Pith. sign in

REVIEW 3 major objections 4 minor 30 references

A control-barrier-function correction injected into flow-matching denoising provably keeps generated robot action chunks safe while keeping the output distribution within a bounded 2-Wasserstein distance of the unmodified model.

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

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

T0 review · deepseek-v4-flash

2026-08-03 04:23 UTC pith:EFDU36HL

load-bearing objection A useful modular safety filter for flow-matching VLAs with solid hardware results, but the formal safety guarantee rests on unverified assumptions about the base flow and the initial noise. the 3 major comments →

arxiv 2607.29569 v1 pith:EFDU36HL submitted 2026-07-31 cs.RO cs.SYeess.SY

Safe Vision Language Action Models via Barrier Enhanced Flow Matching

Kasra Sinaei , Hung-Chieh Wu , Donald Ebeigbe This is my paper
classification cs.RO cs.SYeess.SY
keywords Flow matchingControl barrier functionsVision-language-action modelsSafe trajectory generationLog-sum-exp barrier2-Wasserstein distanceInference-time safetyQuadratic programming
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 prove that a pretrained vision-language-action (VLA) robot policy can be made safe at inference time without retraining or safety-labeled data. Its proposal is to inject a CBF quadratic-program correction into the flow-matching denoising process itself, not to filter the final action. The central claim is that this correction enforces the barrier condition h(q_i) ≥ 0 across the whole action chunk while keeping the generated distribution close to the original model's distribution in 2-Wasserstein distance. If the proof holds, any flow-matching-based VLA can be wrapped with formal safety guarantees and still complete its tasks. The experiments on two robotic manipulation platforms and a 2D navigation benchmark support that claim.

Core claim

The central discovery is that safety can be enforced during denoising rather than after generation. The paper defines an action-chunk barrier as a smooth log-sum-exponential aggregate h(z) = -1/κ ln Σ e^{-κ h(q_i)} and solves a minimal-norm quadratic program at each denoising step whose solution δ_s is added to the flow velocity. Theorem 4 states that if the unmodified flow's worst-case barrier decay rate M is known and denoising starts from a safe noise sample with h0 > 0, then activating the filter from time τ_s ≤ h0/M guarantees h(ψ(z,1)) ≥ (h0 - Mτ_s)e^{-α(1-τ_s)} > 0, so every action in the chunk is safe. Corollary 1 shows the modified flow still has finite L2 error and Lipschitz consta

What carries the argument

The load-bearing object is the smooth log-sum-exponential barrier h(z) = -1/κ ln Σ_i e^{-κ h(q_i)}, which under-approximates the min-barrier over an action chunk with error bounded by ln H/κ. Combined with the CBF-QP (12)/(19), its gradient appears in the constraint ∂h/∂z (v_θ + δ) ≥ -α h, and the closed-form correction δ_s = ReLU(-α h - ∇h·v_θ) ∇h/||∇h||² projects the denoising velocity into the safe half-space. This projection is what makes the safe set forward invariant during late denoising, while the aggregate barrier keeps the QP single-constraint and real-time.

Load-bearing premise

The guarantee depends on a known worst-case barrier decay rate M and an initial noise sample that starts inside the safe set with h0 > 0; the paper gives no procedure to measure M, and its own sampling scheme does not guarantee h0 > 0.

What would settle it

Run Algorithm 1 with a wall or spherical barrier on any flow-matching VLA, drawing many z0 from the paper's Beta sampler, and record min_i h(q_i) at the end of each rollout; if even one rollout with τ_s ≤ h0/M and h0 > 0 ends with a negative barrier value, the theorem's guarantee fails. A targeted alternative is to measure the empirical worst-case barrier decay M of the unmodified VLA and check whether trials satisfying τ_s ≤ h0/M always remain safe.

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

If this is right

  • A pretrained flow-matching VLA can be made safe simply by wrapping its denoising loop; no safety-specific dataset or retraining is required.
  • The correction's bounded Wasserstein shift means safety is not bought by pushing outputs out-of-distribution, so task success rate need not drop.
  • Because the whole action chunk is protected by one smooth barrier constraint, the added computation is a single small QP per denoising step, compatible with real-time control.
  • The same QP can include velocity-box constraints and a smoothness penalty through a sparse difference matrix, reducing jerky motions.
  • Safety becomes an attractor during late denoising: even if the trajectory is heading unsafe, the filter exponentially steers the generated chunk back toward the safe set.

Where Pith is reading between the lines

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

  • Editorial inference: The formal guarantee covers the generated action chunk, not the physical robot; an exponentially stable tracking controller is assumed, so a testable extension would close the loop with tracking-error-aware barriers.
  • Editorial inference: The condition h0 > 0 is not ensured by the paper's Beta-distribution noise sampler; one could sample from a distribution conditioned on the safe set or activate the filter from τ = 0 to make the guarantee unconditional.
  • Editorial inference: Since M is a uniform worst-case decay rate, estimating it per-step online would allow earlier and adaptive filter activation, reducing distributional shift and conservatism.
  • Editorial inference: Because the correction acts on the vector field rather than on a specific model architecture, the same construction likely transfers to other flow-based generative planners beyond VLAs, provided the barrier gradient is well-defined.

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

3 major / 4 minor

Summary. The paper proposes a modular safety filter for Vision-Language-Action (VLA) models that use flow matching. Instead of filtering the final output, the method adds a Control Barrier Function (CBF) based quadratic-program correction term δ_s to the denoising ODE, enforcing safety on the whole action chunk via a log-sum-exp aggregate barrier. The paper proves a lower bound on the barrier value at the end of denoising (Theorem 4) and a bounded 2-Wasserstein distance between the modified and target distributions (Corollary 1). Experiments on two robotic manipulation platforms (SO-101, QArm) and a 2D maze benchmark report 100% safety rates without degrading success rates, with lower computational cost and smoother trajectories than SafeDiffuser and SafeFlowMatcher.

Significance. If the theoretical guarantees were established for the implemented algorithm, this would be a practically valuable contribution: a retraining-free, plug-and-play safety layer for state-of-the-art flow-matching VLA policies. The empirical results are encouraging and the method's computational efficiency is a clear plus. The paper also correctly identifies a limitation of post-hoc filtering and provides a concrete alternative that addresses it. The main weakness is that the formal safety and distribution-shift guarantees rest on unverified sufficient conditions, so the 'provable safety' claim is currently stronger than what is demonstrated.

major comments (3)
  1. [§III-C, Theorem 4 and Algorithm 1] Theorem 4's sufficient conditions are not connected to the implemented algorithm. The theorem requires a finite worst-case barrier decay rate M for the unmodified flow (Eq. 16) and a safe initial noise sample h0>0. Algorithm 1 samples z0 from Beta(1.5,1) with no conditioning or rejection; for the wall/sphere barriers of Section III-D (Eqs. 20-23), h(z0) can be negative. The paper provides no procedure to estimate M, and no argument that it is finite for the trained π0 vector field. Thus the formal safety guarantee does not apply to the algorithm as written. Please add a mechanism to ensure h0>0 (e.g., rejection sampling or initial projection), discuss how M can be bounded or estimated, or explicitly restate Theorem 4 as conditional and soften the 'provable safety' claims in the abstract and conclusions.
  2. [§III-B, Corollary 1] The bounded-Wasserstein claim rests on unverified regularity assumptions. The proof sketch asserts that the scalar activation before the ReLU is bounded and Lipschitz because h is twice continuously differentiable and the robot operates in a compact, singularity-free workspace. However, during denoising the intermediate states zτ are not constrained to the singularity-free workspace; they can pass through configurations where the Jacobian loses rank, so ||∇h|| may approach zero and δ_s may blow up. Moreover, the Lipschitz constant of the modified vector field depends on the neural network vθ, whose global Lipschitz constant is not established. A rigorous bound on ||δ_s|| and its Lipschitz constant is needed, or these must be stated as assumptions that are empirically verified in the experiments.
  3. [§III-C and Algorithm 1] Theorem 4 proves safety for the continuous-time ODE (11), but Algorithm 1 solves the QP at discrete Euler steps. The CBF inequality is enforced only at the sample points; between steps the constraint can be violated when h is close to zero. The paper does not analyze discretization error or relate Δτ to the Lipschitz constants. Please state that the guarantee is for the continuous-time flow and either provide a discretization error bound or adopt a more conservative margin in the discrete implementation.
minor comments (4)
  1. [§III-C, Remark 2] The statement that when h0−Mτs<0 the filter 'exponentially drives the trajectory toward the safe set boundary' is potentially misleading: the CBF inequality only gives an asymptotic approach, not a guarantee of reaching h≥0 by τ=1. Consider clarifying that this does not constitute a finite-time safety certificate.
  2. [Table I] The units and meaning of the 'κ' and 'Accel.' columns are unclear. Is κ the trajectory curvature or the smoothing parameter? Please clarify in the table caption or text.
  3. [Eq. (13)] The closed-form expression for δ_s is correct, but it would help to explicitly note that it is the KKT solution of (12) only when the constraint is active; the ReLU operator handles the inactive case. This is implied but not stated.
  4. [Algorithm 1, line 5] The condition 'if τ ≥ τs' is evaluated on the discretized loop; if τs does not coincide with a multiple of Δτ, the filter starts at a slightly different time. A brief note on how τs is rounded would improve reproducibility.

Circularity Check

0 steps flagged

No circularity found: the safety and Wasserstein guarantees are conditional proofs from externally cited results, not self-referential derivations.

full rationale

The paper's central derivation—Theorem 4—is a conditional forward-invariance proof: under the stated assumptions on M and h0, the QP constraint in (12) enforces the CBF inequality dh/dτ ≥ -αh, which by comparison yields h(ψ(z,1)) ≥ (h0 - Mτs)e^{-α(1-τs)}. This is a direct consequence of the designed constraint, not an equivalence with an input. The Wasserstein bound (Corollary 1) applies the external result of Benton et al. [27] to the modified vector field, with bounds expressed in terms of the perturbation norm; no fitted parameter is renamed as a prediction. The LSE barrier approximation (Theorem 2) is cited to an external proof [25]. The authors' prior works [12], [24] are cited for background and baseline construction (model-free CBF, adaptive CBF), but the safe-flow guarantee and the distribution-shift analysis do not rely on those citations. The sufficient conditions of Theorem 4 (finite M, h0 > 0) are assumptions rather than demonstrated facts for the implemented Beta-sampled z0, which is a correctness/validation gap, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 8 axioms · 0 invented entities

The central method rests on standard CBF and flow-matching results plus several application-specific assumptions. The most fragile are the unmeasured worst-case decay rate M, the requirement h0 > 0, and the bounded-away-from-zero barrier gradient. These are not fitted from data but are assumed, and the paper provides no procedure to validate them.

free parameters (7)
  • alpha (CBF decay rate) = not reported in experiments
    Used in QP (12)/(19) and Theorem 4; chosen by user; controls how aggressively the barrier is enforced.
  • kappa (LSE smoothing) = not reported; ablation uses 1, 5, 50, 250
    Smoothing parameter in Eq. (8); larger κ makes LSE closer to min; affects safety margin.
  • tau_s (filter start time) = not reported; ablation uses 0.6-0.9
    Denoising time at which the filter is switched on; Theorem 4 requires τ_s ≤ h0/M.
  • lambda (smoothing weight) = not reported
    Weights the second quadratic term in QP (19); affects smoothness vs minimal correction.
  • M (worst-case barrier decay rate) = not estimated
    Assumed finite in Theorem 4; no method is given to compute or bound it from the flow model.
  • barrier parameters (wall normal/offset or sphere center/radius) = not reported
    Define the user-specified safe set; needed to evaluate h and ∇h in Algorithm 1.
  • velocity limits v_min/v_max = not reported (e.g., ±0.1 rad/s in Fig. 3)
    Box constraints in QP (19); chosen by user for smoothness.
axioms (8)
  • standard math Log-sum-exp under-approximates the min of barrier values (Theorem 2 from [25])
    Used to define h(z) and to conclude h(z) > 0 implies all qi safe.
  • standard math CBF forward invariance theorem (Ames et al. [11])
    Basis for the QP constraint and Theorem 4's second phase.
  • standard math Benton et al. flow-matching Wasserstein bound (Eq. 15)
    Imported as Theorem 3; the paper's distribution-shift claim is an application.
  • standard math KKT conditions for the convex QP (12)
    Gives closed-form δ_s in Eq. (13).
  • domain assumption Control-affine dynamics (2) with Lipschitz f,g and unique solutions
    Standard nonlinear control background used in Definitions and Theorem 1.
  • domain assumption Barrier gradient ∇_z h is bounded away from zero while the filter is active; compact configuration space away from singularities
    Used in Corollary 1 to keep δ_s bounded and the modified flow Lipschitz; not verified experimentally except by setup.
  • ad hoc to paper The unmodified denoising flow has a finite worst-case barrier decay rate M and the initial noise sample satisfies h0 > 0 with τ_s ≤ h0/M
    Theorem 4's sufficient condition; no estimation procedure is provided.
  • domain assumption Position-controlled robot with exponentially stable tracking controller
    Justifies treating generated action chunks as safe reference trajectories; not formalized in the integrated guarantee.

pith-pipeline@v1.3.0-daily-deepseek · 11843 in / 18080 out tokens · 185156 ms · 2026-08-03T04:23:14.714123+00:00 · methodology

0 comments
read the original abstract

This article presents a modular inference framework that integrates Flow Matching generative models with formal Control Barrier Function (CBF) safety guarantees. Unlike existing methods that apply external safety filters to a model's final output, our approach modifies the Flow Matching denoising process within the model to inherently generate safe trajectories. By employing a smooth Log-Sum-Exponential aggregate barrier, we enforce safety over entire action chunks. This aggregate barrier ensures a minimal increase in computational overhead and does not alter the semantic intent of the model. We show that, within the proposed framework, the 2-Wasserstein distance between the generated distribution and the target distribution remains bounded. Our method eliminates the need for safety-specific datasets or costly model retraining, providing a versatile solution for safe inference. We validate the approach on two robotic manipulation platforms and a 2D navigation benchmark, verifying that our framework achieves reliable safety without degrading the success rate of the model.

Figures

Figures reproduced from arXiv: 2607.29569 by Donald Ebeigbe, Hung-Chieh Wu, Kasra Sinaei.

Figure 1
Figure 1. Figure 1: High-level block diagram of a safe VLA with an additional CBF [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation study demonstrating the impact of hyper-parameters on the safety margin. Trajectories were recorded during a physical manipulation task using the SO-101 robot. Provided that the sampled noise is safe (h0 > 0) and denois￾ing starts such that τs ≤ h0 M then (12) enforces safe flows. Proof. Let us assume the worst-case scenario for the first part of the flow matching, where the CBF filter is not acti… view at source ↗
Figure 3
Figure 3. Figure 3: Effective velocity constraint enforcement via CBF QP within one [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sample trajectories generated by different methods for Maze 2D [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rollout of safe VLA with CBF-Enhanced flow matching in an [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Rollout of safe VLA with wall barriers performing table cleanup task. Barrier-enhanced FM does not allow the gripper to enter unsafe region. TABLE II HARDWARE MANIPULATION RESULTS ACROSS TWO ROBOTIC PLATFORMS. Platform Metric No Filter E2E-CBF CBF-FM (Ours) SO-101 (5-DoF) Pick-and-Place Safety (%) 15.0 68.2 100.0 Success (%) 75.0 68.2 77.4 QArm (4-DoF) Table Cleanup Safety (%) 0.00 70.0 100.0 Success (%) 1… 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

30 extracted references · 9 linked inside Pith

  1. [1]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  2. [2]

    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

  3. [3]

    Robot learning: A tutorial,

    F. Capuano, C. Pascal, A. Zouitine, T. Wolf, and M. Aractingi, “Robot learning: A tutorial,”arXiv preprint arXiv:2510.12403, 2025

    arXiv 2025

  4. [4]

    Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,”Advances in Neural Information Processing Systems, vol. 36, pp. 44 776–44 791, 2023

    2023

  5. [5]

    π 0.5: a vision-language-action model with open-world generalization,

    P. Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusaiet al., “π 0.5: a vision-language-action model with open-world generalization,”arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,”arXiv preprint arXiv:2304.13705, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    2023

  8. [8]

    Open- vla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketiet al., “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Smolvla: A vision-language-action model for affordable and efficient robotics,

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafiotiet al., “Smolvla: A vision-language-action model for affordable and efficient robotics,”arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,”arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Control barrier function based quadratic programs for safety critical systems,

    A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function based quadratic programs for safety critical systems,”IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3861–3876, 2016

    2016

  12. [12]

    Safe adaptive control with vanishing conservativeness for robotic systems with unknown dynamics via barrier functions,

    K. Sinaei and D. Ebeigbe, “Safe adaptive control with vanishing conservativeness for robotic systems with unknown dynamics via barrier functions,”IEEE Robotics and Automation Letters, vol. 11, no. 5, pp. 6161–6168, 2026

    2026

  13. [13]

    Robust data-driven receding horizon control1,

    J. Zheng, S. Kiani, M. Sznaier, and C. Lagoa, “Robust data-driven receding horizon control1,”IFAC-PapersOnLine, vol. 59, no. 16, pp. 25–30, 2025

    2025

  14. [14]

    Cobl-diffusion: Diffusion-based conditional robot planning in dynamic environments using control barrier and lyapunov functions,

    K. Mizuta and K. Leung, “Cobl-diffusion: Diffusion-based conditional robot planning in dynamic environments using control barrier and lyapunov functions,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 13 801– 13 808

    2024

  15. [15]

    Safediffuser: Safe planning with diffusion probabilistic models,

    W. Xiao, T.-H. Wang, C. Gan, R. Hasani, M. Lechner, and D. Rus, “Safediffuser: Safe planning with diffusion probabilistic models,” in The thirteenth international conference on learning representations, 2023

    2023

  16. [16]

    Safeflowmatcher: Safe and fast planning using flow matching with control barrier functions,

    J. Yang, S. Jang, and S. Han, “Safeflowmatcher: Safe and fast planning using flow matching with control barrier functions,”arXiv preprint arXiv:2509.24243, 2025

    arXiv 2025

  17. [17]

    Vlsa: Vision- language-action models with plug-and-play safety constraint layer,

    S. Hu, Z. Liu, S. Liu, J. Cen, Z. Meng, and X. He, “Vlsa: Vision- language-action models with plug-and-play safety constraint layer,” arXiv preprint arXiv:2512.11891, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Safevla: Towards safety alignment of vision-language-action model via constrained learning,

    B. Zhang, Y . Zhang, J. Ji, Y . Lei, J. Dai, Y . Chen, and Y . Yang, “Safevla: Towards safety alignment of vision-language-action model via constrained learning,”arXiv preprint arXiv:2503.03480, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    From demonstrations to safe deployment: Path-consistent safety filtering for diffusion policies,

    R. R ¨omer, J. Balletshofer, J. Thumm, M. Pavone, A. P. Schoel- lig, and M. Althoff, “From demonstrations to safe deployment: Path-consistent safety filtering for diffusion policies,”arXiv preprint arXiv:2511.06385, 2025

    arXiv 2025

  20. [20]

    Safeflow: Safe robot motion planning with flow matching via control barrier functions,

    X. Dai, Z. Yang, D. Yu, F. Liu, H. Sadeghian, S. Haddadin, and S. Hirche, “Safeflow: Safe robot motion planning with flow matching via control barrier functions,”arXiv preprint arXiv:2504.08661, 2025

    arXiv 2025

  21. [21]

    Flow matching on general geometries,

    R. T. Chen and Y . Lipman, “Flow matching on general geometries,” arXiv preprint arXiv:2302.03660, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    H. K. Khalil,Nonlinear systems. Upper Saddle River, N.J.: Prentice Hall, 2002

    2002

  23. [23]

    Model-free safety-critical control for robotic systems,

    T. G. Molnar, R. K. Cosner, A. W. Singletary, W. Ubellacker, and A. D. Ames, “Model-free safety-critical control for robotic systems,”IEEE robotics and automation letters, vol. 7, no. 2, pp. 944–951, 2021

    2021

  24. [24]

    Safety-critical position con- trol of robots: A model-free approach,

    K. Sinaei, H.-C. Wu, and D. Ebeigbe, “Safety-critical position con- trol of robots: A model-free approach,” in2025 American Control Conference (ACC). IEEE, 2025, pp. 2689–2694

    2025

  25. [25]

    Composing control barrier functions for complex safety specifications,

    T. G. Molnar and A. D. Ames, “Composing control barrier functions for complex safety specifications,”IEEE Control Systems Letters, vol. 7, pp. 3615–3620, 2023

    2023

  26. [26]

    Boyd and L

    S. Boyd and L. Vandenberghe,Convex optimization. Cambridge University Press, 2004

    2004

  27. [27]

    Error bounds for flow matching methods,

    J. Benton, G. Deligiannidis, and A. Doucet, “Error bounds for flow matching methods,”arXiv preprint arXiv:2305.16860, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Control barrier functions for multi-agent systems under conflicting local signal temporal logic tasks,

    L. Lindemann and D. V . Dimarogonas, “Control barrier functions for multi-agent systems under conflicting local signal temporal logic tasks,”IEEE control systems letters, vol. 3, no. 3, pp. 757–762, 2019

    2019

  29. [29]

    Obstacle avoidance for low-speed autonomous vehicles with barrier function,

    Y . Chen, H. Peng, and J. Grizzle, “Obstacle avoidance for low-speed autonomous vehicles with barrier function,”IEEE Transactions on Control Systems Technology, vol. 26, no. 1, pp. 194–206, 2017

    2017

  30. [30]

    Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,

    R. Cadene, S. Alibert, A. Soare, Q. Gallouedec, A. Zouitine, S. Palma, P. Kooijmans, M. Aractingi, M. Shukor, D. Aubakirova, M. Russi, F. Capuano, C. Pascal, J. Choghari, J. Moss, and T. Wolf, “Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,” https://github.com/huggingface/lerobot, 2024

    2024