Pith. sign in

REVIEW 4 major objections 6 minor 31 references

A distributed version of model-based diffusion lets each robot denoise its own trajectory while conditioning on teammates' current plans, and derives a bound on the gap from centralized planning.

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

2026-08-01 08:46 UTC pith:KWZX5V2G

load-bearing objection The reviewer's main theoretical objection doesn't survive contact with the algebra — the 1/|R| weights do sum to one — but the missing comparison with the closest baseline and lack of error bars are real problems. the 4 major comments →

arxiv 2607.20992 v1 pith:KWZX5V2G submitted 2026-07-23 cs.RO

Distributed Model-Based Diffusion For Scalable Multi-Robot Trajectory Optimization

Haejoon Lee , Xinyi Wang , Taekyung Kim , Dimitra Panagou This is my paper
classification cs.RO
keywords distributed multi-robot trajectory optimizationmodel-based diffusionconditional denoisingsampling-based optimizationscore function approximationnon-convex planningcollision avoidanceserver-robot coordination
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.

The paper sets out to make sampling-based multi-robot trajectory optimization scalable by distributing the reverse diffusion process. Instead of denoising a single high-dimensional joint trajectory, each robot denoises its own control trajectory while conditioning on the other robots' current state trajectories, which a server aggregates and broadcasts. The central theoretical claim is Proposition 1: at every denoising step, the per-robot conditional score is close to the centralized score, with the gap bounded by the Lipschitz sensitivity of the local denoiser to the other robots' trajectories times how far those trajectories still are from their final values. If that bound holds, the approach would give communication-light, learning-free coordination that does not require any robot to reveal its objective or dynamics to others, and the paper's simulations report sub-second planning for teams of up to twenty robots in goal-swapping, coverage, parking, and rush-hour tasks.

Core claim

DMBD replaces the centralized reverse diffusion of model-based diffusion with, for each robot k, a local conditional reverse process over that robot's own control space. At each step, robot k samples candidate trajectories from a Gaussian around its current estimate, weights them by a local conditional likelihood p_k^0(Y_k | τ_{-k}) that uses only its own objective, dynamics, and constraints, and updates its trajectory. Proposition 1 states that the difference between this local conditional score and the centralized score is bounded by (√ᾱ_i/(1−ᾱ_i)) H_k E[‖τ_{-k}^{(0)} − τ_{-k}^{(i)}‖ | Y^{(i)}], where H_k is the Lipschitz constant of robot k's denoiser with respect to the other robots' tra

What carries the argument

The central object is the local conditional score function ∇_{Y_k^{(i)}} log p_k^i(Y_k^{(i)} | τ_{-k}^{(i)}), estimated by self-normalized weighted sampling of candidate control trajectories in robot k's own subspace. The paper's Proposition 1 is the load-bearing result: it derives the bound (√ᾱ_i/(1−ᾱ_i)) H_k E[‖τ_{-k}^{(0)} − τ_{-k}^{(i)}‖ | Y^{(i)}] using the standard Gaussian posterior identity and a Lipschitz assumption on the local denoiser. This bound is what connects the distributed updates to the centralized diffusion objective and justifies the claim that local conditional denoising approximates global denoising.

Load-bearing premise

The proof assumes that the global objective can be split into robot k's local objective plus a term that does not depend on robot k; that split, asserted as the equality used in the proof of Proposition 1, holds only when every constraint shared by several robots enters the local objective with the same weight it has in the global objective, which is not the case for a pairwise collision constraint.

What would settle it

Evaluate the equality used in the proof of Proposition 1 for a two-robot problem with one shared collision constraint: the global cost weights that constraint by 1/2, the local objective weights it by 1, so the two conditional densities are not equal. A numerical companion check would compare the centralized score and the local conditional score at a fixed denoising step in the large-sample limit; if the difference exceeds the Proposition 1 bound, or if DMBD and centralized MBD converge to different solutions, the equivalence claim is falsified.

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

If this is right

  • Multi-robot trajectory optimization can be run in sub-second time for teams of tens of robots, without sampling in the joint trajectory space.
  • Each robot needs to send and receive only one trajectory per denoising step, not a set of candidate samples, keeping communication low.
  • Robots never need to share their local objectives, dynamics, or constraint functions with the server or teammates, so heterogeneous teams can coordinate while keeping information private.
  • The method applies to non-convex, non-smooth, non-differentiable costs and constraints, where gradient-based planners typically fail.
  • If the Proposition 1 bound holds, the quality of coordination degrades gracefully: the approximation error at each step is controlled by how far the other robots' trajectories are from their final solutions.

Where Pith is reading between the lines

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

  • A natural modification would be to rescale each coupled constraint in the local objective by 1/|R_{k,p}|, matching the global cost; doing so would make the density equality in the proof exact and is testable on a two-robot collision problem.
  • The bound depends on how far other robots' trajectories are from their final values, so it could serve as a practical stopping criterion: halt denoising once the estimated gap falls below a tolerance.
  • The server-robot architecture suggests an asynchronous variant where robots broadcast at different rates; the paper's analysis assumes synchronized steps, so an asynchronous extension would need its own bound.

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

4 major / 6 minor

Summary. The paper proposes Distributed Model-Based Diffusion (DMBD), a server-robot framework that distributes the reverse denoising process of Model-Based Diffusion across robots. Each robot denoises its own control trajectory in a local conditional diffusion process, conditioning on the current state-trajectory estimates of other robots aggregated by a server. The paper derives Proposition 1, an upper bound on the discrepancy between the centralized score and the local conditional score, and reports simulations in goal swapping, multi-floor coverage, parking, and rush-hour scenarios claiming strong scalability and superior performance over CEM, MPPI, MBD, and D4orm.

Significance. The algorithmic idea of decomposing the reverse diffusion across robots is natural and potentially useful: it is learning-free, zeroth-order, and communication-light (one trajectory per robot per denoising step). The theoretical bound in Proposition 1 is a reasonable characterization of the approximation error, though it depends on a Lipschitz assumption and a deviation term that is not controlled. I have checked the reader's concern about the 1/|R_{k,p}| weights in Eq. (7): under Assumption 1, the weights cancel across participating robots, so the decomposition used in the proof of Proposition 1 is algebraically valid for constraints satisfying the symmetry assumption. If the empirical claims are substantiated, the paper would offer a practical alternative to centralized multi-robot MBD. However, the current experimental validation does not yet support the headline claims, because the closest baseline [21] is not evaluated and no statistical uncertainty information is provided.

major comments (4)
  1. [Section IV, Fig. 3 and Table I] The central empirical claim that DMBD 'significantly outperform[s] existing baselines' is not supported by the experiments because the most relevant comparator, [21] (individual MBD with centralized deconfliction), is never evaluated. The introduction identifies [21] as the closest existing method; omitting it from the comparisons makes the improvement claim unsubstantiated. Additionally, no comparison is made with the distributed sampling-based methods cited in the introduction ([11], [24], [25]). The authors should add these baselines or substantially qualify the claim.
  2. [Section IV, Fig. 3 and Table I] Success rates are reported as percentages over 50 trials without confidence intervals, error bars, or significance tests. With n=50, the standard error for a binary outcome can reach ~7 percentage points, so reported differences of a few points between DMBD and D4orm in the circular goal-swap scenario may not be statistically meaningful. Planning-time averages are reported without variance. The strong comparative conclusions require error bars and a paired statistical test across the 50 seeds.
  3. [Section IV] The empirical results are not reproducible as reported. No code is released, and the key hyperparameters—temperature λ, noise schedule, cost weights, constraint penalty coefficients, and candidate counts S_k—are not specified; the text only says they were 'tuned once using the MBD on the case with N=2 for each scenario.' To support the empirical claims, the authors should release code or provide a complete parameter table and clarify whether the N=2 tuning uses a separate set of trials.
  4. [Section IV, Q2 (Parking and Rush Hour)] Table I reports only DMBD's absolute success rates and planning times for the parking and rush-hour tasks; no baseline is evaluated in these scenarios. Consequently, these results do not demonstrate that the distributed mechanism outperforms alternatives; they only show that the method can solve the tasks. A comparison to centralized MBD and to [21] on these tasks would be needed to support the claim that distributed conditioning recovers globally coordinated behavior.
minor comments (6)
  1. [Eq. (7) and proof of Proposition 1] The 1/|R_{k,p}| normalization is exactly what makes the decomposition work under Assumption 1; please add a sentence explicitly explaining this, since a reader might otherwise think the weights are ad hoc.
  2. [Proposition 1] The bound (19) contains the term E[||τ_-k^(0) - τ_-k^(i)|| | Y^(i)], which is not shown to be small; the proposition is a characterization of the discrepancy rather than a convergence guarantee. This should be stated clearly in the discussion to avoid over-interpretation.
  3. [Figure 3 caption] 'Baseline methods fail at N=20 due to out-of-memory errors' — please clarify whether DMBD succeeds at N=20 and report the success rate or memory usage at that point, since the figure appears to cut off at N=20.
  4. [Section IV] The planning horizon T is increased with N by 5 steps per additional robot, so the per-robot problem dimension increases with N. Please discuss how this affects the scalability interpretation.
  5. [Table I] Planning-time values are given without variance; report standard deviations or box plots.
  6. [Abstract and Section IV] The phrase 'sub-seconds' is a bit strong given Fig. 3 shows some planning times above 1 second for larger N; please rephrase or add a qualifier.

Circularity Check

0 steps flagged

No significant circularity: Proposition 1's Eq. (20) follows from the weighted constraint sum, not from an assumed equivalence.

full rationale

The derivation chain is self-contained. The central theoretical step, Proposition 1, is not circular: the local conditional target distribution p_k^0(Y_k | tau_-k) is defined independently in Eq. (13)-(14), and Eq. (20) is derived from the global cost in Eq. (7) under Assumption 1. For a coupled constraint with r participants, the global cost contains r identical terms each weighted 1/r, so the total contribution is exactly one copy of g_{k,p}; conditioning on Y_-k leaves every Y_k-dependent term with coefficient 1, matching the local cost in Eq. (14). Thus p0(Y_k^(0)|Y_-k^(0)) = p_k^0(Y_k^(0)|tau_-k^(0)) is a consequence of algebra, not an input. The subsequent Tweedie, tower-property, and Lipschitz argument provides an upper bound on the score discrepancy; it does not assume the conclusion. Self-citations [10] and [18] appear only as background and are not load-bearing, and the convergence citation [27] is external. Remark 3 explicitly states the finite-sample bound may be violated, which is a limitation rather than a circular step. The lack of comparison to [21] and absence of released code are empirical evidence gaps, not circularity. No fitted parameter is relabeled as a prediction; cost parameters are tuned once on MBD at N=2 for baselines, while DMBD's results are independent demonstrations. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claim leans on several unstated or free choices: temperature λ, noise schedule, denoising steps and sample counts are set per scenario; cost weights are tuned once on N=2. The proof also implicitly assumes that the global constraint normalization can be ignored and that the local denoiser is Lipschitz in peers' trajectories. No new physical entities are introduced.

free parameters (6)
  • temperature λ
    Sharpness of target distributions p0 and p_k^0 (Eqs. 6, 13); value not reported; affects score landscape and results.
  • noise schedule {β_i, ᾱ_i}
    Defines forward noising and denoising (Eq. 8); schedule not reported.
  • denoising steps M = 150 (goal-swap), 200 (multi-floor)
    Set per scenario; no ablation; larger M increases compute.
  • candidate samples per robot S_k = 500
    Set to match baselines; chosen by hand.
  • cost weights in J_k and constraint penalties = tuned on N=2 via MBD
    Tuned once per scenario using MBD on N=2; values not reported; all methods share them.
  • Lipschitz constants H_k
    Assumed finite in Prop. 1; not estimated or bounded; central to the discrepancy bound.
axioms (5)
  • standard math Tweedie's formula and Gaussian forward noising are valid for the score estimates (Eqs. 9, 21-22).
    Background result used to express scores as conditional expectations.
  • domain assumption Assumption 1: coupled constraint g_{k,p}(Y) is symmetric and shared by involved robots.
    Required for any decomposition of global cost; restricts settings with heterogeneous/private coupling costs.
  • ad hoc to paper The global cost can be written as local cost plus terms independent of Y_k (proof of Prop. 1).
    This premise is false for |R_{k,p}|>1 because Eq. (7) normalizes constraints by |R| while Eq. (14) does not; the proof silently ignores the factor.
  • ad hoc to paper The local denoiser μ_{k,i} is H_k-Lipschitz in the peers' trajectories τ_{-k}.
    Assumed in Prop. 1; no bound or verification is given; needed for the discrepancy bound.
  • domain assumption Each robot has access only to its own dynamics, cost, and constraints, while the server aggregates only state trajectories.
    Architecture assumption; the practical privacy/scalability claims rely on it.

pith-pipeline@v1.3.0-alltime-deepseek · 12141 in / 19386 out tokens · 190028 ms · 2026-08-01T08:46:01.869836+00:00 · methodology

0 comments
read the original abstract

Trajectory optimization for multi-robot systems remains a critical challenge, particularly when navigating highly non-convex, non-linear, and non-differentiable environments. While Model-Based Diffusion (MBD) has recently emerged as a promising sampling-based optimization paradigm for single-robot trajectory generation, extending it to multi-robot systems results in a centralized, high-dimensional inference problem that (i) suffers from poor sample efficiency due to the curse of dimensionality and (ii) requires global access to all robots' dynamics, constraints, and objectives. To address this, we propose Distributed Model-Based Diffusion (DMBD), a distributed server-robot framework that decomposes the reverse diffusion process into local conditional reverse diffusion processes. This decomposition enables each robot to iteratively perform denoising independently within its own control subspace while conditioning on the current trajectory estimates of the other robots that are aggregated and broadcast by the server. Extensive simulations in goal swapping, multi-floor coverage, parking, and rush-hour scenarios demonstrate that DMBD achieves strong scalability, solving many challenging coordination tasks in sub-seconds and significantly outperforming existing baselines.

Figures

Figures reproduced from arXiv: 2607.20992 by Dimitra Panagou, Haejoon Lee, Taekyung Kim, Xinyi Wang.

Figure 1
Figure 1. Figure 1: Visualization of Distributed Model-Based Diffusion (DMBD). For every denoising step, each robot independently (i) performs local conditional [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: DMBD’s local conditional denoising process for goal swapping [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Success rates and average planning time across [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualizations of representative planned trajectories through DMBD for the (a) parking and (b) rush hour scenarios. [PITH_FULL_IMAGE:figures/full_fig_p008_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

31 extracted references · 5 linked inside Pith

  1. [1]

    Nocedal and S

    J. Nocedal and S. J. Wright,Numerical optimization. Springer, 2006

    2006

  2. [2]

    Distributed mpc for multi-vehicle systems moving in formation,

    R. Van Parys and G. Pipeleers, “Distributed mpc for multi-vehicle systems moving in formation,”Robotics and Autonomous Systems, vol. 97, pp. 144–152, 2017

    2017

  3. [3]

    Distributed nonlinear trajectory optimization for multi-robot motion planning,

    L. Ferranti, L. Lyons, R. R. Negenborn, T. Keviczky, and J. Alonso- Mora, “Distributed nonlinear trajectory optimization for multi-robot motion planning,”IEEE Transactions on Control Systems Technology, vol. 31, no. 2, pp. 809–824, 2023

    2023

  4. [4]

    Distributed optimization methods for multi-robot systems: Part 1—a tutorial [tutorial],

    O. Shorinwa, T. Halsted, J. Yu, and M. Schwager, “Distributed optimization methods for multi-robot systems: Part 1—a tutorial [tutorial],”IEEE Robotics & Automation Magazine, vol. 31, no. 3, pp. 121–138, 2024

    2024

  5. [5]

    The cross-entropy method for combinatorial and continuous optimization,

    R. Rubinstein, “The cross-entropy method for combinatorial and continuous optimization,”Methodology and computing in applied probability, vol. 1, no. 2, pp. 127–190, 1999

    1999

  6. [6]

    Information-theoretic model predictive control: Theory and applica- tions to autonomous driving,

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou, “Information-theoretic model predictive control: Theory and applica- tions to autonomous driving,”IEEE Transactions on Robotics, vol. 34, no. 6, pp. 1603–1622, 2018

    2018

  7. [7]

    Model-based diffusion for trajectory optimization,

    C. Pan, Z. Yi, G. Shi, and G. Qu, “Model-based diffusion for trajectory optimization,”Advances in Neural Information Processing Systems, vol. 37, pp. 57 914–57 943, 2024

    2024

  8. [8]

    Full-order sampling-based mpc for torque-level locomotion control via diffusion-style annealing,

    H. Xue, C. Pan, Z. Yi, G. Qu, and G. Shi, “Full-order sampling-based mpc for torque-level locomotion control via diffusion-style annealing,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 4974–4981

    2025

  9. [9]

    The cross- entropy method for optimization,

    Z. I. Botev, D. P. Kroese, R. Y . Rubinstein, and P. L’ecuyer, “The cross- entropy method for optimization,” inHandbook of statistics. Elsevier, 2013, vol. 31, pp. 35–59

    2013

  10. [10]

    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

  11. [11]

    Distributed sampling-based model predictive control via be- lief propagation for multi-robot formation navigation,

    C. Jiang, “Distributed sampling-based model predictive control via be- lief propagation for multi-robot formation navigation,”IEEE Robotics and Automation Letters, vol. 9, no. 4, pp. 3467–3474, 2024

    2024

  12. [12]

    Sampling-based control via entropy-regularized optimal transport,

    V . Pacelli, A. Ratheesh, and E. A. Theodorou, “Sampling-based control via entropy-regularized optimal transport,”arXiv preprint arXiv:2605.02147, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Denoising diffusion probabilistic models,

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

    2020

  14. [14]

    Plan- ning with diffusion for flexible behavior synthesis,

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Plan- ning with diffusion for flexible behavior synthesis,”arXiv preprint arXiv:2205.09991, 2022

    Pith/arXiv arXiv 2022

  15. [15]

    Guided conditional diffusion for controllable traffic simulation,

    Z. Zhong, D. Rempe, D. Xu, Y . Chen, S. Veer, T. Che, B. Ray, and M. Pavone, “Guided conditional diffusion for controllable traffic simulation,” in2023 IEEE international conference on robotics and automation (ICRA). IEEE, 2023, pp. 3560–3566

    2023

  16. [16]

    Discrete-guided diffusion for scalable and safe multi-robot motion planning,

    J. Liang, S. Koenig, and F. Fioretto, “Discrete-guided diffusion for scalable and safe multi-robot motion planning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 28, 2026, pp. 23 417–23 424

    2026

  17. [17]

    Eb-mbd: Emerging-barrier model- based diffusion for safe trajectory optimization in highly constrained environments,

    R. Mishra and I. R. Manchester, “Eb-mbd: Emerging-barrier model- based diffusion for safe trajectory optimization in highly constrained environments,”arXiv preprint arXiv:2510.07700, 2025

    arXiv 2025

  18. [18]

    Safe model predictive diffusion with shielding,

    T. Kim, K. Majd, H. Okamoto, B. Hoxha, D. Panagou, and G. Fainekos, “Safe model predictive diffusion with shielding,”arXiv preprint arXiv:2512.06261, 2025

    arXiv 2025

  19. [19]

    On the complexity of motion planning for multiple independent objects; pspace-hardness of the

    J. E. Hopcroft, J. T. Schwartz, and M. Sharir, “On the complexity of motion planning for multiple independent objects; pspace-hardness of the” warehouseman’s problem”,”The international journal of robotics research, vol. 3, no. 4, pp. 76–88, 1984

    1984

  20. [20]

    D4orm: Multi-robot trajectories with dynamics-aware diffusion denoised defor- mations,

    Y . Zhang, K. Okumura, H. Woo, A. Shankar, and A. Prorok, “D4orm: Multi-robot trajectories with dynamics-aware diffusion denoised defor- mations,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 14 118–14 123

    2025

  21. [21]

    Motion planning with model-based diffusion via constraint optimization and adaptive scheduling,

    Z. He, B. Li, J. Dou, Y . Zhang, and C. Liu, “Motion planning with model-based diffusion via constraint optimization and adaptive scheduling,”arXiv preprint arXiv:2607.14455, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    Coordinating hundreds of cooperative, autonomous vehicles in warehouses,

    P. R. Wurman, R. D’Andrea, and M. Mountz, “Coordinating hundreds of cooperative, autonomous vehicles in warehouses,”AI magazine, vol. 29, no. 1, pp. 9–9, 2008

    2008

  23. [23]

    A critical review of communi- cations in multi-robot systems,

    J. Gielis, A. Shankar, and A. Prorok, “A critical review of communi- cations in multi-robot systems,”Current robotics reports, vol. 3, no. 4, pp. 213–225, 2022

    2022

  24. [24]

    Cooperative path integral control for stochastic multi-agent systems,

    N. Wan, A. Gahlawat, N. Hovakimyan, E. A. Theodorou, and P. G. V oulgaris, “Cooperative path integral control for stochastic multi-agent systems,” inAmerican Control Conference (ACC), 2021, pp. 1262– 1267

    2021

  25. [25]

    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,”arXiv preprint arXiv:2302.06547, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Stein variational belief propagation for multi-robot coordination,

    J. Pavlasek, J. J. Z. Mah, R. Xu, O. C. Jenkins, and F. Ramos, “Stein variational belief propagation for multi-robot coordination,” IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 4194–4201, 2024

    2024

  27. [27]

    Global convergence of sampling- based nonconvex optimization through diffusion-style smoothing,

    Z. Yi, C. Pan, G. Shi, and G. Qu, “Global convergence of sampling- based nonconvex optimization through diffusion-style smoothing,” arXiv preprint arXiv:2605.16520, 2026

    Pith/arXiv arXiv 2026

  28. [28]

    Villaniet al.,Optimal transport: old and new

    C. Villaniet al.,Optimal transport: old and new. Springer, 2009, vol. 338

    2009

  29. [29]

    JAX: composable transforma- tions of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, Y . Katariya, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, and Q. Zhang, “JAX: composable transforma- tions of Python+NumPy programs,” 2018

    2018

  30. [30]

    The kinematic bicycle model: A consistent model for planning feasible tra- jectories for autonomous vehicles?

    P. Polack, F. Altch ´e, B. d’Andr ´ea Novel, and A. de La Fortelle, “The kinematic bicycle model: A consistent model for planning feasible tra- jectories for autonomous vehicles?” in2017 IEEE intelligent vehicles symposium (IV). IEEE, 2017, pp. 812–818

    2017

  31. [31]

    Ericson,Real-Time Collision Detection

    C. Ericson,Real-Time Collision Detection. Morgan Kaufmann, 2005

    2005