Pith. sign in

REVIEW 2 major objections 5 minor 42 references

CoRL-MPPI steers MPPI sampling with a learned cooperative policy while preserving its theoretical safety guarantees.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 22:38 UTC pith:JCWRI5HR

load-bearing objection Useful hybrid with strong empirical gains, but the 'provably-safe' claim does not cover the actual weighted-average control; needs a re-derivation of the safety argument and matched baselines. the 2 major comments →

arxiv 2511.09331 v3 pith:JCWRI5HR submitted 2025-11-12 cs.RO cs.MA

CoRL-MPPI: Enhancing MPPI With Learnable Behaviours For Efficient And Provably-Safe Multi-Robot Collision Avoidance

Stepan Dergachev , Artem Pshenitsyn , Aleksandr Panov , Alexey Skrynnik , Konstantin Yakovlev This is my paper
classification cs.RO cs.MA
keywords multi-robot collision avoidancemodel predictive path integralreinforcement learningsampling-based controldecentralized navigationsafety guaranteesORCAdifferential-drive robots
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper claims that a reinforcement-learned policy can be embedded into the Model Predictive Path Integral (MPPI) controller as an additional proposal distribution, biasing random rollouts toward cooperative collision-avoidance maneuvers without weakening the algorithm's formal safety guarantees. The authors argue that standard MPPI samples control sequences from an uninformed Gaussian around the previous solution, which is inefficient in dense multi-robot settings where most random rollouts are uncooperative. Their hybrid, CoRL-MPPI, trains a decentralized policy via independent proximal policy optimization in simulation, then uses that policy to generate a parallel control sequence whose mean and covariance are safety-adjusted alongside the MPPI branch. Experiments in dense circle, mesh, and random scenarios show that CoRL-MPPI achieves near-100% success rates and roughly half the makespan of the MPPI-ORCA baseline, while remaining comparable in sparse settings. A sympathetic reader would care because the method offers a practical recipe for combining learned behavior with provable safety in decentralized multi-robot navigation.

Core claim

The central discovery is that MPPI's sampling distribution can be guided by a learned policy without sacrificing its theoretical collision-safety guarantees, provided the policy's output is treated as just another Gaussian proposal and both proposal distributions are passed through the same chance-constrained optimization step. Concretely, the method uses two predictive trajectories at each planning step: one from the previous MPPI solution and one from the RL policy. For the first Hsafe steps, the means and covariances of both Gaussian distributions are adjusted by solving a convex second-order cone program that enforces ORCA-derived linear velocity constraints with a user-specified probabi

What carries the argument

The key mechanism is a dual-proposal sampling scheme with safety-constrained distribution updates. At each control iteration, two Gaussian control distributions are maintained: one centered on the shifted previous MPPI solution, the other on the output of a pre-trained RL policy. For the safety horizon, each distribution's mean and covariance are adjusted by solving a convex optimization problem that minimizes deviation from the nominal parameters while enforcing probabilistic safety constraints derived from ORCA's velocity obstacles, so that with probability at least a threshold the sampled controls lie in the safe region. The adjusted distributions then feed MPPI's weighted averaging, pres

Load-bearing premise

The safety proof assumes that each robot predicts its neighbors' future positions by assuming they keep moving at their current constant velocity, and that each robot computes its safety constraints independently; in the real closed loop, all robots change velocity at the same time, so those predicted positions are systematically wrong and the proof does not account for the coupling.

What would settle it

Run a closed-loop experiment with two differential-drive robots heading directly toward each other, both controlled by CoRL-MPPI with zero actuation noise, and record the minimum inter-robot distance over many trials. If the constant-velocity prediction is used, the robots may plan to pass at a separation that appears safe under straight-line extrapolation but is not when both turn into the encounter; observing any trial where the distance falls below the guaranteed threshold (or collisions occur at a rate exceeding the specified δ) would falsify the claim that theoretical guarantees are prese

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, learned policies and formal safety guarantees are compatible in decentralized multi-robot control, not mutually exclusive.
  • The safety-constrained update is not tied to a specific learned policy; any proposal distribution that outputs a Gaussian over controls could be substituted, suggesting a modular design.
  • Dense, symmetric scenarios that cause deadlocks in classical methods (circles, grids) become tractable, with potential application to warehouse logistics and swarm navigation.
  • Because MPPI's weighted average can down-weight poor rollouts, the learned policy need not be perfect; this may explain the method's generalization to unseen numbers of agents.
  • The probabilistic guarantee is explicit: the safety threshold δ is close to but not exactly one, so occasional rare collisions are expected and bounded by design.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 'provably safe' claim inherits the constant-velocity neighbor prediction model (Eq. 10); in a closed loop where every robot runs the same controller, the predicted neighbor positions are systematically wrong, so extending the guarantee to the coupled system would require propagating the policy's own predicted trajectories or a joint uncertainty model.
  • The 30%/70% split between RL-guided and MPPI rollouts is presented as a fixed choice; an adaptive mixing ratio based on online rollout scores or a learned gating function could further improve sample efficiency.
  • The safety adjustment relies on ORCA's reciprocity assumption that both agents share avoidance responsibility; testing against non-reciprocal or adversarial neighbors would empirically bound when the guarantee degrades.
  • The policy was trained on 32-agent scenarios; the paper shows graceful scaling to 50 agents, but the upper limit of the guarantee and the policy's generalization to much denser swarms remain untested.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces CoRL-MPPI, a hybrid controller that uses a pretrained RL policy to bias the sampling distribution of MPPI for decentralized multi-robot collision avoidance. Two sampling branches are maintained: one driven by the RL policy and one by the previous MPPI solution. The means and covariances of both Gaussian proposal distributions are adjusted through ORCA-based chance constraints taken from the authors' prior work, and the executed control is the MPPI-style weighted average of all sampled rollouts. The authors claim that CoRL-MPPI preserves all theoretical guarantees of regular MPPI and report simulation results showing higher success rates and lower makespans than ORCA-DD, B-UAVC, and MPPI-ORCA in Circle, Mesh (Dense), and Random scenarios.

Significance. If the theoretical claims were supported, the paper would make a useful contribution: it demonstrates a computationally feasible way to inject learned cooperative priors into a sampling-based MPC framework while retaining a safety-constrained sampling distribution. The empirical study is comparatively broad: four scenario families, varying agent counts, a learned baseline, and a multi-agent MPPI baseline, with reported success rates, collision rates, and makespans. The fast IPPO training pipeline and the C++/ONNX implementation are also practical strengths. However, the two main theoretical pillars of the paper are not established: the safety guarantee is only given for individual samples, not for the control actually executed, and the MPPI weighting formula is not re-derived for a mixture of two proposal distributions. These are central to the abstract's 'provably-safe' and 'preserves all theoretical guarantees' claims, so the manuscript in its current form cannot be accepted without substantial revision.

major comments (2)
  1. [Sec. IV-B / Algorithm 1 (lines 16–22)] The safety guarantee in Eq. (7) applies to individual samples from the adjusted Gaussian distributions. The control actually executed, u*_0 = Σ_k ω_k u^k_0, is a cost-weighted average of K samples. Since the ORCA constraints are linear, u*_0 is guaranteed safe only if every sampled control is safe. If the per-sample safe probability is at least δ_c, then, assuming independent samples, P(all K samples safe) ≥ δ_c^K; for K=1500 and δ_c=0.99 this is about 3×10^-7, so the individual-sample guarantee gives essentially no usable bound for the executed control. No theorem bounds the probability that u*_0 violates the constraints. In addition, the expression δ_c = |A_i| δ_ν δ_u can exceed 1 when the number of neighbors is large, so the claim is not even well-defined as written. The paper must either prove a high-probability bound on the averaged control, project u*_0 onto the ORCA halfspaces aft
  2. [Sec. IV-A, Eqs. (5)–(6) / Sec. V-B, Eqs. (12)–(14)] The MPPI weight formula is derived for rollouts sampled from a single Gaussian proposal centered at the previous optimal control sequence. CoRL-MPPI samples from two different Gaussian proposals with adjusted means and covariances, N(û^π,Σ̂^π) and N(û^mppi,Σ̂^mppi). Using the original S and ω expressions without including the proposal-density ratio means the weighted average is not the MPPI estimator for the stochastic optimal control problem (4). Therefore the central claim that CoRL-MPPI 'preserves all theoretical guarantees of regular MPPI' is not justified. The authors should either provide the correct importance-sampling weight for the mixture or state clearly that the algorithm is a heuristic sampling scheme whose optimality guarantee is not inherited from standard MPPI.
minor comments (5)
  1. [Sec. VI-A / Eq. (7)] The values of the safety parameters δ, δ_ν, δ_u, H_safe, λ, γ, Σ*, and the RL reward weights are not reported. Without these, the probabilistic safety claim cannot be checked and the experiments cannot be reproduced exactly.
  2. [Algorithm 1, line 17] The notation {u^k}_{k=Kπ} is ambiguous; it should specify the range Kπ+1,...,K. Line 18 also iterates k∈1,...,H although k indexes rollouts, not the horizon.
  3. [Sec. VI, Table I] The name 'B-UAVC' appears both as 'B-UAVC' and 'B-UA VC', and the abstract mentions BVC while the experiments compare B-UAVC. Please make the baseline naming consistent.
  4. [Fig. 4] The makespan plots show only point averages; reporting standard errors or confidence intervals would help assess whether the observed differences are statistically meaningful.
  5. [Sec. V-B] The safety analysis relies on constant-velocity predictions of neighbors for t>0 (Eq. (10)), while in the closed loop all agents re-plan simultaneously. The paper should state explicitly that only the first executed control is safety-filtered with respect to current observations and that no formal multi-step closed-loop guarantee is derived.

Circularity Check

1 steps flagged

Provably-safe claim is inherited by self-citation to the authors' own prior MPPI-ORCA work; the empirical benchmark results are independent and non-circular.

specific steps
  1. self citation load bearing [Section IV-B (Eq. 7) and contribution list in Section I; Algorithm 1, lines 9-10]
    "To ensure collision-free behavior during control sampling, we adopt an approach proposed in [4], [5]. ... A detailed derivation of the constraint formulation, as well as the transformation of the problem into SOCP and LP forms with all theoretical justifications, is provided in [5]. ... We provide theoretical justification that CoRL-MPPI preserves safety guarantees"

    The paper's theoretical contribution is not derived here; it is explicitly delegated to references [4] and [5], both authored by present authors Dergachev and Yakovlev. The claim that 'CoRL-MPPI preserves all the theoretical guarantees of regular MPPI' is therefore load-bearing on a self-citation. If [5] were not accepted, this paper would contain no proof of its headline safety guarantee. Moreover, the cited guarantee applies to individual samples drawn from the adjusted Gaussian distributions, whereas Algorithm 1 executes a cost-weighted average of samples (line 22); the paper gives no argument that the averaged control inherits the sampled safety property. Thus the central theoretical claim reduces to an assertion imported from the authors' own prior work, and even that import does not

full rationale

The only significant circularity is the safety guarantee. Section IV-B states that the safety-constrained distribution update is 'an approach proposed in [4], [5]' and that all 'theoretical justifications' are 'provided in [5]'. Since [4] and [5] are by two of the same authors, the paper's contribution 'We provide theoretical justification that CoRL-MPPI preserves safety guarantees' is an appeal to the authors' own prior work rather than a self-contained derivation. This is load-bearing because the paper's title and abstract rest on the 'provably-safe' claim. It is not, however, a case of a fitted parameter being renamed a prediction: the empirical success-rate and makespan results come from simulation comparisons against ORCA-DD, B-UAVC, and MPPI-ORCA, and are independent of the theoretical inheritance. The learned RL policy is a proposal generator, not a parameter fitted to the reported metrics. The constant-velocity neighbor prediction and the gap between sampled-control safety and weighted-average executed control are correctness/soundness concerns rather than circularity: they do not make an equation equal to its input, but they do reinforce the point that the cited guarantee is doing the work and may not even cover the actual closed-loop execution. Overall score 4 reflects one load-bearing self-citation with the central empirical claim remaining independently supported.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical entities. The main free parameters are the trained RL policy and several unstated control/safety hyperparameters that the central results depend on. The axioms include the imported safety proof, the constant-velocity prediction model, the Gaussian assumption for the RL policy, and the unverified use of the MPPI weighting formula for a mixture proposal.

free parameters (7)
  • RL policy network weights = not provided
    The entire method depends on the learned mapping from observations to action-distribution parameters, trained via IPPO on specific scenarios (Circle, Mesh Sparse). The checkpoint is not released.
  • Mixing ratio of RL-guided rollouts = 0.3
    30% of the 1500 rollouts are sampled from the RL-guided distribution; no ablation is shown for this choice.
  • Safety thresholds δ, δ_ν, δ_u = not specified
    Required in the safety-constraint optimization (Eq. 7); actual values are never given, so the probability bound δ_c cannot be computed.
  • Safety horizon H_safe = not specified
    Determines for how many planning steps the safety constraints (7) are enforced; no value is stated.
  • MPPI cost function weights = not specified
    The components of S(x,u) are listed qualitatively (goal deviation, collision penalty, etc.) but no numerical weights are provided.
  • MPPI parameters λ, γ, Σ* = not specified
    Inverse temperature and control cost weight affect the weighting of trajectories; no values are given.
  • RL reward weights = 0.2, 1, 0.5
    Reward shaping terms for goal arrival, collision, and distance progress are chosen without sensitivity analysis.
axioms (4)
  • domain assumption The probabilistic safety-constraint optimization (Eq. 7) is valid and guarantees the claimed per-step collision probability, as proven in the authors' prior work [5].
    The paper does not re-derive this result; it is the foundation of the 'provably-safe' claim and is cited from a self-authored arXiv preprint.
  • domain assumption Neighbor positions over the planning horizon follow a constant-velocity model (Eq. 10).
    Algorithm 1 predicts neighbor positions this way and uses them in the safety constraints; in the closed-loop system, all agents re-plan and this assumption is violated.
  • domain assumption The RL policy outputs a Gaussian control distribution (mean and covariance) that is compatible with the safety-constraint update in Eq. (7).
    The paper writes (u^π, Σ^π) = π(x,o). If the policy is deterministic or outputs a non-Gaussian distribution, the SOCP safety update is not directly applicable.
  • ad hoc to paper The MPPI weighted-average update (Eq. 5-6) remains the optimal control estimator when sampling from a mixture of two different Gaussian proposal distributions.
    Standard MPPI importance-sampling derivations assume a single Gaussian proposal centered at the previous control sequence; using a mixture changes the measure without a re-derived correction term. The paper does not address this gap.

pith-pipeline@v1.3.0-alltime-deepseek · 13592 in / 12779 out tokens · 136939 ms · 2026-08-03T22:38:01.758264+00:00 · methodology

0 comments
read the original abstract

Decentralized collision avoidance is a core challenge for scalable multi-robot systems. A promising approach to this problem is Model Predictive Path Integral (MPPI) control - a framework that naturally handles arbitrary motion models and provides strong theoretical guarantees. Still, in practice an MPPI-based controller may produce suboptimal trajectories because its performance relies heavily on uninformed random sampling. We introduce CoRL-MPPI, a fusion of Cooperative Reinforcement Learning and MPPI that addresses this limitation. We train an action policy, approximated by a deep neural network, in simulation to learn local cooperative collision-avoidance behaviors. This learned policy is then embedded into the MPPI framework to guide its sampling distribution, biasing it toward more intelligent and cooperative actions in scenarios that may differ substantially from those used during training. Moreover, CoRL-MPPI preserves the theoretical guarantees of regular MPPI. We evaluate our approach in dense, dynamic setups against classical and learning-based state-of-the-art baselines. Our results demonstrate that CoRL-MPPI outperforms competing methods and significantly improves navigation efficiency, measured by success rate and delay, as well as safety, enabling agile and robust multi-robot navigation.

Figures

Figures reproduced from arXiv: 2511.09331 by Aleksandr Panov, Alexey Skrynnik, Artem Pshenitsyn, Konstantin Yakovlev, Stepan Dergachev.

Figure 1
Figure 1. Figure 1: The figure illustrates the core idea of our method for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of safety-constrained update of distri [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustrative visualization of the experimental scenar [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The average makespan of the evaluated algorithms across the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 7 linked inside Pith

  1. [1]

    Reciprocal n-body collision avoidance,

    J. Van Den Berg, S. J. Guy, M. Lin, and D. Manocha, “Reciprocal n-body collision avoidance,” inRobotics research, 2011, pp. 3–19

    2011

  2. [2]

    Fast, on- line collision avoidance for dynamic vehicles using buffered voronoi cells,

    D. Zhou, Z. Wang, S. Bandyopadhyay, and M. Schwager, “Fast, on- line collision avoidance for dynamic vehicles using buffered voronoi cells,”IEEE Robotics and Automation Letters, vol. 2, no. 2, pp. 1047– 1054, 2017

    2017

  3. [3]

    Aggressive driving with model predictive path integral control,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Aggressive driving with model predictive path integral control,” in 2016 IEEE International Conference on Robotics and Automation (ICRA), 2016, pp. 1433–1440

    2016

  4. [4]

    Model predictive path integral for de- centralized multi-agent collision avoidance,

    S. Dergachev and K. Yakovlev, “Model predictive path integral for de- centralized multi-agent collision avoidance,”PeerJ Computer Science, vol. 10, p. e2220, 2024

    2024

  5. [5]

    Decentralized uncertainty-aware multi-agent collision avoidance with model predictive path integral,

    S. Dergachev and K. Yakovlev, “Decentralized uncertainty-aware multi-agent collision avoidance with model predictive path integral,”

  6. [6]

    Information theoretic mpc for model-based reinforcement learning,

    G. Williams, N. Wagener, B. Goldfain, P. Drews, J. M. Rehg, B. Boots, and E. A. Theodorou, “Information theoretic mpc for model-based reinforcement learning,” in2017 IEEE International Conference on Robotics and Automation (ICRA), 2017, pp. 1714–1721

    2017

  7. [7]

    Smooth model predictive path integral control without smoothing,

    T. Kim, G. Park, K. Kwak, J. Bae, and W. Lee, “Smooth model predictive path integral control without smoothing,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 10 406–10 413, 2022

    2022

  8. [8]

    Trajectory dis- tribution control for model predictive path integral control using covariance steering,

    J. Yin, Z. Zhang, E. Theodorou, and P. Tsiotras, “Trajectory dis- tribution control for model predictive path integral control using covariance steering,” in2022 International Conference on Robotics and Automation (ICRA), 2022, pp. 1478–1484

    2022

  9. [9]

    Constrained covariance steering based tube-mppi,

    I. M. Balci, E. Bakolas, B. Vlahov, and E. A. Theodorou, “Constrained covariance steering based tube-mppi,” in2022 American Control Conference (ACC), 2022, pp. 4197–4202

    2022

  10. [10]

    Robust model predictive path integral control: Analysis and performance guarantees,

    M. S. Gandhi, B. Vlahov, J. Gibson, G. Williams, and E. A. Theodorou, “Robust model predictive path integral control: Analysis and performance guarantees,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 1423–1430, 2021

    2021

  11. [11]

    Control barrier function augmentation in sampling-based control algorithm for sample efficiency,

    C. Tao, H. Kim, H. Yoon, N. Hovakimyan, and P. V oulgaris, “Control barrier function augmentation in sampling-based control algorithm for sample efficiency,” in2022 American Control Conference (ACC), 2022, pp. 3488–3493

    2022

  12. [12]

    Path integral methods with stochastic control barrier functions,

    C. Tao, H.-J. Yoon, H. Kim, N. Hovakimyan, and P. V oulgaris, “Path integral methods with stochastic control barrier functions,” in2022 IEEE 61st Conference on Decision and Control (CDC), 2022, pp. 1654–1659

    2022

  13. [13]

    Sampling-based optimization for multi-agent model predictive con- trol,

    Z. Wang, A. D. Saravanos, H. Almubarak, O. So, and E. A. Theodorou, “Sampling-based optimization for multi-agent model predictive con- trol,”arXiv preprint arXiv:2211.11878, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Safety embedded stochastic optimal control of networked multi-agent systems via barrier states,

    L. Song, P. Zhao, N. Wan, and N. Hovakimyan, “Safety embedded stochastic optimal control of networked multi-agent systems via barrier states,” in2023 American Control Conference (ACC), 2023, pp. 2554– 2559

    2023

  15. [15]

    Chance-constrained sampling- based mpc for collision avoidance in uncertain dynamic environ- ments,

    I. S. Mohamed, M. Ali, and L. Liu, “Chance-constrained sampling- based mpc for collision avoidance in uncertain dynamic environ- ments,”IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 7492–7499, 2025

    2025

  16. [16]

    Multi-agent path integral control for interaction-aware motion planning in urban canals,

    L. Streichenberg, E. Trevisan, J. J. Chung, R. Siegwart, and J. Alonso- Mora, “Multi-agent path integral control for interaction-aware motion planning in urban canals,” in2023 IEEE International Conference on Robotics and Automation (ICRA), 2023, pp. 1379–1385

    2023

  17. [17]

    Temporal difference learning for model predictive control,

    N. Hansen, X. Wang, and H. Su, “Temporal difference learning for model predictive control,”arXiv preprint arXiv:2203.04955, 2022

    Pith/arXiv arXiv 2022

  18. [18]

    Td-mpc2: Scalable, robust world models for continuous control,

    N. Hansen, H. Su, and X. Wang, “Td-mpc2: Scalable, robust world models for continuous control,”arXiv preprint arXiv:2310.16828, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    RL-driven MPPI: Accelerating online control laws calculation with offline policy,

    Y . Qu, H. Chu, S. Gao, J. Guan, H. Yan, L. Xiao, S. E. Li, and J. Duan, “RL-driven MPPI: Accelerating online control laws calculation with offline policy,”IEEE Transactions on Intelligent Vehicles, vol. 9, no. 2, pp. 3605–3616, 2024

    2024

  20. [20]

    Smooth and collision-free navigation for multiple robots under differential-drive constraints,

    J. Snape, J. Van Den Berg, S. J. Guy, and D. Manocha, “Smooth and collision-free navigation for multiple robots under differential-drive constraints,” in2010 IEEE/RSJ international conference on intelligent robots and systems, 2010, pp. 4584–4589

    2010

  21. [21]

    Smooth coordination and navigation for multiple differential-drive robots,

    J. Snape, S. J. Guy, J. Van Den Berg, and D. Manocha, “Smooth coordination and navigation for multiple differential-drive robots,” inExperimental Robotics: The 12th International Symposium on Experimental Robotics, 2014, pp. 601–613

    2014

  22. [22]

    Optimal reciprocal collision avoidance for multiple non- holonomic robots,

    J. Alonso-Mora, A. Breitenmoser, M. Rufli, P. Beardsley, and R. Sieg- wart, “Optimal reciprocal collision avoidance for multiple non- holonomic robots,” inDistributed autonomous robotic systems: The 10th international symposium, 2013, pp. 203–216

    2013

  23. [23]

    Cooperative collision avoidance for nonholonomic robots,

    J. Alonso-Mora, P. Beardsley, and R. Siegwart, “Cooperative collision avoidance for nonholonomic robots,”IEEE Transactions on Robotics, vol. 34, no. 2, pp. 404–420, 2018

    2018

  24. [24]

    Prvo: Probabilistic reciprocal velocity obstacle for multi robot navigation under uncertainty,

    B. Gopalakrishnan, A. K. Singh, M. Kaushik, K. M. Krishna, and D. Manocha, “Prvo: Probabilistic reciprocal velocity obstacle for multi robot navigation under uncertainty,” in2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017, pp. 1089– 1096

    2017

  25. [25]

    Multi-robot col- lision avoidance with localization uncertainty,

    D. Hennes, D. Claes, W. Meeussen, and K. Tuyls, “Multi-robot col- lision avoidance with localization uncertainty,” inProceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems-Volume 1, 2012, pp. 147–154

    2012

  26. [26]

    Collision avoidance under bounded localization uncertainty,

    D. Claes, D. Hennes, K. Tuyls, and W. Meeussen, “Collision avoidance under bounded localization uncertainty,” in2012 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems, 2012, pp. 1192– 1198

    2012

  27. [27]

    Distributed collision avoidance of multiple robots with probabilistic buffered voronoi cells,

    M. Wang and M. Schwager, “Distributed collision avoidance of multiple robots with probabilistic buffered voronoi cells,” in2019 in- ternational symposium on multi-robot and multi-agent systems (MRS), 2019, pp. 169–175

    2019

  28. [28]

    Decentralized probabilis- tic multi-robot collision avoidance using buffered uncertainty-aware voronoi cells,

    H. Zhu, B. Brito, and J. Alonso-Mora, “Decentralized probabilis- tic multi-robot collision avoidance using buffered uncertainty-aware voronoi cells,”Autonomous Robots, vol. 46, no. 2, pp. 401–420, 2022

    2022

  29. [29]

    Decentralized non-communicating multiagent collision avoidance with deep rein- forcement learning,

    Y . F. Chen, M. Liu, M. Everett, and J. P. How, “Decentralized non-communicating multiagent collision avoidance with deep rein- forcement learning,” inProceedings of the 2017 IEEE International Conference on Robotics and Automation ({ICRA}2017), 2017, pp. 285–292

    2017

  30. [30]

    To- wards optimally decentralized multi-robot collision avoidance via deep reinforcement learning,

    P. Long, T. Fan, X. Liao, W. Liu, H. Zhang, and J. Pan, “To- wards optimally decentralized multi-robot collision avoidance via deep reinforcement learning,” in2018 IEEE international conference on robotics and automation (ICRA), 2018, pp. 6252–6259

    2018

  31. [31]

    Distributed multi-robot collision avoidance via deep reinforcement learning for navigation in complex scenarios,

    T. Fan, P. Long, W. Liu, and J. Pan, “Distributed multi-robot collision avoidance via deep reinforcement learning for navigation in complex scenarios,”The International Journal of Robotics Research, vol. 39, no. 7, pp. 856–892, 2020

    2020

  32. [32]

    Least-restrictive multi-agent collision avoidance via deep meta reinforcement learning and optimal control,

    S. Asayesh, M. Chen, M. Mehrandezh, and K. Gupta, “Least-restrictive multi-agent collision avoidance via deep meta reinforcement learning and optimal control,” inInternational Conference on Robot Intelli- gence Technology and Applications, 2022, pp. 213–225

    2022

  33. [33]

    Re- inforcement learned distributed multi-robot navigation with reciprocal velocity obstacle shaped rewards,

    R. Han, S. Chen, S. Wang, Z. Zhang, R. Gao, Q. Hao, and J. Pan, “Re- inforcement learned distributed multi-robot navigation with reciprocal velocity obstacle shaped rewards,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 5896–5903, 2022

    2022

  34. [34]

    The complexity of decentralized control of markov decision processes,

    D. S. Bernstein, R. Givan, N. Immerman, and S. Zilberstein, “The complexity of decentralized control of markov decision processes,” Mathematics of operations research, vol. 27, no. 4, pp. 819–840, 2002

    2002

  35. [35]

    Planning and acting in partially observable stochastic domains,

    L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, “Planning and acting in partially observable stochastic domains,”Artificial intelli- gence, vol. 101, no. 1-2, pp. 99–134, 1998

    1998

  36. [36]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  37. [37]

    Is independent learning all you need in the starcraft multi-agent challenge?

    C. S. De Witt, T. Gupta, D. Makoviichuk, V . Makoviychuk, P. H. Torr, M. Sun, and S. Whiteson, “Is independent learning all you need in the starcraft multi-agent challenge?”arXiv preprint arXiv:2011.09533, 2020

    Pith/arXiv arXiv 2011

  38. [38]

    High- dimensional continuous control using generalized advantage estima- tion,

    J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “High- dimensional continuous control using generalized advantage estima- tion,”arXiv preprint arXiv:1506.02438, 2015

    Pith/arXiv arXiv 2015

  39. [39]

    Camar: Continuous actions multi-agent routing,

    A. Pshenitsyn, A. Panov, and A. Skrynnik, “Camar: Continuous actions multi-agent routing,”arXiv preprint arXiv:2508.12845, 2025

    arXiv 2025

  40. [40]

    Sample factory: Egocentric 3d control from pixels at 100000 fps with asynchronous reinforcement learning,

    A. Petrenko, Z. Huang, T. Kumar, G. Sukhatme, and V . Koltun, “Sample factory: Egocentric 3d control from pixels at 100000 fps with asynchronous reinforcement learning,” inInternational Conference on Machine Learning. PMLR, 2020, pp. 7652–7662

    2020

  41. [41]

    Onnx runtime,

    O. R. developers, “Onnx runtime,” https://onnxruntime.ai/, 2021, ver- sion: x.y.z

    2021

  42. [2025]

    Available: https://arxiv.org/abs/2507.20293

    [Online]. Available: https://arxiv.org/abs/2507.20293

    Pith/arXiv arXiv