Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

NeHMO: Neural Hamilton-Jacobi Reachability Learning for Decentralized Safe Multi-Arm Motion Planning

T0 review · 4 major / 4 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 Neural Hamilton-Jacobi reachability learns a per-arm safety value function so that each manipulator in a shared workspace plans against worst-case behavior of the others, achieving safe real-time decentralized motion planning.

desk verdict A useful engineering combination of neural HJR and decentralized trajectory optimization, but the worst-case safety claim outruns the evidence. read the letter →

arxiv 2507.13940 v2 pith:Z2VMJBYS submitted 2025-07-18 cs.RO

Qingyi Chen , Zachary Kingston , Ahmed H. Qureshi This is my paper
classification cs.RO
keywords Hamilton-Jacobireachabilitymulti-armmotionplanningdecentralizedneuralvaluefunctionworst-casesafetytrajectoryoptimizationmanipulatorrobots
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 claims that a neural approximation of the Hamilton-Jacobi-Isaacs (HJI) value function—the solution to a zero-sum game where one robot arm avoids collision while the other acts as a worst-case adversary—can serve as a per-arm safety constraint in decentralized motion planning. If correct, this would let each manipulator plan its own trajectory in a shared workspace without needing to know or predict the other arms' control policies, closing a gap left by methods that rely on behavior prediction or explicit coordination. NeHMO learns this value function with two physics-informed priors—symmetry reduction and a boundary-condition residual—and embeds it in a real-time receding-horizon optimizer. The empirical case is that on a dual-UR5 system it reaches 92% success and 2% collisions, and on a five-UR5 (30-DoF) system it keeps 66% success and 8% collisions, while baselines deteriorate sharply.

What carries the argument

The load-bearing object is the Hamilton-Jacobi-Isaacs value function $V(t,x)$ for the pairwise collision-avoidance game, defined as the viscosity solution of the HJI PDE with terminal condition $V(T,x) = \ell(x)$, where $\ell$ is a signed-distance collision function. Its zero level set separates states from which the controlled arm can guarantee avoidance for the horizon from states where collision is inevitable under worst-case disturbance. The learning machinery is a self-supervised neural network trained against the HJI PDE and terminal condition, augmented by two priors: (i) symmetry invariance—if a differentiable bijection $f$ preserves both the boundary condition and the Hamiltonian, then $V(t,x) = V(t,f(x))$, so training samples only need to cover half the state space; and (ii) residual parameterization—the network fits $V_{\mathrm{res}} = V - \ell(x)$ rather than the full value, reducing the burden of representing complex configuration-space geometry. The planner layer turns the learned function into a constraint: at each receding-horizon step, it requires $V_\theta(T - t_{\mathrm{safe}}, \cdot) > \epsilon$ for each other arm and uses the closed-form worst-case disturbance derived from the value gradient. This coupling of the learned representation to the optimization variable is what makes decentralized defensive planning possible.

What would settle it

Run NeHMO on a simple 2-link arm system where the true HJI value can be computed numerically, and count the states on planned trajectories where the learned value has the wrong sign relative to ground truth (e.g., $V_\theta \le 0$ while true $V > 0$); if such sign mismatches occur on or near the paths NeHMO chooses, the safety constraint is not enforcing the claimed worst-case guarantee. Alternatively, on the dual-UR5 setup, let the other arm execute the exact worst-case policy derived from $V_\theta$ (the minimizer in (15)) rather than a naive policy, and test whether NeHMO's collision rate rises significantly above the reported 2%.

Watch

Extended reading notes

Core claim

The central discovery is that the avoid-set value function of a pairwise HJI reachability game can be learned in high-dimensional manipulator configuration spaces and then used directly as a constraint in decentralized trajectory optimization. The paper shows that by modeling one arm as the control and the other as a disturbance in the coupled dynamics, the HJI value function $V(t,x)$ encodes, for every joint configuration, whether the first arm can guarantee collision avoidance for the remaining horizon against any behavior of the second. The learning pipeline enforces the terminal condition and the HJI PDE while exploiting two structural priors: a symmetry map $f$ that leaves value invariant (so training can be halved) and a residual decomposition $V_{\mathrm{res}} = V - \ell(x)$ that isolates the geometric boundary term. The planner then solves a receding-horizon problem that keeps $V_\theta(T - t_{\mathrm{safe}}, x_1(t_{\mathrm{plan}}), x_i(t_{\mathrm{plan}})) > \epsilon$ for every other arm $i$, using the value gradient to pick the worst-case action of each neighbor. The paper reports that this scheme outperforms centralized and learning-based baselines, stays real-time at roughly 25 Hz for dual-UR5, and provides defensive behavior even when the other arm follows a reckless policy.

Load-bearing premise

The load-bearing premise is that the learned neural value function $V_\theta$ accurately approximates the true HJI value in high-dimensional configuration spaces, where no ground-truth solution exists and no formal certificate is supplied; if $V_\theta$ drifts, the safety constraint and the inferred worst-case disturbance are not the true ones, and the paper's collision-avoidance claim loses its foundation.

Editorial extensions

If this is right

  • Multi-arm systems can be planned without central coordination: each arm solves its own optimization using the shared learned value function, so runtime scales with the number of arms rather than the joint configuration space.
  • Safety no longer depends on predicting or trusting other agents' policies; treating them as worst-case disturbances makes the plan valid against arbitrary behavior, including reckless or adversarial actors.
  • The same learned value function transfers across different numbers of arms and obstacle layouts because it is learned pairwise and decoupled from the global scene.
  • Physics-informed priors make neural HJR practical in high-dimensional articulated systems, where grid-based solvers are intractable and prior neural methods struggle.
  • The conservative worst-case model trades path length for safety, but the framework still achieves real-time control (roughly 25 Hz on dual-UR5), suggesting it can be used in dynamic environments.

Reading between the lines

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

  • If the learned value function could be certified (e.g., with Lipschitz bounds or conformal margin selection), the same planner would yield formal safety guarantees rather than the empirical ones reported; the paper identifies this as future work.
  • The pairwise worst-case game is likely conservative in dense scenes; a general-sum or intent-aware variant that estimates other arms' goals while keeping a worst-case fallback could recover path efficiency without sacrificing safety.
  • The symmetry and residual priors are not UR5-specific; they should extend to other articulated platforms, and the framework may apply to heterogeneous teams and mobile manipulators where pairwise reachability is tractable.
  • A direct test of the safety claim would be to run NeHMO against an adversary that actively seeks the network's blind spots—configurations where $V_\theta$ overestimates the true value—and measure whether collisions rise accordingly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes NeHMO, a decentralized multi-arm motion planning framework built on a learned neural approximation of the Hamilton-Jacobi-Isaacs (HJI) reachability value function. The value function is trained self-supervised with two physics-informed priors: exploiting symmetry to halve the training state space, and learning only the time-dependent residual relative to the boundary condition. This learned value function is then used in a receding-horizon trajectory optimization problem as a safety constraint, with the worst-case other-agent action inferred from the value gradient. The paper validates the training pipeline on Air3D and a 2-link planar arm against numerical HJI solvers, and evaluates the full planner in simulation on dual-UR5, three- and five-UR5, and heterogeneous UR5-Kinova systems, including a scenario where one arm runs a fixed reckless policy. The main claim is that NeHMO provides scalable, data-efficient, real-time decentralized safety that outperforms existing baselines.

Significance. If the central claim is substantiated, the paper would make a useful contribution: it combines a principled worst-case differential-game formulation with a scalable neural approximation and demonstrates a decentralized planner that does not rely on behavior prediction or coordination protocols. The low-dimensional validation against numerical solvers is a real strength, and the planning results on 12-30 DoF systems are internally consistent and suggestive. The authors are also honest about the loss of formal safety guarantees and about the conservatism of the approach. However, the key load-bearing evidence for the high-dimensional claim is the planning experiments, and these currently do not directly stress-test the worst-case property that the HJI formulation is meant to provide. The main contribution is therefore promising but incompletely validated as stated.

major comments (4)
  1. [§IV-C, Eq. (15), Table IV] The adversarial evaluation uses a single fixed reckless Kinova policy, not the worst-case disturbance d* defined in Eq. (15) that the HJI game is quantified over. A value function can be substantially wrong and still avoid one particular fixed policy, so this experiment does not validate the central worst-case safety claim. Moreover, because the same V_theta is used to define both the safety constraint (16) and the inferred adversary (15), the method is never checked for self-consistency. Please evaluate NeHMO in closed-loop against the adversarial policy generated from Eq. (15) (or from a separately optimized worst-case policy), and report the resulting success and collision rates; without such an experiment, the claims in Sections III and IV that the learned value function encodes worst-case inter-arm safety are not supported.
  2. [Theorem 1 and Appendix I, Eq. (21)] The symmetry property that halves the training data is asserted for the dual-UR5 system and the UR5-Kinova system, but the hypotheses of Theorem 1 are not verified for these systems. In particular, the equality l(f(x)) = l(x) is not demonstrated for the signed-distance collision boundary under the point reflection in Eq. (21); for heterogeneous arm geometries this invariance is not obvious. If the symmetry does not hold, training only on X_train and using the inference rule (9) is invalid. Please provide a numerical or analytic verification of l(f(x)) = l(x) (and of the Hamiltonian symmetry condition) on a validation grid, or restrict the symmetry reduction to systems where it provably applies.
  3. [§IV-B, Tables II-IV] Success rate and collision rate are binomial proportions, but they are reported as single numbers without confidence intervals, multiple seeds, or a per-condition trial breakdown. In Table II, for example, NeHMO's 92% versus CHOMP's 86% over what appears to be 100 trials is within plausible sampling noise. The comparison claims against baselines therefore need error bars or confidence intervals and a statement of the number of independent runs per scenario. Additionally, the safety margin epsilon and the horizons t_plan and t_safe in Section III-C are fixed at 0.05, 0.1, and 0.3 without any sensitivity analysis, so it is unclear how dependent the reported success and collision rates are on these hand-chosen values.
  4. [Appendix I, last paragraph; Section IV-A] For the UR5 and UR5-Kinova systems, the boundary condition l(x) used in the HJI training is itself an MLP approximation of the signed distance function. The error of this network is not reported, even though the residual formulation V_res = V - l(x) and the terminal condition V(T,x) = l(x) mean that any error in the learned l directly shifts the zero level set of the value function and hence the safety constraint (16). Since Section IV-A explicitly states that ground-truth HJI solutions are intractable for these systems, the only check on this approximation chain is the planning experiments. Please report the approximation error of the learned l against analytic link-pair distances, or include it in the validation study in Section IV-A.
minor comments (4)
  1. [Table II] The NeHMO row reads '92 236 ± 35 8.5 ± 4.6', which is ambiguous because the collision rate and planning time are not visually separated; insert explicit column separators so that SR, CR, Time, and MPL are clearly distinguishable.
  2. [Algorithm 1] The 'Loop' on line 2 has no termination condition; specify the stopping criterion for the planning loop (e.g., until the goal is reached or a timeout occurs).
  3. [Appendix I] The software dependencies hj_reachability and PyTorch Kinematics are cited without version numbers; for reproducibility, please provide the exact versions or commits used.
  4. [Section III-C] Equation (13) uses d_i in the evolution of x_i but the constraint in (16) is evaluated for each pair (1,i); clarify that the worst-case disturbance is computed pairwise for every other arm and that the optimization problem simultaneously enforces all m-1 safety constraints.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the value function is trained against the HJI PDE and checked against external numerical solvers, not fitted to downstream success.

full rationale

NeHMO's derivation chain is self-contained. The learned value function V_theta is trained by minimizing the HJI PDE residual subject to the terminal boundary condition V(T,x)=l(x), and the low-dimensional validation in Table I is compared directly against numerical ground truth from hj_reachability and PyTorch Kinematics, which are external solvers. The symmetry prior is justified by the standard uniqueness of viscosity solutions, not by a self-citation, and the residual parameterization V_res = V - l(x) is a network reparameterization rather than a fitted target. The downstream planner imposes V_theta(T - t_safe, ...) > epsilon with user-set margins (epsilon=0.05, t_plan=0.1, t_safe=0.3), and no reported success or collision metric is obtained by fitting V_theta to that metric. The use of V_theta's gradient in Eq. (15) to infer an adversarial action and in Eq. (16) as a safety certificate is a standard HJI closed-loop construction, not a definitional reduction: if V_theta were inaccurate, the guarantee would degrade, which is precisely the limitation conceded in Section V and left for future certification. This is an approximation-accuracy and evaluation-coverage concern, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

NeHMO introduces learned function approximators (the HJR value network and the SDF MLP), but these are models, not new physical entities. There are no new forces, particles, or conserved quantities. The main burden on the reader is a set of priors and assumptions about symmetry, pairwise decomposition, and convergence of self-supervised PDE training.

free parameters (4)
  • safety margin epsilon = 0.05
    User-defined threshold in constraint (16); no sensitivity analysis or principled selection is reported.
  • planning horizon t_plan = 0.1 s
    User-defined receding horizon in equations (11)-(14); the choice affects how far ahead each arm plans and thus conservatism.
  • safety horizon t_safe = 0.3 s
    User-defined look-ahead safety duration in constraint (16); the authors note conservatism is governed by t_plan and t_safe.
  • neural network hyperparameters = not reported
    Architecture, learning rate, loss weights, and training schedule for V_theta and the SDF MLP are omitted, which blocks reproduction.
assumptions (5)
  • standard math Standard HJI conditions: Lipschitz continuous dynamics, compact control and disturbance sets, nonanticipative strategies, and the Isaacs condition.
    Invoked in Section III-A to guarantee existence and uniqueness of the viscosity solution of the HJI PDE, and the validity of the optimal control expression (6).
  • domain assumption Pairwise interaction decomposition: safety of the m-arm system follows from enforcing pairwise HJI constraints (16) for every pair of arms.
    The planner only evaluates V_theta on pairwise states, treating one arm as evader and the other as adversarial disturbance; no theorem shows that pairwise worst-case avoidance implies global multi-arm safety.
  • ad hoc to paper Exact symmetry of the value function for the dual-manipulator systems, equation (21), including the UR5-Kinova pair.
    Equation (21) is assumed rather than verified for UR5 and Kinova geometry. The boundary condition relies on a learned SDF whose exact invariance under f is not established, and the appendix models the Kinova with 6 joints although the Kinova Gen3 has 7.
  • domain assumption The learned SDF MLP approximates the signed distance function and its gradient accurately enough to define the HJI boundary condition.
    Appendix I-C states the signed distance and gradient are hard to obtain, so an MLP is trained following [37]; no approximation error bound is given.
  • domain assumption The self-supervised training converges to the viscosity solution of (4) in high-dimensional systems where no ground truth is available.
    Section IV-A says numerical ground truth is intractable for UR5 systems and evaluation is indirect through planning experiments; convergence is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NeHMO: Neural Hamilton-Jacobi Reachability Learning for Decentralized Safe Multi-Arm Motion Planning." pith.science (2026). https://pith.science/paper/Z2VMJBYS

@misc{pith2026250713940,
  author       = {Pith},
  title        = {Pith review of: NeHMO: Neural Hamilton-Jacobi Reachability Learning for Decentralized Safe Multi-Arm Motion Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z2VMJBYS}},
  note         = {Machine review of arXiv:2507.13940}
}
read the original abstract

Safe multi-arm motion planning is a challenging problem in robotics due to its high dimensionality, coupled configuration space, and complex collision constraints. Centralized planners are capable of coordinating all arms but often face scalability limitations, restricting applicability in real-time settings. On the other hand, decentralized methods are scalable and recent deep learning-based approaches have shown promising results. However, these depend on accurate behavior prediction or coordination protocols and may fail when other arms act unpredictably. To address these challenges, we introduce a neural Hamilton-Jacobi Reachability (HJR) learning-based approach to approximate a safety value function that captures worst-case inter-arm safety constraints. We further develop a decentralized trajectory optimization framework that uses the learned HJR representation for real-time planning. The proposed method is scalable and data-efficient, generalizes across multi-manipulator systems, and outperforms state-of-the-art baselines on challenging multi-arm motion planning tasks.

Figures

Figures reproduced from arXiv: 2507.13940 by the authors.

Figure 1
Figure 1. Our method, NeHMO, controlling a 12-dimensional dual-UR5 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. System setups and their symmetries. The primal state is depicted in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An example of NeHMO planning for the dual-UR5 system in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: An example five-UR5 scenario. Table III reports the results. As the number of manipula￾tors increases, the joint configuration space becomes more constrained due to the increase of inter-arm interactions, resulting in a high collision rate for both the naive planner an…
Figure 5
Figure 5. Figure 5: An example of NeHMO avoiding collision for the UR5 arm while the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Manifold-constrained Hamilton-Jacobi Reachability Learning for Decentralized Multi-Agent Motion Planning

    cs.RO 2025-11 conditional novelty 6.0 of 10

    HaMMAR learns manifold-constrained Hamilton-Jacobi reachability value functions and uses them for decentralized collision-free multi-robot motion planning under task constraints.

Reference graph

Works this paper leans on

37 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    dRRT*: Scalable and informed asymptotically-optimal multi-robot motion planning,

    R. Shome, K. Solovey, A. Dobson, D. Halperin, and K. E. Bekris, “dRRT*: Scalable and informed asymptotically-optimal multi-robot motion planning,”Autonomous Robots, vol. 44, no. 3, pp. 443–467, 2020

    2020

  2. [2]

    Representation- optimal multi-robot motion planning using conflict-based search,

    I. Solis, J. Motes, R. Sandstr ¨om, and N. M. Amato, “Representation- optimal multi-robot motion planning using conflict-based search,” IEEE Robot. Autom. Lett., vol. 6, no. 3, pp. 4608–4615, 2021

    2021

  3. [3]

    CHOMP: Gradient optimization techniques for efficient motion planning,

    N. Ratliff, M. Zucker, J. A. Bagnell, and S. Srinivasa, “CHOMP: Gradient optimization techniques for efficient motion planning,” in IEEE Int. Conf. Robot. Autom., 2009, pp. 489–494

    2009

  4. [4]

    Learning a decentralized multi-arm motion planner,

    H. Ha, J. Xu, and S. Song, “Learning a decentralized multi-arm motion planner,” inConf. on Robotic Learning, 2020

    2020

  5. [5]

    Hamilton-Jacobi reachability: A brief overview and recent advances,

    S. Bansal, M. Chen, S. Herbert, and C. J. Tomlin, “Hamilton-Jacobi reachability: A brief overview and recent advances,” inIEEE Annu. Conf. on Decision and Control, 2017, pp. 2242–2253

    2017

  6. [6]

    Multi-vehicle collision avoidance via Hamilton-Jacobi reachability and mixed integer pro- gramming,

    M. Chen, J. C. Shih, and C. J. Tomlin, “Multi-vehicle collision avoidance via Hamilton-Jacobi reachability and mixed integer pro- gramming,” inIEEE Conf. on Decision and Control, 2016, pp. 1695– 1700

    2016

  7. [7]

    Infusing reachability-based safety into planning and control for multi-agent interactions,

    X. Wang, K. Leung, and M. Pavone, “Infusing reachability-based safety into planning and control for multi-agent interactions,” in IEEE/RSJ Int. Conf. on Intell. Robots and Syst., 2020, pp. 6252–6259

    2020

  8. [8]

    Deepreach: A deep learning approach to high-dimensional reachability,

    S. Bansal and C. J. Tomlin, “Deepreach: A deep learning approach to high-dimensional reachability,” inIEEE Int. Conf. Robot. Autom., 2021, pp. 1817–1824

    2021

Show all 37 references
  1. [9]

    RRT-Connect: An efficient approach to single-query path planning,

    J. J. Kuffner and S. M. LaValle, “RRT-Connect: An efficient approach to single-query path planning,” inIEEE Int. Conf. Robot. Autom., vol. 2, 2000, pp. 995–1001

    2000

  2. [10]

    A review of path-planning approaches for multiple mobile robots,

    S. Lin, A. Liu, J. Wang, and X. Kong, “A review of path-planning approaches for multiple mobile robots,”Machines, vol. 10, no. 9, p. 773, 2022

    2022

  3. [11]

    Reciprocal velocity obstacles for real-time multi-agent navigation,

    J. Van den Berg, M. Lin, and D. Manocha, “Reciprocal velocity obstacles for real-time multi-agent navigation,” inIEEE Int. Conf. Robot. Autom., 2008, pp. 1928–1935

    2008

  4. [12]

    Efficient multi- robot motion planning for manifold-constrained manipulators by ran- domized scheduling and informed path generation,

    W. Guo, Z. Kingston, K. Hang, and L. E. Kavraki, “Efficient multi- robot motion planning for manifold-constrained manipulators by ran- domized scheduling and informed path generation,”IEEE Robot. Autom. Lett., 2026

    2026

  5. [13]

    Safe multiagent motion planning under uncertainty for drones using filtered reinforcement learning,

    S. Safaoui, A. P. Vinod, A. Chakrabarty, R. Quirynen, N. Yoshikawa, and S. D. Cairano, “Safe multiagent motion planning under uncertainty for drones using filtered reinforcement learning,”IEEE Trans. Robot., vol. 40, pp. 2529–2542, 2024

    2024

  6. [14]

    Reachability-based trajectory safeguard (RTS): A safe and fast reinforcement learning safety layer for continuous control,

    Y . S. Shao, C. Chen, S. Kousik, and R. Vasudevan, “Reachability-based trajectory safeguard (RTS): A safe and fast reinforcement learning safety layer for continuous control,”IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 3663–3670, 2021

    2021

  7. [15]

    A toolbox of level set methods,

    I. M. Mitchellet al., “A toolbox of level set methods,”UBC Depart- ment of Computer Science Technical Report TR-2007-11, vol. 1, p. 6, 2007

    2007

  8. [16]

    Provably safe and scalable multivehicle trajectory planning,

    S. Bansal, M. Chen, K. Tanabe, and C. J. Tomlin, “Provably safe and scalable multivehicle trajectory planning,”IEEE Trans. on Control Syst. Technology, vol. 29, no. 6, pp. 2473–2489, 2021

    2021

  9. [17]

    Isaacs: Iterative soft adversarial actor-critic for safety,

    K.-C. Hsu, D. P. Nguyen, and J. F. Fisac, “Isaacs: Iterative soft adversarial actor-critic for safety,” inLearning for Dynamics and Control Conf.PMLR, 2023, pp. 90–103

    2023

  10. [18]

    Iterative reachability estimation for safe reinforcement learning,

    M. Ganai, Z. Gong, C. Yu, S. Herbert, and S. Gao, “Iterative reachability estimation for safe reinforcement learning,”Advances in Neural Information Processing Syst., vol. 36, 2024

    2024

  11. [19]

    On infusing reachability-based safety assurance within planning frameworks for human–robot vehicle interactions,

    K. Leung, E. Schmerling, M. Zhang, M. Chen, J. Talbot, J. C. Gerdes, and M. Pavone, “On infusing reachability-based safety assurance within planning frameworks for human–robot vehicle interactions,” Int. J. of Robotics Research, vol. 39, no. 10-11, pp. 1326–1345, 2020

    2020

  12. [20]

    Safe multi-agent interaction through robust control barrier functions with learned uncertainties,

    R. Cheng, M. J. Khojasteh, A. D. Ames, and J. W. Burdick, “Safe multi-agent interaction through robust control barrier functions with learned uncertainties,” inIEEE Conf. on Decision and Control, 2020, pp. 777–783

    2020

  13. [21]

    Parameterized fast and safe tracking (fastrack) using deepreach,

    H. J. Jeong, Z. Gong, S. Bansal, and S. Herbert, “Parameterized fast and safe tracking (fastrack) using deepreach,” inAnnu. Learning for Dynamics & Control Conf.PMLR, 2024, pp. 1006–1017

    2024

  14. [22]

    Reachability-based trajectory design with neural implicit safety constraints

    J. Michaux, Q. Chen, Y . Kwon, and R. Vasudevan, “Reachability-based trajectory design with neural implicit safety constraints.” Robotics: Science and Syst., 2023

    2023

  15. [23]

    Efficient motion planning for manipulators with control barrier function-induced neural controller,

    M. Yu, C. Yu, M.-M. Naddaf-Sh, D. Upadhyay, S. Gao, and C. Fan, “Efficient motion planning for manipulators with control barrier function-induced neural controller,” inIEEE Int. Conf. Robot. Autom., 2024, pp. 14 348–14 355

    2024

  16. [24]

    Reach-avoid problems with time-varying dynamics, targets and constraints,

    J. F. Fisac, M. Chen, C. J. Tomlin, and S. S. Sastry, “Reach-avoid problems with time-varying dynamics, targets and constraints,” inInt. Conf. on Hybrid Syst.: Computation and Control, 2015, pp. 11–20

    2015

  17. [25]

    I. M. Mitchell,Application of level set methods to control and reachability problems in continuous and hybrid systems. Stanford University, 2002

    2002

  18. [26]

    A time-dependent Hamilton-Jacobi formulation of reachable sets for continuous dynamic games,

    I. M. Mitchell, A. M. Bayen, and C. J. Tomlin, “A time-dependent Hamilton-Jacobi formulation of reachable sets for continuous dynamic games,”IEEE Trans. on Automatic Control, vol. 50, no. 7, pp. 947– 957, 2005

    2005

  19. [27]

    Exact imposition of safety boundary conditions in neural reachable tubes,

    A. Singh, Z. Feng, and S. Bansal, “Exact imposition of safety boundary conditions in neural reachable tubes,” inIEEE Int. Conf. Robot. Autom., 2025, pp. 5489–5495

    2025

  20. [28]

    On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,

    A. W ¨achter and L. T. Biegler, “On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,”Mathematical programming, vol. 106, pp. 25–57, 2006

    2006

  21. [29]

    Safe Planning for Articulated Robots Using Reachability-based Obstacle Avoidance With Spheres,

    J. Michaux, A. Li, Q. Chen, C. Chen, and R. Vasudevan, “Safe Planning for Articulated Robots Using Reachability-based Obstacle Avoidance With Spheres,” inRobotics: Science and Syst., Delft, Netherlands, July 2024

    2024

  22. [30]

    hj reachability,

    E. Schmerling, “hj reachability,” 2024

    2024

  23. [31]

    PyTorch Kinematics,

    S. Zhong, T. Power, A. Gupta, and P. Mitrano, “PyTorch Kinematics,” Feb. 2024

    2024

  24. [32]

    Reducing the barrier to entry of complex robotic software: a MoveIt! case study,

    D. Coleman, I. Sucan, S. Chitta, and N. Correll, “Reducing the barrier to entry of complex robotic software: a MoveIt! case study,”arXiv preprint arXiv:1404.3785, 2014

    2014 arXiv

  25. [33]

    Certifiable reachability learning using a new Lipschitz continuous value function,

    J. Li, D. Lee, J. Lee, K. S. Dong, S. Sojoudi, and C. Tomlin, “Certifiable reachability learning using a new Lipschitz continuous value function,”IEEE Robot. Autom. Lett., vol. 10, no. 4, pp. 3582– 3589, 2025

    2025

  26. [34]

    Generating formal safety assurances for high- dimensional reachability,

    A. Lin and S. Bansal, “Generating formal safety assurances for high- dimensional reachability,” inIEEE Int. Conf. Robot. Autom., 2023, pp. 10 525–10 531

    2023

  27. [35]

    Multi-robot local motion planning using dynamic optimization fabrics,

    S. Bakker, L. Knoedler, M. Spahn, W. B ¨ohmer, and J. Alonso- Mora, “Multi-robot local motion planning using dynamic optimization fabrics,” in2023 International Symposium on Multi-Robot and Multi- Agent Systems (MRS). IEEE, 2023, pp. 149–155

    2023

  28. [36]

    A robust controlled backward reach tube with (almost) analytic solution for two dubins cars

    I. Mitchell, “A robust controlled backward reach tube with (almost) analytic solution for two dubins cars.” inARCH, 2020, pp. 242–258

    2020

  29. [37]

    Neural joint space implicit signed distance functions for reactive robot manipulator control,

    M. Koptev, N. Figueroa, and A. Billard, “Neural joint space implicit signed distance functions for reactive robot manipulator control,”IEEE Robotics and Automation Letters, vol. 8, no. 2, pp. 480–487, 2023. APPENDIXI DETAILEDSYSTEMSETUP This section presents the details of the...

    2023

Pith tools

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