fabrics teacher that refines joint targets using privileged obstacle information. At inference, DCP-RMP identifies dynamic obstacle points by KD-tree frame differencing, computes the closest point to the robot, and applies a Riemannian Motion Policy repulsive acceleration that virtually moves the joint-space goal away from the obstacle before IMPACT tracks it. The paper rep

Load-bearing premise

The biggest performance gains come from DCP-RMP classifying points as dynamic if they appear in one camera frame but not the previous one; if sensor noise or occlusion changes create spurious frame differences, the policy pushes away from static geometry and the reactive advantage can turn into collisions.

Editorial extensions

If this is right

  • DRP runs closed-loop at 300 Hz with a 3.48 ms cold-start time, while NeuralMP's test-time optimization takes about 2970 ms, so learned reactive control can operate in real time on live point clouds.
  • Classical planners that require full environment knowledge fail on dynamic and goal-blocking tasks, and the paper shows that adding a reactive goal-proposal module recovers most of the lost performance.
  • DCP-RMP is method-agnostic: adding it to cuRobo, NeuralMP, and IMPACT improves their success rates on dynamic tasks, so the reactive goal-shifting idea transfers beyond the specific policy.
  • Including goal-blocked training scenarios and scaling to 10 million expert trajectories yields a policy that generalizes zero-shot to the MπNets benchmark and to real-world scenes despite simulation-only training.
  • The final policy outperforms its own expert data generator on static scenes, indicating that the student-teacher finetuning stage contributes beyond imitation of the original planner.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • DCP-RMP's frame-differencing assumption is the fragile link: under camera noise, jitter, or a moving obstacle that changes occlusions of static geometry, static points will be mislabeled as dynamic and the robot will receive an incorrect repulsive push. A controlled static-scene test with camera vibration would isolate this failure mode.
  • The reactive module only modifies the goal rather than the policy's internal representation, so it could be attached to any goal-conditioned planner or policy; the architecture contribution and the reactive contribution are separable and independently reusable.
  • The gains on Floating Dynamic Obstacles and Dynamic Goal Blocking likely depend on the repulsion metric's tuning and the number of Euler integration steps, meaning the method may be sensitive to velocity estimation of dynamic points from frame differencing alone.
  • The real-world evaluation uses one robot arm and multiple calibrated cameras; whether the method degrades gracefully with a single cheap camera or in occlusion-heavy narrow spaces is untested and follows directly from the paper's stated limitation on perception accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes Deep Reactive Policy (DRP), a closed-loop visuo-motor motion policy for manipulators in dynamic, partially observable environments. DRP consists of three components: IMPACT, a transformer-based policy pretrained by behavior cloning on 10M cuRobo-generated trajectories; an iterative student-teacher finetuning stage that distills a Geometric Fabrics teacher into the point-cloud-conditioned student; and DCP-RMP, a point-cloud-based Riemannian Motion Policy module that modifies the goal at inference time based on dynamically classified points. The authors evaluate DRP on a new benchmark, DRPBench, containing static, suddenly appearing, floating dynamic, goal-blocking, and dynamic-goal-blocking tasks in simulation and on a real Franka Panda setup, reporting success rates that outperform prior classical (AIT*, cuRobo, RMP) and learning-based (MpiNets, MpiFormer, NeuralMP) baselines. The strongest reported gains occur in dynamic tasks, e.g., FDO 75.50% vs. 32.00% for IMPACT alone and DGB 65.25% vs. 0.25%.

Significance. If the empirical claims are fully supported, DRP would be a meaningful advance: it combines large-scale pretraining on a modern GPU planner, a scalable transformer architecture, and a lightweight reactive goal-proposal module, and it demonstrates sim-to-real transfer on a multi-camera, point-cloud-only system. The authors also provide useful ablations (architecture, finetuning, DCP-RMP) and report fast inference (300 Hz, cold-start 3.48 ms). However, the load-bearing evidence for the reactive component is incomplete. The dynamic-point extraction in DCP-RMP is not validated against ground truth or noise, the headline success rates are point estimates without confidence intervals or task-level trial counts, and the hand-designed DCP-RMP gains are not reported. These gaps currently prevent a full endorsement of the central claim that the learned-plus-reactive combination is responsible for the reported superiority in dynamic scenarios.

major comments (4)
  1. [Appendix C, Eq. (DCP-RMP dynamic-point extraction); Table 1] Dynamic obstacle points are extracted by building a KD-tree from the previous point-cloud frame and labeling current-frame points not found in it as dynamic. This equates 'newly visible' with 'moving'. In real depth streams, spurious points, sensor dropout, occlusion changes, and static geometry that reappears after a moving obstacle passes will all be classified as dynamic. These false positives enter the repulsive policy f_r and metric M_r and alter qmg before IMPACT. The largest reported gains in Table 1 (FDO 75.50 vs. 32.00; DGB 65.25 vs. 0.25) depend directly on this classification. The paper does not state whether robot self-points are masked before the KD-tree query, and it does not provide a simulation ablation with a ground-truth dynamic-object mask. Without such validation or a false-positive analysis, the reactive-superiority claim is not fully established.
  2. [Sec. 4.1 and 4.2, Tables 1 and 3] All success rates are single point estimates with no confidence intervals, no number of trials per task, and no indication of how many evaluation seeds or environment instances underlie each number. The real-world text states 'over 50 real-world instances' but gives no per-task denominators; the real-world FDO and DGB rows (70.00 and 93.33, respectively) are thus impossible to interpret statistically. Since the paper's central claim is empirical superiority over baselines, the absence of variance information is load-bearing. Please report trial counts, confidence intervals (e.g., Wilson intervals), and, where feasible, multiple seeds or environment resampling for simulation.
  3. [Appendix C, equations for f_r and M_r; Sec. 3.3] DCP-RMP is a hand-designed module with multiple free gains: kp, ℓp, kv, lv, ℓd, ℓm, µr, r, and the number of Euler integration steps per control loop. None of these values are reported in the main text or in Table 7. DCP-RMP is responsible for the largest improvements in Table 1, so the missing values preclude reproducibility and make it impossible to assess whether the gains were obtained by tuning the module on the evaluation benchmark. Please report all DCP-RMP hyperparameters and include a sensitivity study (e.g., varying the gains across a reasonable range) to show that the reported improvements are not artifacts of a single tuned configuration.
  4. [Sec. 3.2, sentence citing 'Table 6'] The text states that iterative student-teacher finetuning 'improves the success rate over the pretrained model by 45%, as shown in Table 6.' Table 6 does not contain a pretrained-only IMPACT row, and its columns are labeled as reaching/collision rates rather than the success rates in Table 1. The claimed 45% improvement therefore cannot be verified from the cited table. Please add the missing pretrained-only row, correct the citation, and specify how the percentage is computed.
minor comments (5)
  1. [Appendix C, first paragraph] The name 'DRP-RMP' appears in the first paragraph but the module is called DCP-RMP elsewhere; please unify the notation.
  2. [Table 6 and accompanying text] Table 6 is titled 'DRPBench (R/C)' but the entries include malformed values such as '94.00/36/33'. Define R and C in the caption, correct the formatting, and clarify the relationship between R/C and the success rates in Table 1.
  3. [Table 2 and Sec. 4.1] The text says DCP-RMP improves performance 'across all dynamic tasks,' but Table 2 reports only FDO and DGB. Either add the other dynamic tasks or qualify the claim.
  4. [Sec. 4.2] The real-world benchmark reports 'over 50 real-world instances' but Table 3 percentages are given with two decimal places. Please state exact per-task trial counts, since percentages alone are misleading with small denominators.
  5. [Appendix D, Table 4] The row labeled 'NeuralMP [9]' actually appears to be the LSTM-GMM architecture trained with the authors' data and finetuning; the original NeuralMP numbers in Table 1 are much lower. Please relabel the row (e.g., 'LSTM-GMM (NeuralMP arch)') to avoid confusion.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: central claims are empirical performance comparisons, not derivations that reduce to their inputs.

full rationale

The paper's central claim is an empirical one: DRP achieves higher success rates than prior classical and learning-based methods on DRPBench, MπNets, and real-world tasks. These are benchmark comparisons, not mathematical derivations, so the main self-definitional and fitted-input circularity patterns do not apply. The DCP-RMP module is an inference-time, hand-designed reactive controller with equations given in Appendix C; while its dynamic-obstacle classification via KD-tree frame differencing is a heuristic assumption that could be fragile under sensor noise, nothing in the paper shows that its parameters were fitted to the benchmark success metric or that the benchmark success is used in its design. The student-teacher finetuning uses Geometric Fabrics as an external privileged-information teacher and distills it into a point-cloud student; this is not circular because the teacher contains an external, independently defined controller. Pretraining with cuRobo trajectories and then comparing against cuRobo is a teacher-vs-student comparison, not a hidden reuse of the test outcome. The only notable self-citation is [9] (NeuralMP), which is cited for the data-generation pipeline; the paper replaces the expert planner with cuRobo and scales to 10M trajectories, and the cited pipeline does not by itself imply the reported results. This self-citation is minor and not load-bearing. The paper also evaluates zero-shot on the external MπNets benchmark, providing independent evidence. Therefore, no exhibited circular step meets the evidentiary bar; concerns about DCP-RMP robustness belong to correctness risk, not circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The system rests on four domain assumptions about point-cloud sufficiency, static-to-dynamic transfer, the Geometric Fabrics teacher, and frame-differencing for dynamic obstacles. The DCP-RMP gains are free parameters not specified in the paper, and no new physical entities are introduced.

free parameters (1)
  • DCP-RMP gains = not reported
    Appendix C defines 13 hand-chosen gains (kg, kd, μg, kp, ℓp, kv, lv, ℓd, εd, µr, ℓm, εm, r) with no numerical values. These govern the repulsive acceleration that produces the reported FDO/DGB improvements (Table 2), so they are free parameters whose values are essential to reproducing the results.
assumptions (4)
  • domain assumption Point cloud plus joint positions is a sufficient state representation for reactive motion planning.
    Section 3.1 conditions IMPACT only on point clouds and joint angles; real-world deployment uses calibrated depth cameras and mesh-based robot points. Any state information needed for collision avoidance must be recoverable from these inputs.
  • domain assumption Simulated static-scene expert trajectories transfer to dynamic and partially observable deployment.
    All 10M pretraining trajectories and the Geometric Fabrics finetuning occur in static scenes (Sections 3.1, 3.2). Dynamic tasks are not re-planned by an expert (Section 3.3 states this is computationally infeasible), so the policy must transfer static expertise to dynamic settings.
  • domain assumption Geometric Fabrics with privileged SDFs provides a reliable teacher for point-cloud-conditioned student finetuning.
    Section 3.2 uses this teacher to refine IMPACT's joint targets. If the teacher's local avoidance behavior is not transferable through distillation, the claimed 45% finetuning improvement would not materialize.
  • domain assumption Dynamic obstacles are detectable as points present in the previous point-cloud frame but not the current frame.
    Appendix C describes KD-tree frame differencing for dynamic point extraction. This assumption is load-bearing for DCP-RMP and is validated only implicitly through the overall success rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Reactive Policy: Learning Reactive Manipulator Motion Planning for Dynamic Environments." pith.science (2026). https://pith.science/paper/72SFRXOD

@misc{pith2026250906953,
  author       = {Pith},
  title        = {Pith review of: Deep Reactive Policy: Learning Reactive Manipulator Motion Planning for Dynamic Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/72SFRXOD}},
  note         = {Machine review of arXiv:2509.06953}
}
read the original abstract

Generating collision-free motion in dynamic, partially observable environments is a fundamental challenge for robotic manipulators. Classical motion planners can compute globally optimal trajectories but require full environment knowledge and are typically too slow for dynamic scenes. Neural motion policies offer a promising alternative by operating in closed-loop directly on raw sensory inputs but often struggle to generalize in complex or dynamic settings. We propose Deep Reactive Policy (DRP), a visuo-motor neural motion policy designed for reactive motion generation in diverse dynamic environments, operating directly on point cloud sensory input. At its core is IMPACT, a transformer-based neural motion policy pretrained on 10 million generated expert trajectories across diverse simulation scenarios. We further improve IMPACT's static obstacle avoidance through iterative student-teacher finetuning. We additionally enhance the policy's dynamic obstacle avoidance at inference time using DCP-RMP, a locally reactive goal-proposal module. We evaluate DRP on challenging tasks featuring cluttered scenes, dynamic moving obstacles, and goal obstructions. DRP achieves strong generalization, outperforming prior classical and neural methods in success rate across both simulated and real-world settings. Video results and code available at https://deep-reactive-policy.com

Figures

Figures reproduced from arXiv: 2509.06953 by the authors.

Figure 1
Figure 1. We present Deep Reactive Policy (DRP), a point cloud conditioned motion policy capable of performing reactive, collision-free goal reaching in diverse complex and dynamic environments. Abstract: Generating collision-free motion in dynamic, partially observable envi￾ronments is a fundamental challenge for robotic manipulators. Classical motion planners can compute globally optimal trajectories but require full enviro… view at source ↗
Figure 2
Figure 2. Deep Reactive Policy (DRP) is a visuo-motor neural motion policy designed for dynamic, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The IMPACT policy is combined with a locally reac [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: DRPBench introduces five challenging evaluation scenarios in simulation and real. In addition to complex Static Environments (SE), we propose Suddenly Appearing Obstacle (SAO) where obstacles appear suddenly ahead of the robot, Floating Dynamic Obstacles (FDO) where ob…

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. OmniRobotHome: A Multi-Camera Platform for Real-Time Multiadic Human-Robot Interaction

    cs.RO 2026-04 unverdicted novelty 7.0 of 10

    A 48-camera residential platform delivers real-time occlusion-robust 3D perception and coordinated actuation for multi-human multi-robot interaction in a shared home workspace.

  2. Flow Motion Policy: Manipulator Motion Planning with Flow Matching Models

    cs.RO 2026-04 unverdicted novelty 7.0 of 10

    Flow Motion Policy uses flow matching to model distributions over feasible manipulator paths, enabling best-of-N sampling with post-generation collision filtering to improve success and efficiency over prior neural an...

  3. ELMP: Efficient Learning for Motion Planning via Analytical Policy Gradients

    cs.RO 2026-06 unverdicted novelty 5.0 of 10

    ELMP performs data-efficient self-supervised adaptation of neural motion planners via analytical policy gradients and point-cloud tool encoding, raising success from 57.3% zero-shot to 89.8% in unseen environments.

  4. Redefining End-of-Life: Intelligent Automation for Electronics Remanufacturing Systems

    eess.SY 2026-04 unverdicted novelty 2.0 of 10

    A literature review of intelligent automation approaches using robotics, AI, and control for disassembly, inspection, sorting, and reprocessing of end-of-life electronics.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages · cited by 4 Pith papers

  1. [1]

    P. E. Hart, N. J. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum cost paths.IEEE transactions on Systems Science and Cybernetics, 4(2):100–107, 1968

    work page 1968

  2. [2]

    M. P. Strub and J. D. Gammell. Adaptively informed trees (ait*): Fast asymptotically optimal path planning through adaptive heuristics. In2020 IEEE International Conference on Robotics and Automation (ICRA), pages 3191–3198. IEEE, 2020

    work page 2020

  3. [3]

    O. Khatib. Real-time obstacle avoidance for manipulators and mobile robots. InProceedings. 1985 IEEE international conference on robotics and automation, volume 2, pages 500–505. IEEE, 1985

    work page 1985

  4. [4]

    N. D. Ratliff, J. Issac, D. Kappler, S. Birchfield, and D. Fox. Riemannian motion policies.arXiv preprint arXiv:1801.02854, 2018

    arXiv 2018

  5. [5]

    Van Wyk, M

    K. Van Wyk, M. Xie, A. Li, M. A. Rana, B. Babich, B. Peele, Q. Wan, I. Akinola, B. Sun- daralingam, D. Fox, et al. Geometric fabrics: Generalizing classical mechanics to capture the physics of behavior.IEEE Robotics and Automation Letters, 7(2):3202–3209, 2022

    work page 2022

  6. [6]

    Bhardwaj, B

    M. Bhardwaj, B. Sundaralingam, A. Mousavian, N. D. Ratliff, D. Fox, F. Ramos, and B. Boots. Storm: An integrated framework for fast joint-space model-predictive control for reactive manipulation. InConference on Robot Learning, pages 750–759. PMLR, 2022

    work page 2022

  7. [7]

    Fishman, A

    A. Fishman, A. Murali, C. Eppner, B. Peele, B. Boots, and D. Fox. Motion policy networks. In conference on Robot Learning, pages 967–977. PMLR, 2023

    work page 2023

  8. [8]

    A. H. Qureshi, A. Simeonov, M. J. Bency, and M. C. Yip. Motion planning networks. In2019 International Conference on Robotics and Automation (ICRA), pages 2118–2124. IEEE, 2019

    work page 2019

Show all 36 references
  1. [9]

    Dalal, J

    M. Dalal, J. Yang, R. Mendonca, Y . Khaky, R. Salakhutdinov, and D. Pathak. Neural mp: A generalist neural motion planner.arXiv preprint arXiv:2409.05864, 2024

    2024 arXiv

  2. [10]

    Fishman, A

    A. Fishman, A. Walsman, M. Bhardwaj, W. Yuan, B. Sundaralingam, B. Boots, and D. Fox. Avoid everything: Model-free collision avoidance with expert-guided fine-tuning. InCoRL Workshop on Safe and Robust Robot Learning for Operation in the Real World, 2024

    2024

  3. [11]

    Sundaralingam, S

    B. Sundaralingam, S. K. S. Hari, A. Fishman, C. Garrett, K. Van Wyk, V . Blukis, A. Millane, H. Oleynikova, A. Handa, F. Ramos, et al. Curobo: Parallelized collision-free robot motion generation. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 811...

    2023

  4. [12]

    Likhachev, G

    M. Likhachev, G. J. Gordon, and S. Thrun. Ara*: Anytime a* with provable bounds on sub-optimality.Advances in neural information processing systems, 16, 2003

    2003

  5. [13]

    Likhachev, D

    M. Likhachev, D. I. Ferguson, G. J. Gordon, A. Stentz, and S. Thrun. Anytime dynamic a*: An anytime, replanning algorithm. InICAPS, volume 5, pages 262–271, 2005

    2005

  6. [14]

    Koenig and M

    S. Koenig and M. Likhachev. A new principle for incremental heuristic search: Theoretical results. InICAPS, pages 402–405, 2006

    2006

  7. [15]

    L. E. Kavraki, P. Svestka, J.-C. Latombe, and M. H. Overmars. Probabilistic roadmaps for path planning in high-dimensional configuration spaces.IEEE transactions on Robotics and Automation, 12(4):566–580, 1996

    1996

  8. [16]

    S. LaValle. Rapidly-exploring random trees: A new tool for path planning.Research Report 9811, 1998. 10

    1998

  9. [17]

    Bohlin and L

    R. Bohlin and L. E. Kavraki. Path planning using lazy prm. InProceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), volume 1, pages 521–528. IEEE, 2000

    2000

  10. [18]

    S. M. LaValle and J. J. Kuffner. Rapidly-exploring random trees: Progress and prospects: Steven m. lavalle, iowa state university, a james j. kuffner, jr., university of tokyo, tokyo, japan. Algorithmic and computational robotics, pages 303–307, 2001

    2001

  11. [19]

    J. J. Kuffner and S. M. LaValle. Rrt-connect: An efficient approach to single-query path planning. InProceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No. 00CH37065), volume 2, pages 995–1001. IEEE, 2000

    2000

  12. [20]

    J. D. Gammell, S. S. Srinivasa, and T. D. Barfoot. Batch informed trees (bit*): Sampling-based optimal planning via the heuristically guided search of implicit random geometric graphs. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 3067–3074,

    2015

  13. [21]

    Zucker, N

    M. Zucker, N. Ratliff, A. D. Dragan, M. Pivtoraiko, M. Klingensmith, C. M. Dellin, J. A. Bagnell, and S. S. Srinivasa. Chomp: Covariant hamiltonian optimization for motion planning. The International journal of robotics research, 32(9-10):1164–1193, 2013

    2013

  14. [22]

    Schulman, Y

    J. Schulman, Y . Duan, J. Ho, A. Lee, I. Awwal, H. Bradlow, J. Pan, S. Patil, K. Goldberg, and P. Abbeel. Motion planning with sequential convex optimization and convex collision checking. The International Journal of Robotics Research, 33(9):1251–1270, 2014

    2014

  15. [23]

    A. D. Dragan, N. D. Ratliff, and S. S. Srinivasa. Manipulation planning with goal sets using constrained trajectory optimization. In2011 IEEE International Conference on Robotics and Automation, pages 4582–4588. IEEE, 2011

    2011

  16. [24]

    Kumar, A

    R. Kumar, A. Mandalika, S. Choudhury, and S. Srinivasa. Lego: Leveraging experience in roadmap generation for sampling-based planning. In2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1488–1495. IEEE, 2019

    2019

  17. [25]

    Zhang, J

    C. Zhang, J. Huh, and D. D. Lee. Learning implicit sampling distributions for motion planning. in 2018 ieee. InRSJ International Conference on Intelligent Robots and Systems (IROS), pages 3654–3661, 2018

    2018

  18. [26]

    Chamzas, Z

    C. Chamzas, Z. Kingston, C. Quintero-Pe ˜na, A. Shrivastava, and L. E. Kavraki. Learning sampling distributions using local 3d workspace decompositions for motion planning in high dimensions. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1283–12...

    2021

  19. [27]

    Ichter, P

    B. Ichter, P. Sermanet, and C. Lynch. Broadly-exploring, local-policy trees for long-horizon task planning.arXiv preprint arXiv:2010.06491, 2020

    2010 arXiv

  20. [28]

    A. H. Qureshi, J. Dong, A. Choe, and M. C. Yip. Neural manipulation planning on constraint manifolds.IEEE Robotics and Automation Letters, 5(4):6089–6096, 2020

    2020

  21. [29]

    Huang, B

    H. Huang, B. Sundaralingam, A. Mousavian, A. Murali, K. Goldberg, and D. Fox. Diffusion- seeder: Seeding motion optimization with diffusion for rapid motion planning.arXiv preprint arXiv:2410.16727, 2024

    2024 arXiv

  22. [30]

    J. J. Johnson, L. Li, F. Liu, A. H. Qureshi, and M. C. Yip. Dynamically constrained motion planning networks for non-holonomic robots. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6937–6943. IEEE, 2020

    2020

  23. [31]

    Carvalho, A

    J. Carvalho, A. T. Le, M. Baierl, D. Koert, and J. Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1916–1923. IEEE, 2023. 11

    1916

  24. [32]

    K. Saha, V . Mandadi, J. Reddy, A. Srikanth, A. Agarwal, B. Sen, A. Singh, and M. Krishna. Edmp: Ensemble-of-costs-guided diffusion for motion planning. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 10351–10358. IEEE, 2024

    2024

  25. [33]

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    2017

  26. [34]

    S. Ross, G. Gordon, and D. Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. InProceedings of the fourteenth international conference on artifi- cial intelligence and statistics, pages 627–635. JMLR Workshop and Conference Proce...

    2011

  27. [35]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021. 12 Appendix A Detailed Task Description...

    2021 arXiv

  28. [2015]

    doi:10.1109/ICRA.2015.7139620

    2015

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.