betaFeedback

WorksComputer Science

Learn-To-Propagate: Distributionally Robust Motion Planning for Nonlinear Uncertain Systems

Abstract

Motion planning for stochastic systems aims to find a sequence of control inputs that drives a system to a goal set while avoiding unsafe states with high probability. Existing approaches, however, are either restricted to linear dynamics with perfect knowledge of the underlying distributions or lack formal correctness guarantees. To address these limitations, we propose Learn-To-Propagate Distributionally-Robust Forest (L2P-DR-Forest), a distributionally-robust sampling-based planning algorithm for nonlinear stochastic systems with unknown initial state and process noise distributions. Our approach leverages recent advances in optimal transport to learn ambiguity sets that contain the true system distributions with high confidence. We develop an efficient method for propagating these sets through the system dynamics and for verifying their safety. We then integrate these techniques into a novel sampling-based planning algorithm to grow a forest of search trees in the system's state space while accounting for and mitigating the growth of ambiguity-sets radii. We prove that the algorithm is correct, i.e., every returned plan satisfies the specifications with the prescribed confidence, and, under some feasibility and distributional assumptions, is probabilistically complete. Evaluations on several robotic benchmarks demonstrate that our algorithm significantly outperforms existing methods while providing formal guarantees with high confidence.

CCS Concepts

• Computing methodologies → Motion path planning; • Theory of computation → Stochastic control and optimization; • Computer systems organization → Robotic control.

Introduction

Motion planning has become a central problem in robotics, enabling autonomous systems to operate safely in complex environments while avoiding obstacles. Applications span domains such as autonomous driving, aerial navigation, and robotic surgery. In this context, sampling-based algorithms have emerged as particularly successful approaches [7], where samples from the state and control spaces are used to efficiently construct search trees and compute collision-free trajectories. However, real-world robotic systems are often subject to uncertainty arising from process and measurement noise, external disturbances, and imperfect knowledge of the system dynamics. Unfortunately, existing approaches either rely on linear-Gaussian assumptions to obtain tractable safety guarantees or accommodate more general uncertainty at the expense of formal correctness guarantees [3, 8, 20, 30, 31, 34, 37]. This work aims to bridge this gap by developing a motion planning framework for uncertain systems that provides both correctness and completeness guarantees.

Among approaches that seek formal safety guarantees under uncertain disturbances, two broad directions have emerged. The first assumes that disturbances have bounded support, reducing safety checks to a worst-case reachability problem [9, 40]. While effective in some settings, these methods often become overly conservative when the support of the uncertainty is large. A second direction relies on distributionally robust optimization (DRO) [32], constructing sets of plausible probability distributions, referred to as ambiguity sets, and evaluating worst-case probabilities over these sets [18, 19, 29, 35, 36]. Yet, most existing DRO-based approaches are limited to linear or locally linearized systems, can be overly conservative, and either require a predefined ambiguity set or face challenges in learning sufficiently tight ambiguity sets from data. To address these limitations, we show how recent advances in Wasserstein-based distribution learning [13, 14] and nonlinear uncertainty propagation [12] can be combined to enable safe motion planning for nonlinear stochastic systems with unknown disturbance distributions.

In this paper, we propose Learn-To-Propagate Distributionally-Robust Forest (L2P-DR-Forest), a novel and computationally efficient motion-planning algorithm that combines DRO, sampling-based planning, and nonlinear uncertainty propagation to handle stochastic systems with unknown initial state and disturbance distributions. Leveraging recent results in optimal transport, our approach estimates a compact ambiguity set centered around a discrete distribution that is guaranteed, with high probability, to contain the true underlying system distributions. Critically, Wasserstein ambiguity sets admit a highly tractable propagation rule through Lipschitz continuous dynamics, with their radii expanding at most linearly with the Lipschitz constant of the dynamics. Consequently, these sets can be efficiently propagated while preserving computable bounds errs? on distributional uncertainty, avoiding the need for relearning at every propagation step. This property enables rapid exploration of the planning tree and substantially reduces the computational burden associated with exhaustive sampling. Combined with existing DRO results [15], this allows us to synthesize motion plans that carry formal collision-avoidance guarantees despite underlying distributional uncertainty.

L2P-DR-Forest grows a forest of trees in stages. Each new tree is rooted at a node of an existing tree, and the ambiguity set at this root is re-learned from samples rolled out through the controller leading to that node. Thus, the planner propagates ambiguity sets cheaply within each tree and re-learns tight ones only at the selected restart nodes. We prove that the algorithm is correct, i.e., any controller it returns solves the planning problem with the prescribed confidence, and probabilistically complete, i.e., if a control sequence is robustly feasible with some margin and the noise is sub-Gaussian, then, run for an unbounded number of stages with growing sample sizes, it returns a controller with probability one. We evaluate L2P-DR-Forest on a range of nonlinear robotics benchmarks. Our experiments demonstrate that the proposed approach achieves high task-success rates while enforcing safe controllers. In contrast, comparable methods that do not provide formal safety guarantees exhibit substantially lower reliability.

In summary, our main contributions are:

  • A novel distributionally robust sampling-based motion planning algorithm, L2P-DR-Forest, which, to the best of our knowledge, is the first motion planner that provides formal, high-confidence safety guarantees for nonlinear systems with unknown distributions.

  • An extension of data-driven distributionally robust algorithms to learn distributions with unbounded support from data, which is critical to formally reduce the uncertainty in our motion planner over forest expansion.

  • A formal proof that L2P-DR-Forest is correct and probabilistically complete.

  • An empirical evaluation including a neural model of a tractor trailer showing that our approach substantially outperforms comparable methods both in terms of success rate and planning time.

Related Works. Motion planning under uncertainty is commonly posed as a chance-constrained problem [3, 8, 24, 30, 31], where the probability of unsafety (e.g., collision) must satisfy a given upper bound at every time step. There are two inherent challenges in such problems: (i) uncertainty propagation through the dynamics, and (ii) bounding the probability of unsafety. Most of the existing approaches assume linear or linearized dynamics and Gaussian uncertainty to obtain tractable probability bounds or approximations. Extensions to non-Gaussian uncertainty use moment-based bounds [22, 37], which provide formal guarantees under restrictive assumptions, such as access to the relevant moments and their dynamics. Even when these assumptions hold, such methods can be overly conservative or computationally expensive.

For nonlinear systems, a common alternative is to assess safety statistically from sampled rollouts. This accommodates general dynamics and noise distributions, but naive Monte Carlo estimation can require many samples to resolve small collision probabilities, motivating variance-reduction techniques [21]. Sampling also underperforms model predictive control methods such as MPPI [39], whose original formulation does not provide finite-sample probabilistic safety certificates. Among sampling-based planners, ϵ\epsilon-RandUP-RRT [28, 40], which we use as a baseline in Section 5, checks collision avoidance using ϵ\epsilon-padded convex hulls of propagated particles. The finite-sample guarantees of its underlying reachable-set estimator depend on bounds on quantities such as Lipschitz constants, packing numbers, and sampling coverage, which can be difficult to obtain tightly in practice.

To mitigate these issues, scenario-based planning [10] has been introduced. It provides distribution-free, finite-sample guarantees from i.i.d. samples drawn from the true uncertainty distribution. These methods have been applied to nonlinear model predictive control and motion planning among uncertain obstacles, but they assume the robot dynamics are deterministic. These guarantees extend to nonconvex decision problems, although certification requires suitable feasibility conditions and bounds on the number of supporting scenarios. In contrast, in this work, we focus on offline, sampling-based planning for nonlinear systems with unknown disturbance distributions, with probabilistic safety guarantees derived from finite data.

Distributionally robust optimization addresses uncertainty about the underlying probability law by optimizing against the worst-case distribution within a set of plausible distributions, called an ambiguity set. In motion planning, ambiguity sets have been defined through means and covariances for linear systems with additive process noise [36], and as Wasserstein balls centered on empirical distributions constructed from samples of obstacle motion within model predictive control [19]. Closely related work [18] learns Wasserstein ambiguity tubes from trajectory data under unknown disturbance distributions and incorporates them into a probabilistically complete sampling-based planner. Its construction exploits linear or feedback-linearizable dynamics to decouple uncertainty propagation from nominal planning, allowing the ambiguity tube to be learned offline and reused during search. In contrast, we propagate learned ambiguity sets through nonlinear Lipschitz continuous dynamics and re-learn them during search, using a forest of trees to mitigate growth in their radii.

Preliminaries

Notation. For a vector x∈Rd\mathbf{x} \in\mathbb{R}^{d}, we denote by x(i)x^{(i)} its ii-th element. For a set X⊆Rd\mathcal{X} \subseteq\mathbb{R}^{d}, the indicator function for X\mathcal{X} is denoted as 1X(x):=1\mathbf{1}_{\mathcal{X}}(x) := 1 if x∈Xx \in\mathcal{X}; otherwise 00. We represent the set of indexes {1,…,M}\{1,\ldots,M\} by [M][M]. Given a Borel measurable set X⊆Rd\mathcal{X} \subseteq\mathbb{R}^{d}, we denote by B(X)\mathcal{B}(\mathcal{X}) the Borel sigma algebra over X\mathcal{X} and by P(X)\mathcal{P}(\mathcal{X}) the set of probability distributions on X\mathcal{X}. The symbol ⊗\otimes is used to represent the product measure between two distributions. The restriction of a distribution P∈P(X)\mathbb{P} \in\mathcal{P}(\mathcal{X}) to the subspace Xc⊆X\mathcal{X}_{c} \subseteq\mathcal{X} is denoted by P∣Xc\mathbb{P}|_{\mathcal{X}_{c}} and defined as P∣Xc(A)=P(A∩Xc)/P(Xc)\mathbb{P}|_{\mathcal{X}_{c}}(A) = \mathbb{P}(A \cap\mathcal{X}_{c})/\mathbb{P}(\mathcal{X}_{c}) for all A∈B(X)A \in\mathcal{B}(\mathcal{X}). For a probability distribution P∈P(X)\mathbb{P} \in\mathcal{P}(\mathcal{X}) and a measurable function g:X→Y⊆Rqg : \mathcal{X} \to\mathcal{Y} \subseteq\mathbb{R}^{q}, we denote the push-forward measure of P\mathbb{P} by gg as g#Pg_{\#}\mathbb{P} such that for all A∈B(Y)A \in\mathcal{B}(\mathcal{Y}), (g#P)(A)=P(g−1(A))(g_{\#}\mathbb{P})(A) = \mathbb{P}(g^{-1}(A)). We note that g#Pg_{\#}\mathbb{P} is still a probability distribution such that g#P∈P(Y)g_{\#}\mathbb{P} \in\mathcal{P}(\mathcal{Y}).

Wasserstein distance. We start by recalling the definition of the Wasserstein distance and Wasserstein ambiguity set, which we use to model distributional uncertainty in our framework. Let ρ≥1\rho\ge1, X⊆Rd\mathcal{X} \subseteq\mathbb{R}^{d}, and define Pρ(X)\mathcal{P}_{\rho}(\mathcal{X}) as the set of probability distributions with finite ρ\rho-th moment, i.e., all P∈P(X)\mathbb{P} \in\mathcal{P}(\mathcal{X}) such that ∫X∥x∥ρdP(x)<∞\int_{\mathcal{X}} \|x\|^{\rho} \mathrm{d}\mathbb{P}(x) < \infty, where ∥⋅∥\|\cdot\| is any fixed norm on Rd\mathbb{R}^{d}.

Definition 1 (ρ\rho-Wasserstein distance). For distributions P,Q∈Pρ(X)\mathbb{P},\mathbb{Q} \in\mathcal{P}_{\rho}(\mathcal{X}), the ρ\rho-Wasserstein distance between P\mathbb{P} and Q\mathbb{Q} is defined as

Wρ(P,Q):=(inf⁡γ∈Γ(P,Q)∫X×X∥x−y∥ρdγ(x,y))1ρ(1)W_{\rho}(\mathbb{P},\mathbb{Q}) := \left(\inf_{\gamma\in\Gamma(\mathbb{P},\mathbb{Q})} \int_{\mathcal{X}\times\mathcal{X}} \|x-y\|^{\rho} \mathrm{d}\gamma(x,y)\right)^{\frac{1}{\rho}} \tag*{(1)}

where Γ(P,Q)⊂Pρ(X×X)\Gamma(\mathbb{P},\mathbb{Q}) \subset\mathcal{P}_{\rho}(\mathcal{X}\times\mathcal{X}) represents the set of joint probability distributions with given marginals P\mathbb{P} and Q\mathbb{Q} (also known as couplings between P\mathbb{P} and Q\mathbb{Q}), i.e., for all γ∈Γ(P,Q)\gamma\in\Gamma(\mathbb{P},\mathbb{Q}), it holds that γ(A×X)=P(A)\gamma(A \times\mathcal{X}) = \mathbb{P}(A), γ(X×A)=Q(A)\gamma(\mathcal{X} \times A) = \mathbb{Q}(A) for all sets A∈B(X)A \in\mathcal{B}(\mathcal{X}).

Additionally, we define the ρ\rho-Wasserstein ambiguity set (ball) of radius ϵ≥0\epsilon\ge0 centered at the probability distribution P∈Pρ(X)\mathbb{P} \in\mathcal{P}_{\rho}(\mathcal{X}) as:

Bϵ(P):={Q∈Pρ(X):Wρ(P,Q)≤ϵ}.(2)\mathcal{B}_{\epsilon}(\mathbb{P}) := \left\{\mathbb{Q} \in\mathcal{P}_{\rho}(\mathcal{X}) : W_{\rho}(\mathbb{P},\mathbb{Q}) \le\epsilon\right\}. \tag*{(2)}

That is, Bϵ(P)\mathcal{B}_{\epsilon}(\mathbb{P}) contains all the probability distributions within ϵ\epsilon distance of P\mathbb{P} according to WρW_{\rho}.

Problem Formulation

We consider a robot with motion uncertainty, whose state evolution is described by the stochastic process

xk+1=f(xk,uk,ωk),x0∼P0,ωk∼Pω,(3)x_{k+1} = f(x_k,u_k,\omega_k), \qquad x_0 \sim\mathbb{P}_0,\qquad\omega_k \sim\mathbb{P}_{\omega}, \tag*{(3)}

where xk∈X⊂Rdx_k \in\mathcal{X} \subset\mathbb{R}^{d} is the state, uk∈U⊂Rqu_k \in\mathcal{U} \subset\mathbb{R}^{q} is the control action, and ωk∈W⊆Rw\omega_k \in\mathcal{W} \subseteq\mathbb{R}^{w} is the process noise at time step kk. Both X\mathcal{X} and U\mathcal{U} are assumed to be bounded sets. Vector field f:X×U×W→Xf:\mathcal{X}\times\mathcal{U}\times\mathcal{W}\to\mathcal{X} is a (possibly nonlinear) function, jointly Lipschitz continuous w.r.t. (x,u,ω)(x,u,\omega). The initial state x0x_0 of the robot is uncertain and distributed according to P0\mathbb{P}_0. In this paper, we make the following assumption about P0\mathbb{P}_0 and noise distribution Pω\mathbb{P}_{\omega}.

Assumption 1. We assume that, for any k≥0k \ge0, ωk∼Pω\omega_k \sim\mathbb{P}_{\omega} are i.i.d. and independent from x0∼P0x_0 \sim\mathbb{P}_0, with P0∈Pρ(X)\mathbb{P}_0 \in\mathcal{P}_{\rho}(\mathcal{X}) and Pω∈Pρ(W)\mathbb{P}_{\omega} \in\mathcal{P}_{\rho}(\mathcal{W}) being unknown, but with finite ρ\rho-moments. Furthermore, we assume access to two fixed, offline datasets, mutually independent of each other: Nx0>0N_{x_0} > 0 i.i.d. samples Dx0\mathcal{D}_{x_0} of x0∼P0x_0 \sim\mathbb{P}_0, and Nω>0N_{\omega} > 0 i.i.d. samples Dω\mathcal{D}_{\omega} of ω∼Pω\omega\sim\mathbb{P}_{\omega}.

The finite ρ\rho-moment assumption ensures that the ρ\rho-Wasserstein distance between the probability measures considered in this work is finite, while the i.i.d. assumption excludes temporal correlations in the process noise. The availability of samples from Pω\mathbb{P}_{\omega}, although common in data-driven stochastic and distributionally robust settings [2, 17, 18], is stronger. Nevertheless, it can be practically satisfied through, e.g., observations of the state and control. In practice, it suffices for ff to be injective over only a subset of X\mathcal{X} as discussed in [17].

To make probabilistic statements about System (3), we define the probability law induced by a fixed control sequence. In particular, for k≥1k \ge1 and a given sequence of control inputs uk−1:=(u0,…,uk−1)∈Uku_{k-1} := (u_0,\ldots,u_{k-1}) \in\mathcal{U}^{k}, the distribution of System (3) at time kk is denoted as Pxkuk−1\mathbb{P}^{u_{k-1}}_{x_k} and is defined by propagating the noise and input distribution through the dynamics kk times using uk−1u_{k-1}, i.e.,

Px1u0=f(P0⊗Pω),Pxkuk−2=f(Pxk−1uk−2⊗Pω)for k>1.\mathbb{P}^{u_0}_{x_1} = f(\mathbb{P}_0 \otimes\mathbb{P}_{\omega}), \qquad \mathbb{P}^{u_{k-2}}_{x_k} = f(\mathbb{P}^{u_{k-2}}_{x_{k-1}} \otimes\mathbb{P}_{\omega}) \quad\text{for } k>1.

The robot operates in a workspace with obstacles and is subject to state constraints, e.g., steering angle or velocity limits. We denote the set of such states O⊂X\mathcal{O} \subset\mathcal{X}. The task of the robot is to avoid these states to get to a target set denoted by T⊂X\mathcal{T} \subset\mathcal{X}. We are now ready to define our problem.

Problem 1 (Motion Planning under Unknown Noise). Given the robotic system in (3), the measurable obstacle and target sets O,T⊆X\mathcal{O},\mathcal{T} \subseteq\mathcal{X} introduced above, unsafety probability thresholds ϵO,ϵT∈(0,1)\epsilon_{\mathcal{O}},\epsilon_{\mathcal{T}} \in(0,1), and confidence-level threshold β∈(0,1)\beta\in(0,1), find a control sequence uH−1∈UHu_{H-1} \in\mathcal{U}^{H} for some horizon H∈NH \in\mathbb{N} such that

Pxkuk−1(O)≤ϵO∀k∈[H],andPxHuH−1(T)≥1−ϵT(4)\mathbb{P}^{u_{k-1}}_{x_k}(\mathcal{O}) \le\epsilon_{\mathcal{O}} \quad\forall k \in[H], \qquad\text{and} \qquad\mathbb{P}^{u_{H-1}}_{x_H}(\mathcal{T}) \ge1-\epsilon_{\mathcal{T}} \tag*{(4)}

References

  1. [1]Alessandro Abate, Omid Akbarzadeh, Henk A.P. Blom, Sofie Haesaert, Sina Hassani, Abolfazl Lavaei, Frederik Baymler Mathiesen, Rahul Misra, Amy Nejati, Mathis Niehane, Fie Ørum, Anne Remke, Behrad Samari, Ruohan Wang, Rafal Wisniewski, Ben Wooding, and Mahdí Zaker. 2025. ARCH-COMP25 Category Report: Stochastic Models. In Proceedings of 12th Int. Workshop on Applied Verification for Continuous and Hybrid Systems (EPiC Series in Computing, Vol. 108). EasyChair, 122–151.
  2. [2]Thomas Badings, Licio Romao, Alessandro Abate, David Parker, Hasan A Poonawala, Marielle Stoelinga, and Nils Jansen. 2023. Robust control for dynamical systems with non-gaussian noise via formal abstractions. Journal of Artificial Intelligence Research 76 (2023), 341–391.
  3. [3]Lars Blackmore, Masahiro Ono, and Brian C Williams. 2011. Chance-constrained optimal path planning with obstacles. IEEE Transactions on Robotics 27, 6 (2011), 1080–1094.DOI
  4. [4]Nicolas Bonneel, Julien Rabin, Gabriel Peyré, and Hanspeter Pfister. 2015. Sliced and radon wasserstein barycenters of measures. Journal of Mathematical Imaging and Vision 51, 1 (2015), 22–45.
  5. [5]Matthias Bos, Bastiaan Vandewal, Wim Decré, and Jan Swevers. 2023. Mpc-based motion planning for autonomous truck-trailer maneuvering. IFAC-PapersOnLine 56, 2 (2023), 4877–4882.
  6. [6]Julien Calbert, Adrien Banse, Benoît Legat, and Raphaël M. Jungers. 2024. Dyonsys.jl: a Modular Platform for Smart Symbolic Control. Proceedings of the JuliaCon Conferences 6, 66 (2024), 160.DOI
  7. [7]Juan Cortés and Thierry Siméon. 2020. Sampling-Based Tree Planners (RRT, EST, and Variations). 1–9.DOI
  8. [8]Siyu Dai, Shawn Schaffert, Ashkan Jasour, Andreas Hofmann, and Brian Williams. 2019. Chance constrained motion planning for high-dimensional robots. In 2019 International Conference on Robotics and Automation (ICRA). IEEE, 8585–8811.arxiv.org/abs/1811.03073
  9. [9]Claus Danielson, Karl Berntorp, Avishai Weiss, and Stefano Di Cairano. 2020. Robust motion planning for uncertain systems with disturbances using the invariant-set motion planner. IEEE Trans. Automat. Control 65, 10 (2020), 4456–4463.DOI
  10. [10]Oscar De Groot, Laura Ferranti, Dariu M Gavrila, and Javier Alonso-Mora. 2025. Scenario-based motion planning with bounded probability of collision. The International Journal of Robotics Research 44, 9 (2025), 1507–1525.arxiv.org/abs/2307.01070
  11. [11]Mahyar Fazlyab, Alexander Robey, Hamed Hassani, Manfred Morari, and George J. Pappas. 2019. Efficient and Accurate Estimation of Lipschitz Constants for Deep Neural Networks. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 32.arxiv.org/abs/1906.04893
  12. [12]Eduardo Figueiredo, Steven Adams, Peyman Mohajerin Esfahani, and Luca Laurenti. 2025. Efficient Uncertainty Propagation with Guarantees in Wasserstein distance. arXiv preprint arXiv:2506.08689 (2025).arxiv.org/abs/2506.08689
  13. [13]Eduardo Figueiredo, Steven Adams, and Luca Laurenti. 2026. Efficient Distribution Learning with Error Bounds in Wasserstein Distance. arXiv preprint arXiv:2602.08063 (2026).arxiv.org/abs/2602.08063
  14. [14]Nicolas Fournier and Arnaud Guillin. 2015. On the rate of convergence in Wasserstein distance of the empirical measure. Probability theory and related fields 162, 3 (2015), 707–738.arxiv.org/abs/1312.2128
  15. [15]Rui Gao and Anton Kleywegt. 2023. Distributionally robust stochastic optimization with Wasserstein distance. Mathematics of Operations Research 48, 2 (2023), 603–655.arxiv.org/abs/1604.02199
  16. [16]Ibon Gracia, Dimitris Boskos, Morteza Lahijanian, Luca Laurenti, and Manuel Mazo Jr. 2025. Efficient strategy synthesis for switched stochastic systems with distributional uncertainty. Nonlinear Analysis: Hybrid Systems 55 (2025), 101554.arxiv.org/abs/2212.14260
  17. [17]Ibon Gracia, Dimitris Boskos, Luca Laurenti, and Morteza Lahijanian. 2024. Data-driven strategy synthesis for stochastic systems with unknown nonlinear disturbances. In 6th Annual Learning for Dynamics & Control Conference. PMLR, 1633–1645.arxiv.org/abs/2406.09704
  18. [18]Ibon Gracia, Qi Heng Ho, Luca Laurenti, and Morteza Lahijanian. 2026. Probably Safe Motion Planning Under Unknown Disturbances. arXiv preprint arXiv:2605.26625 (2026).arxiv.org/abs/2605.26625
  19. [19]Astghik Hakobyan and Insoon Yang. 2020. Wasserstein distributionally robust motion planning and control with safety constraints using conditional value-at-risk. In 2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 490–496.DOI
  20. [20]Qi Heng Ho, Zachary N Sunberg, and Morteza Lahijanian. 2022. Gaussian belief trees for chance constrained asymptotically optimal motion planning. In 2022 International Conference on Robotics and Automation (ICRA). IEEE, 11029–11035.DOI
  21. [21]Lucas Janson, Edward Schmerling, and Marco Pavone. 2017. Monte Carlo motion planning for robot trajectory optimization under uncertainty. In Robotics Research: Volume 2. Springer, 343–361.arxiv.org/abs/1504.08053
  22. [22]Ashkan M Jassour and Brian C Williams. 2019. Risk Contours Map for Risk Bounded Motion Planning under Perception Uncertainties. In Robotics: Science and Systems. 22–26.
  23. [23]Soheil Koulouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo Rohde. 2019. Generalized sliced wasserstein distances. Advances in neural information processing systems 32 (2019).
  24. [24]Mangal Kothari and Ian Postlethwaite. 2013. A probabilistically robust path planning algorithm for UAVs using rapidly-exploring random trees. Journal of Intelligent & Robotic Systems 71, 2 (2013), 231–253.DOI
  25. [25]Tin Lai and Fabio Ramos. 2022. Adaptively Exploits Local Structure With Generalised Multi-Trees Motion Planning. IEEE Robotics and Automation Letters 7, 2 (2022), 1111–1117.DOI
  26. [26]Tin Lai, Fabio Ramos, and Gilad Francis. 2019. Balancing Global Exploration and Local-connectivity Exploitation with Rapidly-exploring Random disjointed-Trees. In 2019 International Conference on Robotics and Automation (ICRA). 5537–5543.arxiv.org/abs/1810.03749
  27. [27]Steven M LaValle and James J Kuffner. 2001. Randomized kinodynamic planning. The International Journal of Robotics Research 20, 5 (2001), 378–400.
  28. [28]Thomas Lew, Lucas Janson, Riccardo Bonalli, and Marco Pavone. 2022. A simple and efficient sampling-based algorithm for general reachability analysis. In Learning for Dynamics and Control Conference. PMLR, 1086–1099.
  29. [29]Lars Lindemann, Matthew Cleaveland, Yiannis Kantaro, and George J Pappas. 2021. Robust motion planning in the presence of estimation uncertainty. In 2021 60th IEEE Conference on Decision and Control (CDC). IEEE, 5205–5212.arxiv.org/abs/2108.11983
  30. [30]Brandon Luders, Mangal Kothari, and Jonathan How. 2010. Chance constrained RRT for probabilistic robustness to environmental uncertainty. In AIAA guidance, navigation, and control conference. 8160.DOI
  31. [31]Brandon D Luders, Sertac Karaman, and Jonathan P How. 2013. Robust sampling-based motion planning with asymptotic optimality guarantees. In AIAA Guidance, Navigation, and Control (GNC) Conference. 5097.DOI
  32. [32]Peyman Mohajerin Esfahani and Daniel Kuhn. 2018. Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations. Mathematical Programming 171, 1 (2018), 115–166.arxiv.org/abs/1505.05116
  33. [33]E. Plaku, K.E. Bekris, B.Y. Chen, A.M. Ladd, and L.E. Kavraki. 2005. Sampling-based roadmap of trees for parallel motion planning. IEEE Transactions on Robotics 21, 4 (2005), 597–608.DOI
  34. [34]Konstantinos Prattis, Luca Laurenti, and Azita Dabiri. 2026. Stochastic Nonlinear Model Predictive Control with Gaussian Mixture Uncertainty Propagation. arXiv preprint arXiv:2608.29272 (2026).arxiv.org/abs/2608.29272
  35. [35]Venkatraman Renganathan, Iman Shames, and Tyler H Summers. 2020. Towards integrated perception and motion planning with distributionally robust risk constraints. IFAC-PapersOnLine 53, 2 (2020), 15350–15536.arxiv.org/abs/2002.02928
  36. [36]Tyler Summers. 2018. Distributionally robust sampling-based motion planning under uncertainty. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 6518–6523.
  37. [37]Allen Wang, Ashkan Jasour, and Brian C Williams. 2020. Non-gaussian chance-constrained trajectory planning for autonomous vehicles under agent uncertainty. IEEE Robotics and Automation Letters 5, 4 (2020), 6041–6048.arxiv.org/abs/2002.10999
  38. [38]Jonathan Weed and Francis Bach. 2019. Sharp asymptotic and finite-sample rates of convergence of empirical measures in Wasserstein distance. Bernoulli 25, 4A (2019), 2620–2648.arxiv.org/abs/1707.00087
  39. [39]Grady Williams, Andrew Aldrich, and Evangelos A Theodorou. 2017. Model predictive path integral control: From theory to parallel computation. Journal of Guidance, Control, and Dynamics 40, 2 (2017), 344–357.DOI
  40. [40]Albert Wu, Thomas Lew, Kiril Solovey, Edward Schmerling, and Marco Pavone. 2022. Robust-RRT: Probabilistically-complete motion planning for uncertain nonlinear systems. In The International Symposium of Robotics Research. Springer, 538–554.arxiv.org/abs/2205.07728

Paper details

vlnc:W_n5j9mg6q
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents