Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Train-Once Plan-Anywhere Kinodynamic Motion Planning via Diffusion Trees

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A diffusion-trained action sampler keeps RRT's completeness and solves unseen mazes more often.

desk verdict A solid empirical combination of a learned flow-matching action prior with RRT-style search that shows real OOD gains, but the probabilistic-completeness proof has a genuine gap and the deployed sampler does not satisfy its assumptions. read the letter →

arxiv 2508.21001 v2 pith:HCTFN23M submitted 2025-08-28 cs.LG cs.AIcs.RO

Yaniv Hassidof , Tom Jurgenson , Kiril Solovey This is my paper
classification cs.LGcs.AIcs.RO
keywords kinodynamicmotionplanningdiffusionpoliciesflowmatchingsampling-basedplannersRRTprobabilisticcompletenessout-of-distributiongeneralizationactionsampling
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 show that a robot planner can learn from one maze and then plan in mazes it has never seen, without giving up the guarantees that classical sampling-based planners provide. The proposal, DiTree, replaces the uniform random action sampling inside an RRT search tree with a conditional flow-matching policy that has been trained once on expert trajectories recorded in a single environment. The tree remains the safety net: every edge is produced by forward-simulating the robot's dynamics and is kept only if it passes collision checking, so any returned path is dynamically feasible and collision-free. The authors prove DiTree stays probabilistically complete—eventual success with high probability, under the usual Lipschitz-dynamics and robust-path assumptions—provided the learned action sampler has full support on action sequences. In tests on fifteen out-of-distribution maps, DiTree reports a 73.3% average success rate on a dynamic car (against 45–47% for RRT, SST, and the standalone diffusion policy) and 35.7% on a 29-dimensional quadruped, where the standalone policy reaches 7.3% and classical planners fail entirely; a real-car corner-turning run also shows fewer collisions and smaller tracking deviation than RRT.

What carries the argument

The key machinery is the conditional flow-matching sampler used as the action selector (Line 4 of the SBP blueprint), conditioned on a local occupancy observation, a goal-relative target, and the current node, with a diffusion goal bias that sometimes replaces the goal by a random exploration target. This sampler replaces uniform action sampling while keeping the RRT node-selector, forward propagation, and collision checking. The theoretical load is carried by the full-support property of diffusion and flow-matching distributions: positive probability on every open set of action sequences, which is exactly what the RRT probabilistic-completeness proof needs to make the per-ball progress prob

What would settle it

Run the released DiTree sampler from a fixed node many thousands of times and estimate the support of the resulting action-sequence distribution; if any non-empty open set of control sequences (say, a narrow cone of steering/throttle pairs) receives zero samples, full support is violated. A targeted planning test is also decisive: build a robust-clearance problem whose only feasible path requires controls inside a very narrow band, and check whether DiTree ever solves it with unbounded sample count; never succeeding would falsify the completeness claim.

Watch

Extended reading notes

Core claim

The central claim is that learning and search can be combined so that the learned part supplies the smart action-selection prior while the search part supplies the guarantees. Concretely, DiTree samples sequences of N controls from a distribution conditioned on the current node, a goal-relative target, and a local occupancy observation, runs them through the dynamic model, and adds the collision-free edge to the tree; the conditioning target is the goal with probability set by a diffusion goal bias, otherwise a random exploration state. Theorem 1 says this RRT-based configuration is probabilistically complete under Lipschitz-continuous dynamics and a robust nominal solution, with the same ex

Load-bearing premise

The load-bearing premise is that the learned sampler assigns positive probability to every non-empty open region of the action-sequence space; the proof gestures at this via a Lipschitz flow map, but only establishes a diffeomorphism onto its image, and the deployed single-step sampler adds no noise, so if any control ball is never sampled the completeness guarantee does not apply to the implemented system.

Editorial extensions

If this is right

  • If the central claim is right, a single training environment is enough to plan usefully in a wide variety of unseen mazes, so the data-collection cost of learning-based planning drops sharply.
  • Collision avoidance is not the learned model's responsibility; because every stored edge is collision-checked, DiTree's returned trajectories are safe by construction even when the policy mispredicts.
  • The probabilistic-completeness guarantee transfers from RRT, so DiTree inherits the standard exponential bound: failure probability decays as exp(-b k) in the number of samples, under the full-support condition.
  • Pairing DiTree with AO-RRT or SST should carry the same argument to asymptotically optimal planning, not just feasible planning.
  • The computational bottleneck becomes generative inference—about 94% of runtime per the appendix—so advances in fast sampling should convert directly into planning success within a time budget.

Reading between the lines

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

  • My inference: the formal guarantee and the released implementation may be separated here. Theorem 1 assumes full support, but the appendix's ODE argument establishes only a diffeomorphism onto its image, and the evaluated sampler uses a single deterministic flow-matching step without added output noise; the paper's own Remark suggests a small final noise injection as the repair. Testing whether th
  • My inference: because the best results came from one flow-matching step inside a tree, the learned model acts more as a direction prior than a trajectory generator; a cheaper one-step guidance model—possibly distilled or quantized—may capture most of the benefit.
  • My inference: the fully-observed local map assumption in the paper could be relaxed; a replanning variant that re-conditions on fresh local observations from a rolling sensor window would test whether the approach generalizes to partially observed environments.
  • My inference: the ant-robot results suggest the margin over classical planners grows with state-space dimension; testing on higher-dimensional manipulation or legged locomotion would tell whether that trend continues.
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

3 major / 4 minor

Summary. The paper proposes DiTree, a hybrid kinodynamic motion planner in which a conditional diffusion/flow-matching policy, trained once on a single maze, is used as the action sampler inside an RRT-style sampling-based tree. The claimed contribution is two-fold: (1) empirically, replacing RRT's uniform action sampling with the learned sampler yields substantially higher success rates than standalone diffusion policies and classical SBPs on unseen mazes, for both a car and the 29-DoF MuJoCo ant; (2) theoretically, the paper argues that because diffusion/flow-matching samplers have full support on the action space, the RRT backbone retains probabilistic completeness (Theorem 1), and the framework is therefore 'provably-generalizable' and 'provably-safe.' The experiments cover 15 unseen scenarios, include ablations of denoising steps, propagation duration, and goal bias, and report a real-robot cornering comparison where DiTree avoids collisions that RRT trajectories incur.

Significance. If the theoretical guarantee were established for the deployed system, this would be a valuable contribution: it shows a concrete way to inject learned, context-aware priors into sampling-based kinodynamic planning while preserving the collision-free and completeness properties of search. The empirical results are measured rather than fitted, and the paper is unusually thorough in reporting per-scenario success rates, runtimes, and trajectory lengths, including cases where DiTree fails. The real-robot experiment adds credibility to the claim that the learned sampler produces trackable, safe trajectories. However, the central theoretical claim is not currently supported by the proof as written, and the deployed single-step flow-matching sampler does not obviously satisfy the full-support assumption. This gap directly affects the 'provably-generalizable' framing, so the paper needs substantial revision before the guarantee can be accepted.

major comments (3)
  1. [Appendix A.3.1 (ODE formulation, Eq. 1)] The full-support proof for deterministic ODE sampling is incomplete. The argument that Tθ is a C^1 diffeomorphism onto its image, combined with a Gaussian source, implies positive probability on every open ball in R^D only if Tθ is surjective (or at least its image intersects every ball). A diffeomorphism onto its image can have a proper image, e.g. a contraction mapping into a bounded set, in which case the preimage of a ball outside the image is empty. Surjectivity of the trained flow map is not shown, and the global Lipschitz/C^1 regularity of vθ for a UNet velocity field is assumed rather than established. Thus the claimed derivation of full support for DMs/FMs is not valid as written.
  2. [Sec. 5 (single-step FM) and Appendix A.3.2 (Remark)] The evaluated implementation uses single-step flow matching, which is the deterministic map u1 = u0 + vθ(u0,0). The Remark in A.3.2 explicitly concedes that deterministic/noise-free sampling may lack full support and says output noise can restore it, but the experimental section reports no such output noise. Therefore the action sampler used to produce Table 1 has no established full support, and Theorem 1 does not apply to the shipped system. The Abstract's 'provably-generalizable' claim and Section 4.2's statement that SBP guarantees are preserved are not supported for the implementation as described. Please either add the output-noise term to the implementation and re-evaluate, or restrict the theoretical claim to a variant that is actually deployed.
  3. [Sec. 4.1 vs Appendix A.2 (MPC-style resampling)] The proof of Theorem 1 models a fixed action sequence u_{1:N} sampled once and then propagated with NΔt = trand. Section 4.1, however, describes the deployed edge generation as an MPC-style loop that 'iteratively resamples actions based on the updated local observations' and performs 'perpetual sampling and forward propagation.' The proof does not cover this resampling procedure, since the distribution pθ(u_{1:N} | xnear, xtarget, X_obs^{near}) changes as xnear and the local observation are updated. Either align the theoretical model with the implemented algorithm, or prove that the resampling loop preserves the lower-bound probability required in the RRT completeness argument.
minor comments (4)
  1. [Sec. 4.2 / Theorem 1] The term 'full support' is used in the theorem but not defined in the main text. Please define it explicitly (e.g., positive probability for every nonempty open subset of U^N) so the dependency of the guarantee on this assumption is clear.
  2. [Sec. 5.2 / Fig. 5] In the first ablation panel, the legend uses 'DGB 1' while the text refers to percentages. If this means 100%, please write '100%' for consistency.
  3. [Sec. 4.2 / Appendix A.4] The main text states that asymptotic optimality is achieved when pairing DiTree with SST and AO-RRT, but Appendix A.4 says the SST AO proof is left for future work. Please temper the main-text wording to avoid implying a formal AO guarantee for SST.
  4. [Appendix A.3.1] The phrase 'full support on R^D' could be misinterpreted when the control space U is a bounded subset. Clarify whether outputs are projected/clipped to U and how the support argument interacts with such a projection.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DiTree's claims are empirically measured and the PC theorem is conditional on an independently-established RRT result; the full-support proof gap is a correctness issue, not circularity.

full rationale

DiTree's central empirical claim (higher success rates on unseen mazes after training on one maze, Table 1) is a measured outcome against external baselines, not a fitted quantity renamed as a prediction; hyperparameters were chosen on a validation split, which is standard practice. The theoretical claim (Theorem 1) is explicitly conditional: if the action sampler has full support, then RRT-based DiTree is probabilistically complete. The proof transfers the standard RRT PC argument from [56] by replacing the uniform-action probability p_u with the product of per-action probabilities, and the only new requirement is positivity of those probabilities. The [56] citation includes a co-author, but it is a published, parameter-free theorem whose assumptions do not include the DiTree result, so under the review rules it counts as independent support rather than self-citation circularity. The full-support derivation in Appendix A.3.1 argues that a Gaussian source and a C^1 flow map give positive preimage measure for every ball. One may object that a diffeomorphism 'onto its image' does not by itself guarantee that every ball intersects the image, and that the deployed single-step FM sampler is deterministic and may lack the output noise mentioned in the Remark; but this is an incompleteness/non-application of a stated assumption, not a reduction of the conclusion to the premise by construction. No equation in the paper defines a predicted quantity as the fitted parameter it is supposed to predict, and no load-bearing claim rests solely on a self-citation. Therefore no circular step is identified.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced. The central theoretical claim rests on the full-support assumption, which is only partially proved and may not hold for the actual single-step deterministic FM sampler used in experiments. Hyperparameters (DGB, N, denoising steps) are fitted to validation, not to the test scenarios.

free parameters (4)
  • Diffusion goal bias (DGB) = 0.85
    Chosen as best in ablation over values 0, 15, 50, 85; controls probability of conditioning on random target instead of goal.
  • Propagation duration N = 64
    Fixed value selected from ablation over 32, 64, 128, scheduled, and random strategies.
  • Denoising iterations = 1
    Single-step flow matching selected as best in ablation over 1, 2, 4, 8, 16 iterations.
  • AntMaze nearest-neighbor distance metric = (x,y) only
    Hand-chosen to encourage exploration and avoid being overshadowed by the other 27 state dimensions.
assumptions (6)
  • standard math The system dynamics f are Lipschitz continuous and a robust (positive-clearance) solution exists.
    Standard assumptions in RRT PC proofs, invoked in Theorem 1 and Appendix A.1 (following [56]).
  • domain assumption The learned flow/diffusion sampler has full support over action sequences.
    Needed for Theorem 1; attempted proof in Appendix A.3, but the proof only establishes full support on the image of the flow map, not on all of R^D.
  • ad hoc to paper The learned vector field is globally Lipschitz and yields a unique global C^1 flow map that is a diffeomorphism onto R^D.
    Implicit in Appendix A.3.1 to conclude full support; not verified for practical neural network samplers.
  • domain assumption The dynamic model f accurately represents the real robot.
    Needed for 'provably-safe' and collision-free execution claims; the paper itself notes in Limitations (Sec. 7) that the real-world system uses an approximate model without system identification.
  • domain assumption Full knowledge of obstacle geometry is available at planning time.
    Stated in Sec. 4.1 and Limitations; the local occupancy grid is extracted from a known map.
  • domain assumption The planner has access to a reliable forward simulator (step function) of the robot dynamics.
    The tree propagates actions through the dynamic model; for the ant, they use the environment's step() function (Appendix C).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Train-Once Plan-Anywhere Kinodynamic Motion Planning via Diffusion Trees." pith.science (2026). https://pith.science/paper/HCTFN23M

@misc{pith2026250821001,
  author       = {Pith},
  title        = {Pith review of: Train-Once Plan-Anywhere Kinodynamic Motion Planning via Diffusion Trees},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCTFN23M}},
  note         = {Machine review of arXiv:2508.21001}
}
read the original abstract

Kinodynamic motion planning is concerned with computing collision-free trajectories while abiding by the robot's dynamic constraints. This critical problem is often tackled using sampling-based planners (SBPs) that explore the robot's high-dimensional state space by constructing a search tree via action propagations. Although SBPs can offer global guarantees on completeness and solution quality, their performance is often hindered by slow exploration due to uninformed action sampling. Learning-based approaches can yield significantly faster runtimes, yet they fail to generalize to out-of-distribution (OOD) scenarios and lack critical guarantees, e.g., safety, thus limiting their deployment on physical robots. We present Diffusion Tree (DiTree): a provably-generalizable framework leveraging diffusion policies (DPs) as informed samplers to efficiently guide state-space search within SBPs. DiTree combines DP's ability to model complex distributions of expert trajectories, conditioned on local observations, with the completeness of SBPs to yield provably-safe solutions within a few action propagation iterations for complex dynamical systems. We demonstrate DiTree's power with an implementation combining the popular RRT planner with a DP action sampler trained on a single environment. In comprehensive evaluations on OOD scenarios, DiTree achieves on average a 30% higher success rate compared to standalone DP or SBPs, on a dynamic car and Mujoco's ant robot settings (for the latter, SBPs fail completely). Beyond simulation, real-world car experiments confirm DiTree's applicability, demonstrating superior trajectory quality and robustness even under severe sim-to-real gaps. Project webpage: https://sites.google.com/view/ditree.

Figures

Figures reproduced from arXiv: 2508.21001 by the authors.

Figure 1
Figure 1. Visualization of DiTree on D4RL’s AntMaze setting in a random maze [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Action sampling in DiTree: (Left) A candidate node is selected for expansion and a local [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Experiment scenarios: We train on a single map, D4RL [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (Left and middle) Average success rate vs. search time across all test scenarios for Car [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of Success Rate vs Runtime for different ablation tests. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: A schematic of the ant robot, borrowed from [ [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Real car tracking results for DiTree. All axes are expressed in meters. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Real car tracking results for RRT, where collisions are marked by a yellow ’X’. Only [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Neural Navigation Functions for Zero-Shot Generalizable Motion Planning

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    Neural-NF learns a mapping from intrinsic Laplacian features to local PDE coefficients whose solution yields a collision-free, monotonically descending navigation function with global goal minimum by construction, ach...

Reference graph

Works this paper leans on

85 extracted references · 57 canonical work pages · cited by 1 Pith paper

  1. [1]

    Paden, M

    B. Paden, M. C ´ap, S. Z. Yong, D. S. Yershov, and E. Frazzoli. A survey of motion planning and control techniques for self-driving urban vehicles. IEEE Transactions on Intelligent Vehicles, 1:33–55, 2016. URL https://api.semanticscholar.org/CorpusID:1229096

    2016

  2. [2]

    Ghambari, M

    S. Ghambari, M. Golabi, L. Vermeulen-Jourdan, J. Lepagnot, and L. Idoumghar. UA V path planning techniques: a survey. RAIRO Oper. Res., 58:2951–2989, 2024. URL https://api. semanticscholar.org/CorpusID:268718404

    2024

  3. [3]

    C. R. de Lima, S. G. Khan, M. Tufail, S. H. Shah, and M. R. O. A. Maximo. Humanoid robot motion planning approaches: a survey. Journal of Intelligent & Robotic Systems , 110(2):86,

  4. [4]

    https://underactuated.csail. mit.edu

    R. ”Tedrake. ”Underactuated Robotics”. 2023. URL "https://underactuated.csail. mit.edu"

    2023

  5. [5]

    S. M. LaValle. Planning Algorithms. Cambridge University Press, Cambridge, UK, 2006

    2006

  6. [6]

    R. S. Sutton. The bitter lesson. urlhttp://www.incompleteideas.net/IncIdeas/BitterLesson.html, 2019

    work page 2019

  7. [7]

    Elbanhawi and M

    M. Elbanhawi and M. Simic. Sampling-based robot motion planning: A review. Ieee access, 2:56–77, 2014. 9

    work page 2014

  8. [8]

    Janner, Y

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine. Planning with diffusion for flexible behavior synthesis. In International Conference on Machine Learning, 2022

    work page 2022

Show all 85 references
  1. [9]

    S. Levine. Reinforcement learning and control as probabilistic inference: Tutorial and review,

  2. [10]

    Silver, J

    D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y . Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and D. Hassabis. Mastering the game of go without human knowledge. Nature, 550:354–35...

    2017

  3. [11]

    Silver, T

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general rein- forcement learning algorithm that masters chess, shogi, and go through self-play. Sci- ence, 362(64...

    2018

  4. [12]

    Openai o1 system card, 2024

    OpenAI. Openai o1 system card, 2024. URL https://openai.com/index/ openai-o1-system-card/

    2024

  5. [13]

    Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2025. URL https://arxiv.org/abs/2501.12948

    2025 arXiv

  6. [14]

    Kavraki, P

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

    1996

  7. [15]

    Bohlin and L

    R. Bohlin and L. Kavraki. Path planning using lazy prm. In Proceedings 2000 ICRA. Mil- lennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), volume 1, pages 521–528 vol.1, 2000

    2000

  8. [16]

    Karaman and E

    S. Karaman and E. Frazzoli. Sampling-based algorithms for optimal motion planning. The International Journal of Robotics Research , 30:846 – 894, 2011. URL https://api. semanticscholar.org/CorpusID:14876957

    2011

  9. [17]

    D. J. Webb and J. van den Berg. Kinodynamic rrt*: Optimal motion planning for systems with linear differential constraints. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2013

    2013

  10. [18]

    H. B. Keller. Numerical methods for two-point boundary-value problems . Courier Dover Publications, 2018

    2018

  11. [19]

    S. M. LaValle and J. J. Kuffner. Randomized kinodynamic planning. I. J. Robotics Res. , 20 (5):378–400, 2001

    2001

  12. [20]

    Y . Li, Z. Littlefield, and K. E. Bekris. Asymptotically optimal sampling-based kinodynamic planning. Int. J. Rob. Res., 35(5):528–564, Apr. 2016. URL https://doi.org/10.1177/ 0278364915614386

    2016

  13. [21]

    Hauser and Y

    K. Hauser and Y . Zhou. Asymptotically optimal planning by feasible kinodynamic planning in a state–cost space. IEEE Transactions on Robotics, 32(6):1431–1443, 2016

    2016

  14. [22]

    I. A. S ¸ucan and L. E. Kavraki. Kinodynamic Motion Planning by Interior-Exterior Cell Exploration, pages 449–464. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010. URL https://doi.org/10.1007/978-3-642-00312-7_28

    2010 doi

  15. [23]

    Kleinbort, E

    M. Kleinbort, E. Granados, K. Solovey, R. Bonalli, K. E. Bekris, and D. Halperin. Refined analysis of asymptotically-optimal kinodynamic planning in the state-cost space. 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6344–6350, 2019. URL https://a...

    2020

  16. [24]

    M. Fu, K. Solovey, O. Salzman, and R. Alterovitz. Resolution-optimal motion planning for steerable needles. In 2022 International Conference on Robotics and Automation (ICRA) , pages 9652–9659, 2022

    2022

  17. [25]

    G. P. d. Carvalho, T. Sawanobori, and T. Horii. Data-driven motion planning: A survey on deep neural networks, reinforcement learning, and large language model approaches. IEEE Access, 13:52195–52245, 2025

    2025

  18. [26]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, 2024

    2024

  19. [27]

    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. In 2024 IEEE International Conference on Robotics and Automation (ICRA), page 10351–10358. IEEE, May 2024. URL http://dx.doi...

    2024

  20. [28]

    Carvalho, A

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

    2023

  21. [29]

    Xiao, T.-H

    W. Xiao, T.-H. Wang, C. Gan, R. Hasani, M. Lechner, and D. Rus. Safediffuser: Safe planning with diffusion probabilistic models. In International Conference on Learning Representations, 2025

    2025

  22. [30]

    Y . Luo, C. Sun, J. B. Tenenbaum, and Y . Du. Potential based diffusion motion planning. In International Conference on Machine Learning, pages 33486–33510. PMLR, 2024

    2024

  23. [31]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learn- ing using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256–2265. pmlr, 2015

    2015

  24. [32]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  25. [33]

    Song and S

    Y . Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  26. [34]

    Lipman, R

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling, 2023. URL https://arxiv.org/abs/2210.02747

    2023 arXiv

  27. [35]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022

    2022 arXiv

  28. [36]

    M. Seo, Y . Cho, Y . Sung, P. Stone, Y . Zhu, and B. Kim. Presto: Fast motion planning using dif- fusion models based on key-configuration environment representation. In IEEE International Conference on Robotics and Automation (ICRA), 2024

    2024

  29. [37]

    K. Lee, S. Kim, and J. Choi. Refining diffusion planner for reliable behavior synthesis by au- tomatic detection of infeasible plans. In Advances in Neural Information Processing Systems, 2023

    2023

  30. [38]

    Huang, Z

    S. Huang, Z. Wang, P. Li, B. Jia, T. Liu, Y . Zhu, W. Liang, and S.-C. Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. arXiv preprint arXiv:2301.06015, 2023

    2023 arXiv

  31. [39]

    L. Feng, P. Gu, B. An, and G. Pan. Resisting stochastic risks in diffusion planners with the trajectory aggregation tree. In International Conference on Machine Learning , volume 235, pages 13175–13198. PMLR, 2024. 11

    2024

  32. [40]

    C. Chen, F. Deng, K. Kawaguchi, C. Gulcehre, and S. Ahn. Simple hierarchical planning with diffusion. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=kXHEBK9uAY

    2024

  33. [41]

    Z. Dong, J. Hao, Y . Yuan, F. Ni, Y . Wang, P. Li, and Y . Zheng. Diffuserlite: Towards real-time diffusion planning, 2024. URL https://arxiv.org/abs/2401.15443

    2024 arXiv

  34. [42]

    J. Liu, M. Stamatopoulou, and D. Kanoulas. Dipper: Diffusion-based 2d path planner applied on legged robots. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 9264–9270. IEEE, 2024

    2024

  35. [43]

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

    2019

  36. [44]

    Ichter, J

    B. Ichter, J. Harrison, and M. Pavone. Learning sampling distributions for robot motion plan- ning. 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 7087– 7094, 2017

    2018

  37. [45]

    J. Wang, W. Chi, C. Li, C. Wang, and M. Q. Meng. Neural rrt*: Learning-based optimal path planning. IEEE Transactions on Automation Science and Engineering , 17:1748–1758, 2020. URL https://api.semanticscholar.org/CorpusID:215841350

    2020

  38. [46]

    B. Chen, B. Dai, Q. Lin, G. Ye, H. Liu, and L. Song. Learning to plan in high dimensions via neural exploration-exploitation trees. In International Conference on Learning Represen- tations, 2019. URL https://api.semanticscholar.org/CorpusID:214300994

    2019

  39. [47]

    Ichter, P

    B. Ichter, P. Sermanet, and C. Lynch. Broadly-exploring, local-policy trees for long- horizon task planning. In Conference on Robot Learning , 2020. URL https://api. semanticscholar.org/CorpusID:222310776

    2020

  40. [48]

    Sivaramakrishnan, E

    A. Sivaramakrishnan, E. Granados, S. Karten, T. McMahon, and K. E. Bekris. Improving kinodynamic planners for vehicular navigation with learned goal-reaching controllers. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 9038– 9043, 2021. ...

    2021

  41. [49]

    McMahon, A

    T. McMahon, A. Sivaramakrishnan, K. Kedia, E. Granados, and K. Bekris. Terrain-aware learned controllers for sampling-based kinodynamic planning over physically simulated ter- rains. pages 2925–2930, 10 2022

    2022

  42. [50]

    H.-T. L. Chiang, J. Hsu, M. Fiser, L. Tapia, and A. Faust. Rl-rrt: Kinodynamic motion planning via learning reachability estimators from rl policies. IEEE Robotics and Automation Letters , 4:4298–4305, 2019. URL https://api.semanticscholar.org/CorpusID:195874334

    2019

  43. [51]

    J. J. Kuffner and S. M. LaValle. RRT-connect: An efficient approach to single-query path planning. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065) , volume 2, pages 995–1001. I...

    2000

  44. [52]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024

    2024

  45. [53]

    A. J. Bose, T. Akhound-Sadegh, G. Huguet, K. Fatras, J. Rector-Brooks, C.-H. Liu, A. C. Nica, M. Korablyov, M. Bronstein, and A. Tong. Se (3)-stochastic flow matching for protein backbone generation. arXiv preprint arXiv:2310.02391, 2023

    2023 arXiv

  46. [54]

    X. Hu, B. Liu, X. Liu, and Q. Liu. Adaflow: Imitation learning with variance-adaptive flow- based policies. ArXiv, abs/2402.04292, 2024. URL https://api.semanticscholar.org/ CorpusID:267523297. 12

    2024 arXiv

  47. [55]

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems (RSS), 2024

    2024

  48. [56]

    Kleinbort, K

    M. Kleinbort, K. Solovey, Z. Littlefield, K. E. Bekris, and D. Halperin. Probabilistic com- pleteness of RRT for geometric and kinodynamic planning with forward propagation. IEEE Robotics and Automation Letters, 4:277–283, 2018. URLhttps://api.semanticscholar. org/CorpusID:52303415

    2018

  49. [57]

    L. Yang, Z. Zhang, Z. Zhang, X. Liu, M. Xu, W. Zhang, C. Meng, S. Ermon, and B. Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv preprint arXiv:2407.02398, 2024

    2024 arXiv

  50. [58]

    Frans, D

    K. Frans, D. Hafner, S. Levine, and P. Abbeel. One step diffusion via shortcut models. ArXiv, abs/2410.12557, 2024. URL https://api.semanticscholar.org/CorpusID: 273375140

    2024 arXiv

  51. [59]

    Dasari, O

    S. Dasari, O. Mees, S. Zhao, M. K. Srirama, and S. Levine. The ingredients for robotic dif- fusion transformers. ArXiv, abs/2410.10088, 2024. URL https://api.semanticscholar. org/CorpusID:273346315

    2024 arXiv

  52. [60]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, editors,Medical Im- age Computing and Computer-Assisted Intervention – MICCAI 2015 , pages 234–241, Cham,

    2015

  53. [61]

    J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl. CasADi - A software framework for nonlinear optimization and optimal control. Mathematical Programming Com- putation, (11):1–36, Mar. 2019

    2019

  54. [62]

    Kloeser, T

    D. Kloeser, T. Schoels, T. Sartor, A. Zanelli, G. Prison, and M. Diehl. Nmpc for racing using a singularity-free path-parametric model with obstacle avoidance. IFAC-PapersOnLine, 53(2): 14324–14329, 2020. URL https://www.sciencedirect.com/science/article/pii/ S2405896320317845

    2020

  55. [63]

    Todorov, T

    E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–5033. IEEE, 2012

    2012

  56. [64]

    de Lazcano, K

    R. de Lazcano, K. Andreas, J. J. Tai, S. R. Lee, and J. Terry. Gymnasium robotics, 2024. URL http://github.com/Farama-Foundation/Gymnasium-Robotics

    2024

  57. [65]

    Minari: A dataset API for Offline Reinforcement Learning, May 2023

    Minari Contributors. Minari: A dataset API for Offline Reinforcement Learning, May 2023

    2023

  58. [66]

    I. A. S ¸ucan, M. Moll, and L. E. Kavraki. The Open Motion Planning Library. IEEE Robotics & Automation Magazine, 19(4):72–82, December 2012. https://ompl.kavrakilab.org

    2012

  59. [67]

    Zheng, R

    Y . Zheng, R. Liang, K. ZHENG, J. Zheng, L. Mao, J. Li, W. Gu, R. Ai, S. E. Li, X. Zhan, and J. Liu. Diffusion-based planning for autonomous driving with flexible guidance. In The Thirteenth International Conference on Learning Representations , 2025. URL https: //openreview.n...

    2025

  60. [68]

    Liang, C

    T. Liang, C. J. Glossner, L. Wang, and S. Shi. Pruning and quantization for deep neu- ral network acceleration: A survey. ArXiv, abs/2101.09671, 2021. URL https://api. semanticscholar.org/CorpusID:231699188

    2021 arXiv

  61. [69]

    Kleinbort, O

    M. Kleinbort, O. Salzman, and D. Halperin. Collision detection or nearest-neighbor search? On the computational bottleneck in sampling-based motion planning. In Workshop on the Algorithmic Foundations of Robotics, pages 624–639. Springer, 2016. 13

    2016

  62. [70]

    Elhafsi, B

    A. Elhafsi, B. Ivanovic, L. Janson, and M. Pavone. Map-predictive motion planning in un- known environments. In IEEE International Conference on Robotics and Automation , pages 8552–8558. IEEE, 2020

    2020

  63. [71]

    Koenig and M

    S. Koenig and M. Likhachev. Fast replanning for navigation in unknown terrain. IEEE Trans- actions on Robotics, 21(3):354–363, 2005

    2005

  64. [72]

    A. Zeng, S. Song, J. Lee, A. Rodriguez, and T. Funkhouser. Tossingbot: Learning to throw arbitrary objects with residual physics. IEEE Transactions on Robotics , 36(4):1307–1319, 2020

    2020

  65. [73]

    Kaufmann, L

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. M ¨uller, V . Koltun, and D. Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620(7976):982–987, 2023

    2023

  66. [74]

    Ghasemi, A

    M. Ghasemi, A. H. Moosavi, and D. Ebrahimi. A comprehensive survey of reinforcement learning: From algorithms to practical challenges, 2025. URL https://arxiv.org/abs/ 2411.18892

    2025 arXiv

  67. [75]

    C. H. Huang, P. Jadhav, B. Plancher, and Z. Kingston. prrtc: Gpu-parallel rrt-connect for fast, consistent, and low-cost motion planning, 2025. URLhttps://arxiv.org/abs/2503. 06757

    2025

  68. [76]

    Thomason, Z

    W. Thomason, Z. Kingston, and L. E. Kavraki. Motions in microseconds via vectorized sampling-based planning. In IEEE International Conference on Robotics and Automation , pages 8749–8756. URL http://arxiv.org/abs/2309.14545

    arXiv

  69. [77]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  70. [78]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score- based generative modeling through stochastic differential equations. In International Con- ference on Learning Representations , 2021. URL https://openreview.net/forum?id= PxTIG12RRHS

    2021

  71. [79]

    V . I. Arnold. Ordinary Differential Equations . Springer, 3rd edition, 1992. ISBN 9783540548133

    1992

  72. [80]

    Nechushtan, B

    O. Nechushtan, B. Raveh, and D. Halperin. Sampling-diagram automata: A tool for analyzing path quality in tree planners. In International Workshop on the Algorithmic Foundations of Robotics, volume 68 ofSpringer Tracts in Advanced Robotics, pages 285–301. Springer, 2010

    2010

  73. [81]

    Towers, A

    M. Towers, A. Kwiatkowski, J. Terry, J. U. Balis, G. De Cola, T. Deleu, M. Goul ˜ao, A. Kallinteris, M. Krimmel, A. KG, et al. Gymnasium: A standard interface for reinforce- ment learning environments. arXiv preprint arXiv:2407.17032, 2024

    2024 arXiv

  74. [82]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey,˙I. Polat, Y . Feng...

    2020

  75. [2015]

    ISBN 978-3-319-24574-4

    Springer International Publishing. ISBN 978-3-319-24574-4

  76. [2018]

    URL https://arxiv.org/abs/1805.00909

    arXiv

  77. [2024]

    URL https://doi.org/10.1007/s10846-024-02117-z

    doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.