Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Growing Trees with an Agent: Accelerating RRTs with Learned, Multi-Step Episodic Exploration

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing the random-point primitive in RRT with a learned, multi-step exploration episode turns diffuse tree growth into directed branches and reports large gains in speed, success rate, and collision-check count…

desk verdict The episodic primitive is genuinely new, but the paper never shows it is what causes the speedups: the One Step Jump is necessary for success and no control is run. read the letter →

arxiv 2507.06605 v2 pith:AL4TFGNP submitted 2025-07-09 cs.RO

Xinyu Wu This is my paper
classification cs.RO
keywords motionplanningRRTdeepreinforcementlearningSoftActor-Criticepisodicexplorationsampling-basedcollisioncheckinghigh-dimensional
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 is trying to establish a new primitive for sampling-based motion planning: instead of drawing one random configuration and attempting to connect to it, a deep reinforcement learning agent proposes a short, multi-step path segment, called an episodic exploration, which then becomes the unit of tree growth. The author argues that this changes RRT's behavior from a diffuse volumetric search into a directed branch-like expansion, countering the curse of dimensionality, reducing expensive collision checks, and improving connectivity. The reported evidence is across 100 randomized trials per setting in 2D, 3D, and a 6D robotic arm, where ERRT and its variants outperform RRT, RRT*, and RRT-Connect on success rate, computation time, collision checks, and initial path length; the 6D results show the largest gap, with a 98% versus 19% success rate. The paper also claims that the asymptotically optimal variant ERRT* converges toward optimality much faster than RRT*, and that a policy trained entirely in simulation transfers to a physical arm. If this holds, learned episodic exploration is a general drop-in replacement for the sampling primitive in RRT-family planners, not merely a heuristic bias on where to sample.

What carries the argument

The load-bearing object is the exploratory episode, defined as a fixed-length sequence of $m$ configuration-space waypoints that the policy proposes in one shot as a locally plausible path segment. It is produced by a Transformer encoder that attends to the agent's self-state, goal-relative state, and local oriented bounding boxes or spheres representing obstacles; a max-pooling layer collapses the token sequence into a fixed vector, and a feedforward network decodes it into the waypoint sequence. The raw output is converted through a cubic B-spline and arc-length re-parameterization into dense, uniformly spaced knot points, then validated by a dynamic bisection routine that searches logarithmically for the furthest collision-free point on the episode and resets the search upper bound whenever the tree grows. A one-step jump mechanism links the tree to the goal when an episode ends near it, and randomized restarts preserve probabilistic completeness. This object carries the argument because it replaces the unit of growth from a point to a connected, locally collision-free segment, which is what reduces collision-check counts and yields shorter initial paths.

What would settle it

Evaluate the frozen policy on obstacle layouts drawn from a distribution different from its training set, such as far denser clutter or obstacle shapes never seen in training, and compare per-trial success rate and computation time against RRT; if ERRT's advantage collapses while RRT's does not, the episodic-exploration gain is distribution-specific rather than a general planning primitive.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that a trained DRL policy can take over exploration inside an RRT and that this changes the growth pattern of the tree. The agent outputs a coherent sequence of waypoints, which after re-sampling and collision validation is attached to the nearest tree node as an extended branch; episodic exploration, randomized restart, adaptive noise, and a one-step jump to the goal replace the sample-connect loop. The paper reports that ERRT, ERRT*, and ERRT-Connect consistently beat their classical counterparts in 2D, 3D, and 6D benchmarks: in a 6D arm scenario, ERRT reaches a 98% success rate versus 19% for RRT, is up to 107 times faster, performs over 99.6% fewer collision checks, and finds initial paths almost 50% shorter. For the optimal variant, ERRT* reaches a path within 10% of the optimal cost 29 times faster than RRT* in 3D, and the planner generalizes to physical hardware with a simulation-trained policy.

Load-bearing premise

The central assumption is that a policy trained on the same obstacle distributions used in the benchmarks continues to generate useful exploratory episodes in new scenes of that class, while the cost of training the policy is not counted in the reported planning times.

Editorial extensions

If this is right

  • RRT-family planners can be accelerated by swapping the sample primitive for a learned episode while keeping the tree structure and its completeness arguments.
  • In high-dimensional or cluttered spaces, the collision-check count falls by orders of magnitude, so the method targets the dominant computational bottleneck of sampling-based planning.
  • Initial paths found by ERRT and ERRT-Connect are shorter than their classical counterparts even before any optimization, which matters for energy-limited robots.
  • ERRT*'s faster anytime convergence means a time-limited planner can return a near-optimal path sooner, useful for real-time or online applications.
  • Because the policy transfers from simulation to a physical arm without GPU acceleration at runtime, the approach is deployable on real hardware.

Reading between the lines

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

  • The reported planning times exclude the cost of training the policy, so a fair single-query comparison would need to include it; in settings where the same policy is reused across many planning queries, the amortized cost shrinks and the advantage grows.
  • Because the policy is trained and tested on the same obstacle distributions, the decisive test of generality is out-of-distribution transfer; a benchmark on unseen clutter densities and obstacle shapes would separate a general primitive from a distribution-specific prior.
  • The same episodic primitive could plausibly be plugged into other sampling-based planners such as PRM or BIT*, and the natural prediction is that the collision-check reduction transfers beyond RRT-style trees.
  • Extending episodes to kinodynamic constraints, where waypoint sequences must respect turning radii and acceleration limits, is the paper's stated next step and a direct test of whether the primitive generalizes to dynamics-aware planning.
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 Episodic RRT (ERRT), a hybrid motion planner that replaces the random sample primitive of RRT-family planners with multi-step path segments ('episodes') generated by a Soft Actor-Critic policy. The authors present variants ERRT, ERRT*, and ERRT-Connect, and report large improvements over OMPL baselines in 2D, 3D, and 6D environments: higher success rates, lower computation time, fewer collision checks, and shorter initial paths, plus faster anytime convergence for ERRT*. A physical UR5e deployment is included. The central empirical claim is that a learned, multi-step exploratory episode is a generally superior primitive for sampling-based planning.

Significance. If the empirical claims hold, the paper offers a genuinely different design point for learning-based planning: instead of biasing samples or steering functions, the learned policy generates whole connected path segments, which naturally reduces collision checks and improves connectivity. The manuscript provides reproducible code, a real-robot demonstration, and an ablation study, which are commendable. However, the significance is currently tempered by a load-bearing attribution problem: the paper's own ablation shows that the single-tree methods fail entirely without the non-learned One Step Jump, so the headline gains cannot be cleanly credited to the learned episodes without additional control experiments. The evaluation also has methodological weaknesses (success-conditioned averages, missing baselines, no out-of-distribution tests) that affect the strength of the quantitative claims.

major comments (4)
  1. [Section V-D, Table II] The NoJump ablation shows that ERRT and ERRT* achieve '×' (complete failure) in every environment when the One Step Jump is disabled. Because Algorithm 2 (lines 18-22) sets qnew directly to qgoal whenever the tree comes within α_jump of the goal, this mechanism is functionally a hard-coded terminal goal bias. The paper's claim that the learned episodes drive the 98% success and 107x speedup in 6D therefore requires a control baseline in which vanilla RRT (and RRT*) are given the same One Step Jump capability. Without such a baseline, the observed gains cannot be attributed to the learned episodic exploration as opposed to the goal-reaching shortcut. This is the central attribution issue and must be addressed.
  2. [Section V-A and Table I] All performance metrics (TIME, COL, LEN) are averaged over successful trials only, while success rates differ by large margins (e.g., 6D: RRT 19% vs ERRT 98%; 3D: RRT* 75% vs ERRT* 100%). A baseline that succeeds on only the easiest 19% of trials will naturally show lower average time and collision checks on those trials, confounding the efficiency comparison with environment difficulty. The paper should report metrics computed over all trials (counting failures as infinite time or using time-to-success with censoring), or provide paired per-trial comparisons, along with error bars or confidence intervals.
  3. [Section V-B and Section IV-A] The evaluation lacks comparisons to learned sampling-bias baselines (e.g., Neural RRT*, DeepSMP) or to classical goal-biased RRT with a comparable goal-biasing probability. The paper's positioning (Section II-C) explicitly contrasts ERRT with sampling-bias methods, yet no such method is included in Table I. As a result, the evidence cannot distinguish the benefit of multi-step episodic generation from the benefit of simply directing samples toward the goal. Adding at least one learned-bias baseline and one goal-bias-controlled RRT is necessary to support the claim that the episodic primitive, rather than goal-directedness, is the source of the advantage.
  4. [Section IV-B and Section V] The policy is trained on the same classes of obstacle distributions used for benchmarking, and the cost of training is not amortized anywhere. The reported speedups are therefore for a planner that has already received distribution-specific prior knowledge. The paper should either report out-of-distribution performance on environments not seen in training (e.g., different obstacle densities, different maze structures) or explicitly restrict the claim to the trained distribution. Without this, the 'general purpose' framing in the abstract and introduction overstates the transferability of the results.
minor comments (5)
  1. [Table I] The column header 'ACC' is undefined; based on the values (1.7X, 9.4X, etc.) it appears to denote the speedup factor of the ERRT variant over its baseline, but this should be stated explicitly in the caption or text.
  2. [Section V-C] Figure 4 reports speedup factors but provides no error bars or statistical significance tests, and the number of trials per environment is only stated indirectly. Please add variance information or at least state the sample size and spread.
  3. [Section V-D, Table II] The ablation table reports changes in success rate and time but not in path length or collision checks; since the paper's claims include path-length improvements, the ablation should also show whether these quality metrics are affected by the NoJump, NoBisection, and Downsample modifications.
  4. [Throughout the manuscript] There are several typographical errors that should be corrected: 'Concecutive' (Algorithm 2 comment), 'V alidation' (Section IV-C heading), 'UA Vs' (Section V-C), 'It’s' in the Algorithm 2 title, and a stray 'T he' in the abstract of the arXiv version.
  5. [Section IV-A3] The reward function (Eq. 2) includes weights α1, α2, α3, β1, β2, but the paper never reports the actual values used, nor the episode length m, the Incremental Bound constant, ddense, α_jump, or Lmax. Providing these parameter values in a table would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ERRT is an empirical planner evaluation with disclosed design components; attribution and generalization concerns are not definitional circularity.

full rationale

This paper does not contain a mathematical derivation whose conclusion is identified with its premises, nor does it fit a parameter and then rename that fit as a prediction. The central evidence is an empirical comparison of ERRT variants against OMPL baselines on measured metrics (success rate, time, collision checks, path length). The reward function in Eq. (2) does include goal-directed terms (r_reach and r_advance), and Algorithm 2 includes a terminal One Step Jump that sets qnew to qgoal when the tree is within alpha_jump. These are openly specified components of the method rather than hidden inputs disguised as findings. The paper's own ablation in Section V-D and Table II states that "Removing it entirely (NoJump) resulted in a complete failure for the single-tree ERRT and ERRT* methods across all environments." This is a genuine attribution and control limitation: the headline success and speed gains cannot be cleanly assigned to the learned episodic exploration without a baseline RRT given the same terminal jump, so the central claim is weaker than presented. However, this is a confound in interpretation, not circularity under the definitions used here: no measured outcome is equal to an input by construction, and no self-citation chain is load-bearing. Training on the same obstacle distributions used for benchmarking and excluding training cost are generalization/fairness risks, not circular steps. The empirical claims are self-contained measurements with disclosed methodology, so the appropriate circularity score is 0.

Assumptions & free parameters 7 free parameters · 3 assumptions · 1 invented entities

The central method rests on a hand-designed reward, a collection of constants such as episode length, bound, resampling density, and jump threshold, and a neural network trained by SAC whose weights are not released. The axioms are the unproved domain assumptions that the trained policy generalizes, that dense resampled paths faithfully represent collision status, and that randomized restarts preserve probabilistic completeness.

free parameters (7)
  • Reward weights alpha1, alpha2, alpha3, beta1, beta2 = not reported
    Hand-chosen coefficients in Eq. (2); no sensitivity analysis is provided for the reward balancing.
  • Action path length m = not reported
    Number of configuration points in each RL-generated episode; fixed before training, no justification of the value.
  • Incremental bound constant = not reported
    Limits each configuration point's displacement in the action path; controls action feasibility.
  • Resample density ddense = not reported
    Arc length between consecutive knot points after B-spline re-parameterization; ablation shows robustness but value is not given.
  • One Step Jump threshold alpha_jump = not reported; described as approximating max feasible policy step
    Determines when to jump directly to the goal; the ablation tests half the threshold but not the full sensitivity range.
  • Maximum episode length Lmax = not reported
    Controls episode termination and restart frequency; ablation only tests Lmax + 2.
  • SAC policy and Q-network weights = trained, not released
    The learned component that generates episodes; architecture, training data, and training hyperparameters are not specified.
assumptions (3)
  • domain assumption The SAC-trained policy generalizes to unseen random environments within each benchmark class.
    Training described in Section IV is followed by evaluation on newly generated layouts of the same scenario types; no out-of-distribution test is reported.
  • domain assumption Dense re-sampled knot points and bisection validation correctly determine the collision-free prefix of every continuous episode path.
    Algorithm 3 checks only discrete points; obstacles that protrude between knot points could be missed, and no formal argument is given.
  • ad hoc to paper Randomized restart preserves probabilistic completeness of the overall planner.
    Section IV-C asserts the random restart 'ensures probabilistic completeness' but provides no proof for the intertwined policy and tree growth.
invented entities (1)
  • Learned multi-step exploratory episode
    purpose: The fundamental tree-growth primitive replacing a random sample point.
    The episode is a methodological construct; its effectiveness is demonstrated only by the paper's own experiments in Table I and Table II, with no external benchmark or theoretical guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Growing Trees with an Agent: Accelerating RRTs with Learned, Multi-Step Episodic Exploration." pith.science (2026). https://pith.science/paper/AL4TFGNP

@misc{pith2026250706605,
  author       = {Pith},
  title        = {Pith review of: Growing Trees with an Agent: Accelerating RRTs with Learned, Multi-Step Episodic Exploration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AL4TFGNP}},
  note         = {Machine review of arXiv:2507.06605}
}
read the original abstract

Classical sampling-based motion planners like the RRTs suffer from inefficiencies, particularly in cluttered or high-dimensional spaces, due to their reliance on undirected, random sampling. This paper introduces the Episodic RRT, a novel hybrid planning framework that replaces the primitive of a random point with a learned, multi-step "exploratory episode" generated by a Deep Reinforcement Learning agent. By making the DRL agent the engine of exploration, ERRT transforms the search process from a diffuse, volumetric expansion into a directed, branch-like growth. This paradigm shift yields key advantages: it counters the curse of dimensionality with focused exploration, minimizes expensive collision checks by proactively proposing locally valid paths, and improves connectivity by generating inherently connected path segments. We demonstrate through extensive empirical evaluation across 2D, 3D, and 6D environments that ERRT and its variants consistently and significantly outperform their classical counterparts without any GPU acceleration. In a challenging 6D robotic arm scenario, ERRT achieves a 98% success rate compared to 19% for RRT, is up to 107x faster, reduces collision checks by over 99.6%, and finds initial paths that are nearly 50% shorter. Furthermore, its asymptotically optimal variant, ERRT*, demonstrates vastly superior anytime performance, refining solutions to near-optimality up to 29x faster than standard RRT* in 3D environments. Code: https://xinyuwuu.github.io/Episodic_RRT/.

Figures

Figures reproduced from arXiv: 2507.06605 by the authors.

Figure 1
Figure 1. Conceptual illustration of the ERRT framework. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Network Architecture and Data Flow. 1) State Space S: A state st ∈ S at time t is a composite vector st = (selft, goalt, envt). selft: The agent’s intrinsic state, such as its velocity vector for 2D/3D point-mass plan￾ning, or a concatenation of joint positions and velocities for a 6D robotic arm. goalt: The target configuration represented relative to the agent, calculated as qgoal − qt. envt: Local environmental c… view at source ↗
Figure 3
Figure 3. Visualization randomly generated environments. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Continuing Optimization Performance. checks each algorithm required to refine its initial path to within a series of quality thresholds of the optimal cost. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

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

    2006

  2. [2]

    A formal basis for the heuristic determination of minimum cost paths,

    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 , vol. 4, no. 2, pp. 100–107, 1968

    1968

  3. [3]

    Rapidly-exploring random trees: A new tool for path planning,

    S. M. LaValle et al. , “Rapidly-exploring random trees: A new tool for path planning,” 1998. [Online]. Available: https://msl.cs.illinois.edu/ lavalle/papers/Lav98c.pdf

    work page 1998

  4. [4]

    Probabilistic roadmaps for path planning in high-dimensional configuration spaces,

    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 Automation , vol. 12, no. 4, pp. 566–580, 1996

    work page 1996

  5. [5]

    Rapidly-exploring random trees: Progress and prospects,

    S. Lavalle and J. Kuffner, “Rapidly-exploring random trees: Progress and prospects,” Algorithmic and computational robotics: New direc- tions, 01 2000

    work page 2000

  6. [6]

    Learning-based near-optimal motion planning for intelligent vehicles with uncertain dynamics,

    Y . Lu, X. Zhang, X. Xu, and W. Yao, “Learning-based near-optimal motion planning for intelligent vehicles with uncertain dynamics,” IEEE Robotics and Automation Letters , vol. 9, no. 2, pp. 1532–1539, 2024

    work page 2024

  7. [7]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” 2018. [Online]. Available: https://arxiv.org/abs/1801.01290

    arXiv 2018

  8. [8]

    Optimal and efficient path planning for partially-known environments,

    A. Stentz, “Optimal and efficient path planning for partially-known environments,” in Proceedings of the 1994 IEEE International Con- ference on Robotics and Automation , 1994, pp. 3310–3317 vol.4

    work page 1994

Show all 33 references
  1. [9]

    Using interpolation to improve path planning: The field d* algorithm,

    D. Ferguson and A. Stentz, “Using interpolation to improve path planning: The field d* algorithm,” Journal of Field Robotics , vol. 23, no. 2, pp. 79–101, 2006

    2006

  2. [10]

    Lifelong planning a*,

    S. Koenig, M. Likhachev, and D. Furcy, “Lifelong planning a*,” Artificial Intelligence, vol. 155, no. 1, pp. 93–146, 2004

    2004

  3. [11]

    The jps pathfinding system,

    D. Harabor and A. Grastien, “The jps pathfinding system,” Proceed- ings of the International Symposium on Combinatorial Search , vol. 3, 08 2021

    2021

  4. [12]

    Improving jump point search,

    ——, “Improving jump point search,” Proceedings International Con- ference on Automated Planning and Scheduling, ICAPS, vol. 2014, pp. 128–135, 05 2014

    2014

  5. [13]

    Theta*: Any-angle path planning on grids,

    K. Daniel, A. Nash, S. Koenig, and A. Felner, “Theta*: Any-angle path planning on grids,” Journal Of Artificial Intelligence Research , vol. 39, pp. 533–579, Sept. 2010

    2010

  6. [14]

    Randomized kinodynamic planning,

    S. M. LaValle and J. J. KuffnerJr., “Randomized kinodynamic planning,” The International Journal of Robotics Research , vol. 20, no. 5, pp. 378–400, 2001. [Online]. Available: https://doi.org/10.1177/02783640122067453

    2001 doi

  7. [15]

    Sampling-based algorithms for optimal motion planning,

    S. Karaman and E. Frazzoli, “Sampling-based algorithms for optimal motion planning,” The International Journal of Robotics Research, vol. 30, no. 7, pp. 846–894, 2011. [Online]. Available: https://doi.org/10.1177/0278364911406761

    2011 doi

  8. [16]

    Rrt-connect: An efficient approach to single-query path planning,

    J. Kuffner and S. LaValle, “Rrt-connect: An efficient approach to single-query path planning,” in Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automa- tion. Symposia Proceedings (Cat. No.00CH37065) , vol. 2, 2000, pp. 995–1001 vol.2

    2000

  9. [17]

    Informed rrt*: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,

    J. D. Gammell, S. S. Srinivasa, and T. D. Barfoot, “Informed rrt*: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,” in 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems , 2014, pp. 2997–3004

    2014

  10. [18]

    Batch informed trees (bit*): Sampling-based optimal planning via the heuristically guided search of implicit random geometric graphs,

    J. Gammell, S. Srinivasa, and T. Barfoot, “Batch informed trees (bit*): Sampling-based optimal planning via the heuristically guided search of implicit random geometric graphs,” Proceedings - IEEE International Conference on Robotics and Automation , vol. 2015, pp. 3067–3074, 06 2015

    2015

  11. [19]

    Advanced bit* (abit*): Sampling- based planning with advanced graph-search techniques,

    M. P. Strub and J. D. Gammell, “Advanced bit* (abit*): Sampling- based planning with advanced graph-search techniques,” in 2020 IEEE International Conference on Robotics and Automation (ICRA) , 2020, pp. 130–136

    2020

  12. [20]

    Adaptively informed trees (ait*): Fast asymptotically optimal path planning through adaptive heuristics,

    ——, “Adaptively informed trees (ait*): Fast asymptotically optimal path planning through adaptive heuristics,” in 2020 IEEE International Conference on Robotics and Automation (ICRA) , 2020, pp. 3191– 3198

    2020

  13. [21]

    Pathrl: An end-to-end path generation method for collision avoidance via deep reinforcement learning,

    W. Yu, J. Peng, Q. Qiu, H. Wang, L. Zhang, and J. Ji, “Pathrl: An end-to-end path generation method for collision avoidance via deep reinforcement learning,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 9278–9284

    2024

  14. [22]

    Neural mp: A generalist neural motion planner,

    M. Dalal, J. Yang, R. Mendonca, Y . Khaky, R. Salakhutdinov, and D. Pathak, “Neural mp: A generalist neural motion planner,” 2024. [Online]. Available: https://arxiv.org/abs/2409.05864

    2024 arXiv

  15. [23]

    Prm-rl: Long-range robotic navigation tasks by combining reinforcement learning and sampling-based planning,

    A. Faust, K. Oslund, O. Ramirez, A. Francis, L. Tapia, M. Fiser, and J. Davidson, “Prm-rl: Long-range robotic navigation tasks by combining reinforcement learning and sampling-based planning,” in 2018 IEEE International Conference on Robotics and Automation (ICRA), 2018, pp. 5113–5120

    2018

  16. [24]

    Rl-rrt: Kinodynamic motion planning via learning reachability estimators from rl policies,

    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, vol. 4, no. 4, pp. 4298–4305, 2019

    2019

  17. [25]

    Mobile robot path plan- ning in dynamic environments through globally guided reinforcement learning,

    B. Wang, Z. Liu, Q. Li, and A. Prorok, “Mobile robot path plan- ning in dynamic environments through globally guided reinforcement learning,” IEEE Robotics and Automation Letters , vol. 5, no. 4, pp. 6932–6939, 2020

    2020

  18. [26]

    Learning-based motion planning in dynamic environments using gnns and temporal encoding,

    R. Zhang, C. Yu, J. Chen, C. Fan, and S. Gao, “Learning-based motion planning in dynamic environments using gnns and temporal encoding,” in Proceedings of the 36th International Conference on Neural Information Processing Systems , ser. NIPS ’22. Red Hook, NY , USA: Curran Ass...

    2022

  19. [27]

    Deeply informed neural sampling for robot motion planning,

    A. H. Qureshi and M. C. Yip, “Deeply informed neural sampling for robot motion planning,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2018, pp. 6582–6588

    2018

  20. [28]

    Learned critical probabilistic roadmaps for robotic motion planning,

    B. Ichter, E. Schmerling, T.-W. E. Lee, and A. Faust, “Learned critical probabilistic roadmaps for robotic motion planning,” in 2020 IEEE International Conference on Robotics and Automation (ICRA) , 2020, pp. 9535–9541

    2020

  21. [29]

    Motion planning networks: Bridging the gap between learning-based and classical motion planners,

    A. H. Qureshi, Y . Miao, A. Simeonov, and M. C. Yip, “Motion planning networks: Bridging the gap between learning-based and classical motion planners,” IEEE Transactions on Robotics , vol. 37, no. 1, pp. 48–66, 2021

    2021

  22. [30]

    Neural rrt*: Learning-based optimal path planning,

    J. Wang, W. Chi, C. Li, C. Wang, and M. Q.-H. Meng, “Neural rrt*: Learning-based optimal path planning,” IEEE Transactions on Automation Science and Engineering , vol. 17, no. 4, pp. 1748–1758, 2020

    2020

  23. [31]

    Neural informed rrt*: Learning-based path planning with point cloud state representations under admissible ellipsoidal constraints,

    Z. Huang, H. Chen, J. Pohovey, and K. Driggs-Campbell, “Neural informed rrt*: Learning-based path planning with point cloud state representations under admissible ellipsoidal constraints,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) , 2024, pp. 8742–8748

    2024

  24. [32]

    The Open Motion Planning Library,

    I. A. S ¸ucan, M. Moll, and L. E. Kavraki, “The Open Motion Planning Library,” IEEE Robotics & Automation Magazine , vol. 19, no. 4, pp. 72–82, December 2012, https://ompl.kavrakilab.org

    2012

  25. [33]

    A new approach to time-optimal path parameterization based on reachability analysis,

    H. Pham and Q. C. Pham, “A new approach to time-optimal path parameterization based on reachability analysis,” IEEE Transactions on Robotics, vol. 34, pp. 645 – 659, 06 2018

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.