Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition

Fangguo Zhao; Lei Xie; Wule Mao; Yunhao Luo; Zhouheng Li

REVIEW 4 major objections 5 minor 1 cited by

Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition

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 composing diffusion models trained separately on static and dynamic obstacle scenes yields a trajectory planner that generalizes to unseen compound scenes without retraining, running in a mean 0.21 s with a 0.57%…

desk verdict A working diffusion-planning system with real hardware results, but the headline generalization claim rests on hand-tuned composition weights that the paper itself shows are fragile. read the letter →

arxiv 2507.04384 v4 pith:B5NGUMZS submitted 2025-07-06 cs.RO cs.SYeess.SY

Wule Mao , Zhouheng Li , Yunhao Luo , Fangguo Zhao , Lei Xie This is my paper · ORCID

classification cs.ROcs.SYeess.SY

keywords diffusionmodelstrajectoryplanningmodelcompositionsafetyfiltergeneralizationtounseenscenescollisionavoidancekinematicfeasibilityF1TENTH

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

The paper sets out to show that a mobile robot can plan trajectories that are both rapid and safe in scenes it has never seen, by composing diffusion models that were trained separately on static-only and dynamic-only environments. The core mechanism is to read each diffusion model as the score of an energy field over trajectories, so the static-obstacle model and the dynamic-obstacle model can be combined at test time into one planner without retraining. In simulation, the composed planner reaches a mean planning time of 0.21 s and a planning failure rate of 0.57%, and on a real F1TENTH vehicle (a 1/10-scale autonomous car) it keeps trajectories collision-free under localization uncertainty, with a minimum obstacle clearance of 0.032 m and a mean of about 0.26 m. The paper also contributes an MPC-based data-generation pipeline that produces kinematically feasible expert trajectories, and a lightweight safety filter that picks the best candidate from a batch. The claim is defended with simulation benchmarks and real-vehicle experiments.

What carries the argument

The load-bearing object is a test-time diffusion-composition rule that treats each trained diffusion model as the score (gradient of the log-density) of an energy field over candidate trajectories $\tau_{1:L}$, and combines constraint-specific fields multiplicatively. Concretely, the composed noise estimate is the unconditional estimate plus a weighted sum of per-model deviations: $\epsilon_{\mathrm{compose}} = \epsilon_{\mathrm{uncond}} + \sum_i \nu_i(\epsilon_i - \epsilon_{\mathrm{uncond}})$, with composition weights $\nu_i$; the same network emits both conditional and unconditional estimates by masking its obstacle encoder. On top of this, a rule-based safety filter draws a batch of $N_{\mathrm{filter}}$ candidate trajectories and selects the one minimizing a normalized cost of path length, acceleration, steering rate, and obstacle distance, while discarding trajectories that violate kinematic limits. This combination is what stitches models trained on static-only and dynamic-only scenes into a planner for scenes that contain both.

What would settle it

Run the composed planner in a scene where static and dynamic obstacles force mutually incompatible constraints, such as a corridor narrower than the vehicle or a required deceleration beyond the dynamics limits, and sweep the composition weights over a grid; if no weight setting achieves near-zero failure rate, the multiplicative-independence composition rule is contradicted. A simpler check is to measure the raw composed samples' collision rate with the safety filter disabled in a new scene: if the raw samples are never safe, the composition itself is not the source of the claimed safety.

Watch

Extended reading notes

Core claim

The central discovery is that generalization to an unseen scene can be obtained by composition rather than retraining. Treating each diffusion model as a score that points toward lower-energy (safer) trajectories, RSTP forms a composed noise estimate from a static-scene model and a dynamic-scene model so that the combined planner avoids both obstacle types. In the paper's experiments the composed planner runs in 0.21 s mean planning time with a 0.57% failure rate, and the same trajectories, when tracked by a simple pure-pursuit controller on a real F1TENTH vehicle, remain collision-free under sensor uncertainty. The paper further claims that kinematic feasibility is learned implicitly from MPC-generated demonstrations, cutting mean tracking error by 47.23% relative to the Hybrid A* baseline, and that the rule-based safety filter is what turns a batch of plausible trajectories into a reliably safe selection.

Load-bearing premise

The load-bearing assumption is that obstacle constraints combine multiplicatively: the conditional trajectory distribution for a scene with both static and dynamic obstacles is assumed to equal the unconditional distribution times the product of per-obstacle-condition ratios, and the composition weights that balance those ratios are hand-tuned per scene, where small changes can raise the failure rate from 0% to 69.62%.

Editorial extensions

If this is right

A robot can be trained once on a static environment and once on a dynamic environment, then deployed where both types of obstacles appear, with no retraining and no new demonstrations.
The reported mean planning time of 0.21 s with a standard deviation of $3.16\times 10^{-4}$ places the planner in the real-time range on commodity hardware, at least for the 6 m by 6 m map and 128-step horizon tested.
Enabling the safety filter reduces the failure rate from 4.50% to 0.57% and the tracking-collision rate from 11.87% to 8.80%, so batch selection is doing real safety work rather than being a cosmetic stage.
Because the composed trajectories already encode kinematic feasibility, they can be executed by a simple pure-pursuit tracker without post-hoc optimization, which is the basis of the real-vehicle clearance results.

Reading between the lines

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

The paper only composes two models (static plus one dynamic); an extension to three or more composed models would test whether the multiplicative independence assumption compounds errors as the number of constraints grows.
Appendix C shows the composition weights must be hand-tuned per scene, with failure rates jumping from 0% to 69.62% at weight 8.0 in one scene, so a practical deployment would likely need an automatic weight-selection mechanism.
The real and simulated scenes all share the same 6 m by 6 m map, so a stronger transfer test would compose models trained on qualitatively different maps to see whether the method recombines scene structure or merely re-weights obstacles from one layout.
Folding the composition weights into the safety filter's cost function, or adjusting them from observed obstacle density, is a concrete extension that could avoid the failure regime the paper reports at large weights.

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.

Desk Editor's Note

read the letter

RSTP is a legitimate systems result: the authors combine an MPC-based data pipeline, an energy-based diffusion planner, and a filter that selects among a batch of sampled trajectories, and they demonstrate the whole thing on an F1TENTH car at 0.21 s mean planning time. The composition of separately trained static and dynamic models at test time is a sensible idea, and the real-world runs show it can work. Credit where due: the data generation pipeline is clean, the planner is fast, and the paper is transparent about weight sensitivity—Appendix C is unusual in showing failure rates jumping from 0% to 69.62% as the dynamic weight goes from 0.8 to 8.0.

The soft spot is the load-bearing composition claim. Eq. (2) is a product-of-experts form that assumes conditional independence of obstacle constraints given the trajectory. But Eq. (3) is a weighted score sum, and it reduces to Eq. (2) only when all weights equal 1, which is never tested. Instead, the weights are free parameters tuned per scene. That turns 'generalization without retraining' into 'generalization at hand-picked weights'—a much weaker statement. The safety filter is also a selection heuristic, not a guarantee: the simulation collision rate during tracking is still 8.8%, and the real-world evaluation covers only two dynamic obstacle trajectories and six start points. The mild circularity of training on MPC data and evaluating with an MPC tracker is worth a footnote but isn't fatal.

Who should read this? Robotics practitioners working on diffusion-based planning will find the system-level details and real-world numbers useful. The generalization story needs more evidence: multiple seeds, more dynamic scenes, and a principled way to set the composition weights. I would send it to peer review, with the expectation that the authors tighten the composition analysis and expand the evaluation.

Referee Report

4 major / 5 minor

Summary. The paper proposes RSTP, a diffusion-based trajectory planner for F1TENTH-style vehicles. It combines an MPC-based data generation pipeline, separate diffusion models trained on static and dynamic scenes, a test-time composition rule that combines the score estimates of the individual models, and a rule-based safety filter that selects one trajectory from a batch of candidates. The reported results include a mean planning time of 0.21 s, a 0.57% planning failure rate in the static evaluation scene, and real-world demonstrations on an F1TENTH vehicle in composed scenes with one moving obstacle.

Significance. If the generalization claim held, the work would be practically significant: real-time planning on commodity hardware, a reusable MPC-based data generation pipeline, and no retraining for new combinations of static and dynamic obstacles would be valuable contributions. The real-world experiments and the lightweight safety filter are useful engineering efforts. However, the central claim of safe generalization to unseen scenes is not established because the composition rule is an untested assumption with per-scene tuned weights, and the safety filter is a heuristic selection rather than an enforcement mechanism. The paper is better characterized as a promising engineering system than as a validated general method.

major comments (4)

[§IV-A, Eq. (2)-(3); Appendix C, Table III] The generalization claim rests on the composition rule, but the paper never tests its Bayes-consistent form. Eq. (3) reduces to Eq. (2) only when all ν_i = 1, and the configuration ν_i = 1 is never evaluated. Instead, Appendix C tunes ν_i per scene and Table III reports the failure rate jumping from 0.00% to 51.90% in CS1 and from 0.00% to 69.62% in CS2 as the dynamic weight increases from 0.8 to 8.0. The 0.00% failure rate at the selected weights is therefore a per-scene tuning artifact, and the claim of generalization without additional tuning is contradicted by the manuscript's own sensitivity analysis.
[§IV-A, Eq. (2)] Eq. (2) requires a single unconditional distribution p_uncond(τ) shared by all composed models, because the product of ratios is taken against one common prior. However, the static and dynamic models are trained on different datasets (Section III), so their unconditional score estimates generally differ. Eq. (3) subtracts each model's own unconditional score, which means it does not implement the product form in Eq. (2) unless the unconditional distributions coincide. The authors should either derive the composition rule under heterogeneous priors or empirically test whether a shared prior is adequate.
[§IV-B; Table I; Table II] The safety filter is a heuristic selection rule, not an enforcement guarantee: it can only choose among samples drawn from the composed distribution, and if that distribution places low mass on safe trajectories, selection cannot recover safety. The paper's own results support this concern: Table I reports a collision rate of 8.8% for RSTP with the filter enabled in simulation, and Table II shows real-world runs with up to 8.54% of trajectory points in the 'Danger' category and a minimum obstacle distance of 0.032 m. The claim that the filter 'ensures safety' should be weakened to 'reduces collision risk,' and the authors should report confidence intervals for the collision and failure rates.
[§V-A, §V-C; Appendix A] The compositional generalization is evaluated on only two dynamic obstacle trajectories (CS1 and CS2) and a 79-pose subset of the evaluation set, with no held-out dynamic obstacle trajectories or additional unseen scenes. This evidence is too narrow to support the claim of generalization across 'diverse scenes.' The authors should evaluate on multiple held-out dynamic obstacle trajectories and report the distribution of failure and collision rates across scenes. In addition, the composition weights should either be selected by a principled automatic procedure or shown to be robust over a wide range, rather than tuned to the two test scenes.

minor comments (5)

[§II-C] The forward noising process is written as q(τ^t | τ^{t-1}) = N(√(1−β_t) τ^t, β_t I); the mean should be √(1−β_t) τ^{t-1}, otherwise the equation is self-referential.
[§III, Eq. (1)] A = [1,1,1,0]^T is a vector but is used inside a norm with ζ_k; specify its role as a selection matrix and check the dimensions in the cost terms.
[§IV-A] The masking notation K=[1,0]^T ⊗ O_i is undefined; clarify how the same network produces conditional and unconditional outputs from this mask.
[Table I] The table reports M.RP for both computation time and mean tracking error without clearly defining which baseline is used; define BLM in the caption and label the two M.RP columns distinctly.
[§IV-A, Eq. (3)] Eq. (3) is the standard classifier-free guidance combination; a citation to the classifier-free guidance literature would help position the novelty of the proposed composition scheme.

Circularity Check

1 steps flagged · score 6.0 of 10

Generalization evidence reduces to per-scene tuning of the free composition weights: the 0% failure rate is selected after observing each test scene's outcomes.

fitted input called prediction [Appendix C, Table III, and Section IV-A, Eq. (3)]
"As indicated in the table, the composed model can safely generalize to unseen scenes when appropriate compositional weights are selected, achieving an F.Rate of 0%. However, increasing the weight assigned to conditional models results in a rise in F.Rate across both test scenes."

Eq. (3) introduces nu_i as free composition parameters. Table III evaluates four weight configurations per scene and reports F.Rate for each, ranging from 0% to 69.62% in CS2 and from 0% to 51.90% in CS1. The 0% F.Rate used as evidence of safe generalization is the best-case configuration selected for that same scene after observing its outcomes. Thus the reported generalization is not a prediction under a fixed composition rule but a per-scene fitted value; the central claim of collision-free generalization without retraining is supported by selecting the weights that already make the test scene succeed.

full rationale

The composition derivation in Eq. (2) is a standard product-of-experts/Bayes-consistent form and is not itself circular; Eq. (3) is a weighted classifier-free-guidance variant. The reuse of architecture and hyperparameters from [10] is a self-citation but not load-bearing, because it supplies an encoder/backbone rather than the generalization result. The MPC-generated training data and MPC-tracker evaluation are a consistency check rather than a definitional reduction. The main circular element is the compositional-weight tuning: Appendix C selects per-scene weights and then reports the resulting 0% failure rate as evidence of generalization, so the headline compositional-safety number is a fitted optimum rather than an out-of-sample prediction. Weight sensitivity is acknowledged in the same appendix, but the paper still presents the tuned 0% result as validation of the generalization claim. This partial reduction of the central evidence to a per-scene fit justifies a score of 6 rather than a clean bill.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities, forces, or dimensions are introduced. The only new components are algorithmic, namely the safety filter and the composition mechanism, which depend on hand-tuned weights and prior-work assumptions.

free parameters (3)

Composition weights nu_i for static/dynamic/unconditional models = e.g., 0.8 to 8.0 in Table III; unconditional weight fixed at 5
Eq. (3) requires weights chosen by hand per scene; Appendix C shows failure rate sensitive to these values.
Safety filter weights omega_1 to omega_4 = all equal (paper states weights in Eq. (4) are set equal)
Chosen by hand, not optimized or ablated; they affect final trajectory selection.
Large penalty V_inf = not specified
Used in Eqs. (4)-(5) to discard collisions and infeasible yaw; value not reported.

assumptions (4)

standard math Diffusion models estimate the gradient of an energy function: grad_tau log p(tau) proportional to epsilon_theta(tau,t,C)
Standard score-matching/EBM result used throughout Section II.
ad hoc to paper Composition via product of conditional densities (Eq. 2) is valid for arbitrary obstacle sets
Assumed without derivation; implies conditional independence of static and dynamic obstacle avoidance.
domain assumption MPC-generated trajectories form an adequate training distribution for collision-free, kinematically feasible planning in unseen scenes
Central to the imitation learning pipeline; only one static and two dynamic scene configurations are used.
domain assumption Geometric distance checks and yaw-rate checks in the safety filter imply safety
Assumes vehicle footprint and tracking error are captured by point-distance to obstacles; min real-world clearance of 0.032m suggests thin margin.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition." pith.science (2026). https://pith.science/paper/B5NGUMZS

@misc{pith2026250704384,
  author       = {Pith},
  title        = {Pith review of: Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5NGUMZS}},
  note         = {Machine review of arXiv:2507.04384}
}

read the original abstract

Achieving safe, efficient, and kinematically feasible planning in dynamic environments remains a significant challenge, as planners must simultaneously handle moving obstacles, sensor uncertainty, and strict motion constraints. To address this problem, we propose an energy-parameterized diffusion planning framework that learns a conservative energy field to realize safe and stable generalization across diverse scenarios. The energy-parameterized diffusion formulation enables flexible integration of multiple constraints, allowing the planner to generalize to previously unseen environments without retraining. To ensure real-time safety during deployment, we further incorporate a lightweight safety filter that enforces safety and kinematic feasibility constraints in real-time. Additionally, we develop a scene-agnostic, MPC-based data generation pipeline to produce large-scale, dynamically feasible training trajectories. In simulation, the proposed method achieves real-time performance with a mean planning time of 0.21s and a low planning failure rate of 0.57%. Real-world experiments on the F1TENTH platform further validate the effectiveness of the proposed framework. Under sensor uncertainty in previously unseen dynamic environments, the planner consistently generates collision-free trajectories, which remain safe after being tracked by a simple controller, maintaining a mean obstacle clearance of 0.26 m, demonstrating strong robustness and practical applicability. Project page: https://rstp-comp-diffuser.github.io.

Figures

Figures reproduced from arXiv: 2507.04384 by the authors.

**Figure 1.** Diffusion Composition Enables Efficient, Safe Planning with Practical Real-world Performance. An individual diffusion model cannot ensure safe trajectory planning in out-of-distribution scenarios, whereas composing multiple diffusion models can achieve safety during generalization. Dashed boxes indicate obstacles that do not exist during training. Validation on the F1TENTH platform shows that trajectories planned by… view at source ↗

**Figure 2.** Diffusion Composition. Trajectory planning using the energy model as a surrogate model for the potential field. The vehicle model f(·) used for MPC is the kinematic model employed in [2]. The state is represented as ζ = [x, y, φ, v] ⊤ and the control input as u = [δ, a] ⊤. δ is the steering angle, and a represents acceleration. v is the velocity and φ represents the vehicle heading. For global collision-free ItCA pa… view at source ↗

**Figure 3.** The Overall Framework of the Proposed Rapid and Safe Trajectory Planning (RSTP) Method. Offline Dataset Generation (Left): the ItCA [3] and MPC-based methods provide kinematically feasible trajectory datasets for training. Diffusion Composition (Middle): individual diffusion models can be flexibly composed to tackle novel scenarios not covered in the training data. Online Inference and Control (Right): The ego vehic… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Diffusion Model Can Ensure Kinematic Constraints. The RSTP method can plan safe trajectories involving large curvature 4a and gear shifting points (GSP) 4b. Deceleration decision made by diffusion composition Time Steps Collision with static obstacles [PITH_FULL_IMAGE…

**Figure 5.** Figure 5: The Composed Model Flexibly Avoids Obstacles in Unseen Scenarios by Adjusting Velocity. Comparison of F1TENTH’s velocity under different diffusion models in CS1. The composed model is able to decelerate to avoid static obstacles, whereas the dynamic model (DyM1) fails …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

SGTP: Sampling-based Game-Theoretic Planning for Real-Time Multi-Vehicle Autonomous Racing
cs.RO 2026-07 conditional novelty 6.0 of 10

SGTP uses GPU-sampled trajectories scored by a manually tuned game-aware cost inside an iterative best-response loop to generate blocking, contesting, and overtaking behaviors in multi-vehicle autonomous racing.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages · cited by 1 Pith paper

[1]

Practical search techniques in path planning for autonomous driving,

D. Dolgov, S. Thrun, M. Montemerlo, and J. Diebel, “Practical search techniques in path planning for autonomous driving,”ann arbor, vol. 1001, no. 48105, pp. 18–80, 2008. 1, 6

work page 2008
[2]

Optimization-based collision avoidance,

X. Zhang, A. Liniger, and F. Borrelli, “Optimization-based collision avoidance,”IEEE Transactions on Control Systems Technology, vol. 29, no. 3, pp. 972–983, 2020. 1, 3, 6

work page 2020
[3]

A rapid iterative trajectory planning method for automated parking through differential flatness,

Z. Li, L. Xie, C. Hu, and H. Su, “A rapid iterative trajectory planning method for automated parking through differential flatness,”Robotics and Autonomous Systems, vol. 182, p. 104816, 2024. 2, 3, 4, 6

work page 2024
[4]

Deep generative models in robotics: A survey on learning from multimodal demonstrations,

J. Urain, A. Mandlekar, Y . Du, M. Shafiullah, D. Xu, K. Fragkiadaki, G. Chalvatzaki, and J. Peters, “Deep generative models in robotics: A survey on learning from multimodal demonstrations,”arXiv preprint arXiv:2408.04380, 2024. 2

arXiv 2024
[5]

Planning with diffu- sion for flexible behavior synthesis,

M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffu- sion for flexible behavior synthesis,”arXiv preprint arXiv:2205.09991,

arXiv
[6]

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. 2

work page 2020
[7]

Diffusion policy: Visuomotor policy learning via ac- tion diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via ac- tion diffusion,”The International Journal of Robotics Research, p. 02783649241273668, 2023. 2

work page 2023
[8]

Autonomous vehicles on the edge: A survey on autonomous vehicle racing,

J. Betz, H. Zheng, A. Liniger, U. Rosolia, P. Karle, M. Behl, V . Krovi, and R. Mangharam, “Autonomous vehicles on the edge: A survey on autonomous vehicle racing,”IEEE Open Journal of Intelligent Trans- portation Systems, vol. 3, pp. 458–488, 2022. 2

work page 2022

Show all 19 references

[9]

Model predictive path-planning controller with potential function for emergency collision avoidance on highway driving,

P. Lin and M. Tsukada, “Model predictive path-planning controller with potential function for emergency collision avoidance on highway driving,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 4662– 4669, 2022. 2

2022
[10]

Potential based diffusion motion planning,

Y . Luo, C. Sun, J. B. Tenenbaum, and Y . Du, “Potential based diffusion motion planning,”arXiv preprint arXiv:2407.06169, 2024. 2, 4, 5

2024 arXiv
[11]

Implicit generation and modeling with energy based models,

Y . Du and I. Mordatch, “Implicit generation and modeling with energy based models,”Advances in neural information processing systems, vol. 32, 2019. 3

2019
[12]

A data-driven aggressive autonomous racing framework utilizing local trajectory planning with velocity prediction,

Z. Li, B. Zhou, C. Hu, L. Xie, and H. Su, “A data-driven aggressive autonomous racing framework utilizing local trajectory planning with velocity prediction,”arXiv preprint arXiv:2410.11570, 2024. 3

2024 arXiv
[13]

Reduce lap time for autonomous racing with curvature-integrated mpcc local trajectory planning method,

Z. Li, L. Xie, C. Hu, and H. Su, “Reduce lap time for autonomous racing with curvature-integrated mpcc local trajectory planning method,”arXiv preprint arXiv:2502.03695, 2025. 3

2025 arXiv
[14]

Global path planning for autonomous vehicles in off-road environment via an a-star algorithm,

Q. Liu, L. Zhao, Z. Tan, and W. Chen, “Global path planning for autonomous vehicles in off-road environment via an a-star algorithm,” International Journal of Vehicle Autonomous Systems, vol. 13, no. 4, pp. 330–339, 2017. 6

2017
[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. 6

2011
[16]

St-fmt*: A fast optimal global motion planning for mobile robot,

Z. Wu, Y . Chen, J. Liang, B. He, and Y . Wang, “St-fmt*: A fast optimal global motion planning for mobile robot,”IEEE Transactions on Industrial Electronics, vol. 69, no. 4, pp. 3854–3864, 2021. 6 THIS WORK HAS BEEN SUBMITTED TO THE IEEE FOR POSSIBLE PUBLICATION. COPYRIGHT MA...

2021
[17]

Sampling-based optimal kinodynamic planning with motion primitives,

B. Sakcak, L. Bascetta, G. Ferretti, and M. Prandini, “Sampling-based optimal kinodynamic planning with motion primitives,”Autonomous Robots, vol. 43, no. 7, pp. 1715–1732, 2019. 6

2019
[18]

Real-time loop closure in 2d lidar slam,

W. Hess, D. Kohler, H. Rapp, and D. Andor, “Real-time loop closure in 2d lidar slam,” in2016 IEEE international conference on robotics and automation (ICRA). IEEE, 2016, pp. 1271–1278. 7

2016
[19]

An aggressive cornering framework for autonomous vehicles combining trajectory planning and drift control,

W. Weng, C. Hu, Z. Li, H. Su, and L. Xie, “An aggressive cornering framework for autonomous vehicles combining trajectory planning and drift control,” in2024 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2024, pp. 2749–2755. 7

2024

Pith tools

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

[1] [1]

Practical search techniques in path planning for autonomous driving,

D. Dolgov, S. Thrun, M. Montemerlo, and J. Diebel, “Practical search techniques in path planning for autonomous driving,”ann arbor, vol. 1001, no. 48105, pp. 18–80, 2008. 1, 6

work page 2008

[2] [2]

Optimization-based collision avoidance,

X. Zhang, A. Liniger, and F. Borrelli, “Optimization-based collision avoidance,”IEEE Transactions on Control Systems Technology, vol. 29, no. 3, pp. 972–983, 2020. 1, 3, 6

work page 2020

[3] [3]

A rapid iterative trajectory planning method for automated parking through differential flatness,

Z. Li, L. Xie, C. Hu, and H. Su, “A rapid iterative trajectory planning method for automated parking through differential flatness,”Robotics and Autonomous Systems, vol. 182, p. 104816, 2024. 2, 3, 4, 6

work page 2024

[4] [4]

Deep generative models in robotics: A survey on learning from multimodal demonstrations,

J. Urain, A. Mandlekar, Y . Du, M. Shafiullah, D. Xu, K. Fragkiadaki, G. Chalvatzaki, and J. Peters, “Deep generative models in robotics: A survey on learning from multimodal demonstrations,”arXiv preprint arXiv:2408.04380, 2024. 2

arXiv 2024

[5] [5]

Planning with diffu- sion for flexible behavior synthesis,

M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffu- sion for flexible behavior synthesis,”arXiv preprint arXiv:2205.09991,

arXiv

[6] [6]

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. 2

work page 2020

[7] [7]

Diffusion policy: Visuomotor policy learning via ac- tion diffusion,

C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via ac- tion diffusion,”The International Journal of Robotics Research, p. 02783649241273668, 2023. 2

work page 2023

[8] [8]

Autonomous vehicles on the edge: A survey on autonomous vehicle racing,

J. Betz, H. Zheng, A. Liniger, U. Rosolia, P. Karle, M. Behl, V . Krovi, and R. Mangharam, “Autonomous vehicles on the edge: A survey on autonomous vehicle racing,”IEEE Open Journal of Intelligent Trans- portation Systems, vol. 3, pp. 458–488, 2022. 2

work page 2022