Pith. sign in

REVIEW 4 major objections 5 minor 257 references

Language-conditioned safety filters can reduce constraint violations and transfer partially to unseen constraint instances.

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

2026-08-04 00:43 UTC pith:WREBLMOW

load-bearing objection Honest and well-scoped, but the VL filter still leans on a privileged failure function and every comparison rests on single-seed 50-episode runs. the 4 major comments →

arxiv 2608.00315 v1 pith:WREBLMOW submitted 2026-07-31 cs.RO cs.LGcs.SYeess.SY

Towards General Language-Conditioned Latent Safety Filters

Ihab Tabbara , Yuxuan Yang , Hussein Sibai This is my paper
classification cs.RO cs.LGcs.SYeess.SY MSC 68T4093C85
keywords language-conditioned safety filterHamilton-Jacobi reachabilitylatent safety filtervision-language-action modelconstraint generalizationrobot manipulationout-of-distribution generalizationvision-language model failure detection
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 show that safety filters for robots can be made general in the same way robot task policies are becoming general: instead of training a separate filter for each safety rule, a single filter reads the rule as natural-language text. Its proposal is a Hamilton-Jacobi reachability critic and actor that take as input a visual observation plus an embedded language constraint, and decide whether the nominal policy's proposed action should be replaced by a safety-preserving one. Evaluated on three simulated manipulation tasks (pick-and-place, table wiping, block stacking), the language-conditioned filter lowers collision and order-violation rates relative to the unfiltered policy, performs comparably to separately trained per-constraint filters, and shows partial generalization to objects, colors, and shapes not seen during filter training. The paper also reports that current vision-language models, when directly asked to judge whether a state violates a language constraint, are not yet reliable enough to serve as the failure signal; in the vision-language experiments the failure function still comes from simulator state. If the central claim holds, safety filtering can move from specification-specific monitors toward pretrained, text-conditioned safety components that are reused and fine-tuned across deployments.

Core claim

The central claim is that a single Hamilton-Jacobi safety critic, conditioned on a latent representation that jointly encodes visual observations and a natural-language safety constraint, can enforce multiple constraints, match the performance of one-filter-per-constraint systems, and extend partly to constraint instances unseen during training. The critic estimates the worst-case safety value of a candidate action under the language-specified failure set; when the nominal action's value falls below a margin, the filter substitutes the closest sampled action whose value is safe. In the paper's three manipulation tasks this design reduced collision rates and improved order correctness relativ

What carries the argument

The load-bearing object is the language-conditioned Hamilton-Jacobi safety critic, a value function over actions, robot observation, and an embedded constraint, trained with the fixed-point Bellman backup approximated by a discounted Q-learning objective. Its zero sublevel set approximates the backward reachable set—states from which violation is unavoidable—so the same critic represents a different failure set whenever the text changes. A learned actor approximates the safety-maximizing action, and at runtime the filter intervenes only when the critic's value for the nominal action is below a margin, choosing the safe candidate closest to the nominal action. The phrase that carries the arch

Load-bearing premise

The load-bearing premise is that every language constraint has a correct failure function computed from the simulator's privileged state and the known constraint—in the vision-language experiments this function still comes from ground truth, not from the image and text; if a user offers only open-ended language with no such oracle, the paper's evidence for generality does not cover that case.

What would settle it

Run the three tasks with a constraint outside the training families, e.g., 'keep the grasped object above the table' or 'do not tilt the milk box', supplying only camera views and text with no privileged failure function. If a filter trained with a vision-language-model-labeled failure function does not drive violation rates below the nominal policy, or if performance collapses on reworded unseen constraints, the generality claim fails. An immediate check within the paper's setup: on out-of-distribution scenes, repeat the obstacle-versus-benign two-rollout protocol and test whether the filter

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

Share X LinkedIn Reddit HN

If this is right

  • Deployment-time safety rules can be changed by changing text alone; a single trained filter covers the constraint family instead of requiring retraining per rule.
  • The same critic can represent multiple simultaneously active constraints and sequentially ordered constraints, with results comparable to or better than composing independent per-constraint filters, notably higher order-correctness in block stacking.
  • Pretrained vision-language representations supply enough object semantics for the critic to partly transfer to unseen objects, colors, and shapes within a constraint family.
  • Direct vision-language-model failure judgment is currently the bottleneck: the best model tested catches only about 56% of failures, so practical filters still require a separate source for the failure function.
  • The sampling-based safe-action selection costs about 220 ms per step, over four times the 50 ms control budget; executing the safety actor directly fits the budget and gives comparable safety metrics.

Where Pith is reading between the lines

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

  • Inference: the architecture would be tested at its limit by replacing the privileged failure function with one learned from vision-language-model single-frame labels; if that closed the gap, the filter could be trained from pixels and text alone, which the paper does not yet claim.
  • Inference: the cue-disentangling results (color used when shape name is absent, shape used when color is ambiguous) suggest the model uses object attributes rather than a single shortcut; a follow-up with adversarial attribute correlations could reveal whether this is robust or driven by spurious features.
  • Inference: the same Bellman backup supports richer failure sets than geometric distances; composing critics for temporal-logic or multi-stage specifications is a natural extension that the paper gestures toward but does not demonstrate.
  • Inference: the honest scope of the claim is constraints of the form 'avoid named object' and 'grasp next named block'; extending to open-ended user constraints with no privileged failure function would require solving the vision-language-model reliability problem the paper documents.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a language-conditioned Hamilton-Jacobi safety filter: a single learned safety critic and actor are conditioned on a vision-language embedding of the observation and a natural-language safety constraint, and are used to intervene on a nominal policy. The approach is evaluated in three RoboSuite manipulation environments (Safe Grab, Safe Wipe, Stack Blocks-3/4) under both a privileged ground-truth state representation and a vision-language (VL) representation built on the pi0.5 VLM backbone. The main empirical claims are that the general filter reduces constraint violations relative to the unfiltered nominal policy, that it is competitive with per-constraint filters, that it partially transfers to unseen objects/colors/shapes, and that current VLMs are not yet reliable as failure functions. The paper also contributes a 5,000-trajectory safety-critical manipulation dataset and VLM evaluation protocols.

Significance. If the empirical claims are statistically secure, the paper makes a useful step toward safety filters that can be reused across multiple language-specified constraints, an important direction as VLA policies become more general. The work's strengths include a clearly formulated extension of latent HJ reachability to language-conditioned inputs, a comparison against per-constraint filters, explicit OOD evaluation, a large released dataset, and a careful latency analysis. The discussion of VLM failure-function limitations is honest and appropriately cautious. However, the central comparisons — General vs. Single and OOD transfer — are not yet statistically supported, and the VL experiments rely on a privileged failure function rather than language-derived supervision. These issues need to be addressed before the paper's claims can be considered established.

major comments (4)
  1. [Table 1 and Table 2; Appendix A.9] All reported metrics are from a single training seed (seed 7) and 50 evaluation scenes per setting. With n=50, the binomial standard error is up to 7 percentage points, so differences such as Safe Grab VL CR 46% (General) vs. 56% (Single), Stack Blocks-4 VL OC 64% vs. 50%, and OOD shifts (Safe Wipe CR 38%→56%; Stack Blocks-3 OC 82%→72% in Geom-Color) are within or near one standard error. No confidence intervals, significance tests, or multiple seeds are reported. The large safety gains over the unfiltered nominal policy are robust, but the novel claims of General-vs-Single competitiveness and partial OOD transfer are load-bearing and currently not statistically secured. Please report multiple seeds and confidence intervals, or at least a significance test on the pooled episodes.
  2. [Experimental Setup, VL failure function (Appendix A.1)] In the VL setting, the failure function used to train the critic is h(z_t)=h(s_t,c), computed from the simulator's privileged state s_t and the known constraint c, not from the language/image input. This means the language condition enters only through the VLM representation that encodes the observation and text; the safety supervision itself is not derived from language or vision. The paper therefore does not evaluate the scenario in which the constraint text is the only specification and the failure set must be inferred from language and images. This is a limitation on the claimed generality of 'language-conditioned' safety filtering, and should be stated more prominently and tested, or the claims should be softened.
  3. [Results, 'The General filter respects the specified constraint'] The two-rollout protocol measures contact with the benign object versus the designated obstacle, but these numbers are also from 50 episodes per condition and no error bars. For example, Safe Grab benign-contact 74% vs. obstacle-contact 26% and Safe Wipe 56% vs. 38% are presented as supporting the claim that the filter is not merely avoiding everything, but the statistical separation is not established. Please provide confidence intervals or additional episodes for this protocol.
  4. [Table 3 and VLM evaluation] The VLM evaluation correctly shows that the best model (Gemini Robotics-ER 1.6) detects only 56.4% of failures while having 95.2% non-failure accuracy, and the paper wisely concludes that raw VLM predictions are not yet usable as a runtime failure oracle. This is an important negative result, but it also reinforces the point that the main experiments' safety signal comes from privileged ground-truth h(s_t,c), not from language understanding. The paper should clarify this relationship in the main text rather than only in the appendix.
minor comments (5)
  1. [References] Several references are duplicated (Ames et al. 2017a and 2017b; Bansal et al. 2017a and 2017b). Please consolidate.
  2. [Appendix A.1, Stack Blocks failure function] The failure set is defined as {s_t | h(s_t,c) ≤ 0} in the appendix, while the main text defines the zero-sublevel set as {s | V(s) < 0} and h(s)<0. Please make the boundary convention consistent.
  3. [Table 2] The SR and CR columns for Safe Grab and Safe Wipe OOD come from two separate rollouts per scene (one driving toward the obstacle, one toward the benign object). This is stated in the text but could be clarified in the table caption to avoid misinterpretation.
  4. [Appendix A.9] Hyperparameter tuning is described qualitatively ('selected by evaluating the resulting filters... and inspecting recorded rollout videos'). Please provide the actual values tried and the selection criterion for the final configuration, especially for d_thresh, r, and epsilon, since these directly affect the safety behavior.
  5. [Appendix A.7] The mean-pooling vs. attention-pooling ablation shows mixed results, and the attention variant increases memory to about 740 GB per replay buffer. This is useful context, but the main text should reference this trade-off when discussing the VL performance gap.

Circularity Check

0 steps flagged

No circularity: central claims rest on held-out evaluation; only a minor, non-load-bearing self-citation.

full rationale

After walking the derivation chain, I find no circular step. The central object is the HJ critic trained with the fixed-point Bellman loss (Eq. 1) using targets built from the privileged failure function h; nothing in Eq. 1 or the runtime filter (Eq. 3 and the sampling procedure in A.4) is defined in terms of the reported CR/SR/OC metrics, so the results are not true by construction. The VL experiments set h(z_t)=h(s_t,c) because current VLMs are too unreliable as failure oracles; that is an external-supervision assumption and a stated limitation, not a circular reduction. The critic still has to learn a mapping from VLM embeddings to the safety value, and on OOD scenes it is evaluated on objects, colors, and shapes that were neither used to set constants nor seen during filter training, so Table 2 is a genuine transfer measurement. The single-seed, 50-scene evaluation makes some reported differences statistically fragile, but that is a sampling limitation, not circularity. The one self-citation (Tabbara et al. 2025) appears in related work and as a pointer to previous latent-filter/classifier ideas; it is not load-bearing, and no uniqueness theorem or ansatz is imported from it to force the present contribution. Hyperparameter selection by inspecting success/collision rates is model selection, not a fitted parameter being renamed as a prediction. Score 2 reflects only the minor, non-load-bearing self-citation; the central derivation is self-contained against its own benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard HJ reachability formalism, a privileged failure-function oracle even in the VL setting, and the pretraining coverage of the frozen VLM. No new physical entities are postulated. The main free parameters are the failure margins, the intervention threshold, and the hyperparameters selected by rollout inspection.

free parameters (5)
  • Failure margin d_thresh (Safe Grab, Safe Wipe) = 0.02
    Minimum allowed distance to the obstacle; chosen by hand and defines the failure set that the critic is trained and evaluated against (Appendix A.1, Table 7).
  • Stack Blocks slackness margin r = 0.0495 (1.3 × block half-width)
    Defines h(s_t,c)=r-d_cube; hand-set slack for the 'grasp the target block' constraint (Table 7).
  • Intervention threshold epsilon = not reported
    Runtime margin in the filter's switching rule Q(z_t,a_nom) >= epsilon; value not given in text, yet it directly controls intervention rate and the safety/task trade-off (Methodology, Runtime HJ reachability-based safety filtering).
  • Discount schedule and SAC hyperparameters = gamma 0.5→0.99, cost scale 0.1, etc.
    Selected by evaluating success/collision rates and inspecting rollout videos (Appendix A.9); these choices affect the learned Q and therefore the filter's decisions.
  • Number of candidate actions sampled = 100 actor + 100 noise-perturbed nominal actions
    Design choice in the minimum-deviation safe-action selection (Appendix A.4); affects task performance and safety.
axioms (5)
  • standard math The Bellman fixed-point equation V(s)=min{h(s), max_a V(f(s,a))} characterizes the optimal safety value function for the HJ reachability problem.
    Used as the foundation for the learned HJ critic (Preliminaries, Bellman equation).
  • domain assumption The environment is a known, fully observable Markov decision process in simulation, with a fixed task-achieving nominal policy.
    All experiments run in ROBOSUITE with scripted nominal controllers; the filter is evaluated in this closed-world setting (Experimental Setup; Appendix A.2).
  • domain assumption A correct constraint-conditioned failure function h(s_t,c) is available from simulator state and the known constraint for training the critic in both GT and VL settings.
    In the VL setting the paper explicitly sets h(z_t)=h(s_t,c) because VLMs are not reliable; without this privileged h the language-conditioned training signal would not exist (Experimental Setup, Tasks).
  • domain assumption The frozen VLM encoder pi0.5 produces embeddings z_t that preserve enough visual-language-proprioceptive structure for the critic to regress the privileged h; OOD objects and colors are assumed to fall inside the VLM's pretraining distribution.
    The paper states that OOD scenes 'likely fall within the pretraining distribution of the VLM backbone,' which is the mechanism proposed for transfer (Results, The General filter transfers to unseen constraints).
  • domain assumption In Stack Blocks, stage completion is observable to the environment ('We assume stage completion is observable').
    The filter switches constraints per stage; this oracle is assumed rather than learned (Appendix A.1).

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards General Language-Conditioned Latent Safety Filters." pith.science (2026). https://pith.science/paper/WREBLMOW

@misc{pith2026260800315,
  author       = {Pith},
  title        = {Pith review of: Towards General Language-Conditioned Latent Safety Filters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WREBLMOW}},
  note         = {Machine review of arXiv:2608.00315}
}
Share X LinkedIn Reddit HN
read the original abstract

Robot policies are becoming increasingly general, with vision-language-action (VLA) models enabling a single policy to execute diverse tasks specified in natural language. Safe deployment, however, requires adapting not only to new tasks but also to varying safety requirements across users, environments, and applications. Existing safety filters remain largely constraint-specific and thus must be redesigned or relearned when safety requirements change. In this paper, we investigate language-conditioned safety filtering, in which a Hamilton-Jacobi safety actor and critic are conditioned on language-specified constraints. We evaluate this formulation across pick-and-place, table-wiping, and block-stacking tasks in the vision-based setting, examining its ability to enforce language-specified constraints and transfer to unseen constraint instances within the evaluated constraint families. Our experiments provide evidence that language-conditioned safety filters reduce constraint violations and exhibit partial transfer to unseen constraint instances.

Figures

Figures reproduced from arXiv: 2608.00315 by Hussein Sibai, Ihab Tabbara, Yuxuan Yang.

Figure 1
Figure 1. Figure 1: Overview of the three benchmark environments. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual examples of the out-of-distribution settings. (a) and (b) show the object-substitution settings for Safe Grab and [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A failure episode in Safe Grab. The manipulator and the designated obstacle overlap in the agent view, so contact [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A failure episode in Safe Wipe. The arm jitters in front of the designated obstacles and makes no progress on the [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A failure episode in Stack Blocks-3. The gripper is driven into a pose from which the target block cannot be grasped [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: A failure episode in Stack Blocks-4. With three blocks already placed, the arm collides with the existing stack while [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Examples from the safety-critical manipulation trajectory dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sample vision input and output response of single [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sample vision input and output response of pair [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sample vision input and output response of short [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: A nominal VLA (π0.5) may fail to modify its behavior when an explicit safety constraint is appended to the task prompt. In both examples, the policy disregards the constraint and acts on the visual observation alone [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

257 extracted references · 4 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    Proceedings of The 7th Conference on Robot Learning , pages =

    RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control , author =. Proceedings of The 7th Conference on Robot Learning , pages =. 2023 , editor =

    2023

  13. [13]

    arXiv preprint arXiv:2406.09246 , year=

    OpenVLA: An Open-Source Vision-Language-Action Model , author=. arXiv preprint arXiv:2406.09246 , year=

    Pith/arXiv arXiv

  14. [14]

    Proceedings of Robotics: Science and Systems , address =

    Octo: An Open-Source Generalist Robot Policy , author =. Proceedings of Robotics: Science and Systems , address =

  15. [15]

    arXiv preprint arXiv:2502.19645 , year=

    Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success , author=. arXiv preprint arXiv:2502.19645 , year=

    Pith/arXiv arXiv

  16. [16]

    2024 , eprint=

    _0 : A Vision-Language-Action Flow Model for General Robot Control , author=. 2024 , eprint=

    2024

  17. [17]

    Black, Kevin and Brown, Noah and Darpinian, James and Dhabalia, Karan and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Galliker, Manuel Y. and Ghosh, Dibya and Groom, Lachy and Hausman, Karol and ichter, brian and Jakubczak, Szymon and Jones, Tim and Ke, Liyiming and LeBlanc, Devin and Levine, Sergey an...

    2025

  18. [18]

    2025 , eprint=

    FAST: Efficient Action Tokenization for Vision-Language-Action Models , author=. 2025 , eprint=

    2025

  19. [19]

    2025 , eprint=

    MolmoAct: Action Reasoning Models that can Reason in Space , author=. 2025 , eprint=

    2025

  20. [20]

    2026 , eprint=

    Steerable Vision-Language-Action Policies for Embodied Reasoning and Hierarchical Control , author=. 2026 , eprint=

    2026

  21. [21]

    2026 , eprint=

    AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots , author=. 2026 , eprint=

    2026

  22. [22]

    2022 , eprint=

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances , author=. 2022 , eprint=

    2022

  23. [23]

    2025 , eprint=

    CAST: Counterfactual Labels Improve Instruction Following in Vision-Language-Action Models , author=. 2025 , eprint=

    2025

  24. [24]

    NeurIPS 2023 Datasets and Benchmarks , year=

    LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning , author=. NeurIPS 2023 Datasets and Benchmarks , year=

    2023

  25. [25]

    2025 , eprint=

    SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Safe Reinforcement Learning , author=. 2025 , eprint=

    2025

  26. [26]

    2026 , eprint=

    HazardArena: Evaluating Semantic Safety in Vision--Language--Action Models , author=. 2026 , eprint=

    2026

  27. [27]

    Proceedings of The 9th Conference on Robot Learning , pages =

    RoboMonkey: Scaling Test-Time Sampling and Verification for Vision-Language-Action Models , author =. Proceedings of The 9th Conference on Robot Learning , pages =. 2025 , editor =

    2025

  28. [28]

    2026 , eprint=

    Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment , author=. 2026 , eprint=

    2026

  29. [29]

    2026 , eprint=

    When Vision Overrides Language: Evaluating and Mitigating Counterfactual Failures in VLAs , author=. 2026 , eprint=

    2026

  30. [30]

    2019 International Conference on Robotics and Automation (ICRA) , pages=

    Bridging hamilton-jacobi safety analysis and reinforcement learning , author=. 2019 International Conference on Robotics and Automation (ICRA) , pages=. 2019 , organization=

    2019

  31. [31]

    Advances in neural information processing systems , volume=

    Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers , author=. Advances in neural information processing systems , volume=

  32. [32]

    Conference on robot learning , pages=

    Daydreamer: World models for physical robot learning , author=. Conference on robot learning , pages=. 2023 , organization=

    2023

  33. [33]

    2024 , eprint=

    Mastering Diverse Domains through World Models , author=. 2024 , eprint=

    2024

  34. [34]

    arXiv preprint arXiv:2411.04983 , year=

    Dino-wm: World models on pre-trained visual features enable zero-shot planning , author=. arXiv preprint arXiv:2411.04983 , year=

    Pith/arXiv arXiv

  35. [35]

    Advances in Neural Information Processing Systems , volume=

    Where are we in the search for an artificial visual cortex for embodied intelligence? , author=. Advances in Neural Information Processing Systems , volume=

  36. [36]

    2024 , eprint=

    Conformal Decision Theory: Safe Autonomous Decisions from Imperfect Predictions , author=. 2024 , eprint=

    2024

  37. [37]

    2023 , eprint=

    Conformal Policy Learning for Sensorimotor Control Under Distribution Shifts , author=. 2023 , eprint=

    2023

  38. [38]

    Trust-based Rate-Tunable Control Barrier Functions for Non-Cooperative Multi-Agent Systems , year=

    Parwana, Hardik and Mustafa, Aquib and Panagou, Dimitra , booktitle=. Trust-based Rate-Tunable Control Barrier Functions for Non-Cooperative Multi-Agent Systems , year=

  39. [39]

    2022 , eprint=

    Adaptive Conformal Prediction for Motion Planning among Dynamic Agents , author=. 2022 , eprint=

    2022

  40. [40]

    2022 , eprint=

    Targeted Adversarial Attacks against Neural Network Trajectory Predictors , author=. 2022 , eprint=

    2022

  41. [41]

    2022 , eprint=

    Adaptation for Validation of a Consolidated Control Barrier Function based Control Synthesis , author=. 2022 , eprint=

    2022

  42. [42]

    2021 , eprint=

    Learning Safe Multi-Agent Control with Decentralized Neural Barrier Certificates , author=. 2021 , eprint=

    2021

  43. [43]

    Wu, Haoze and Barrett, Clark and Sharif, Mahmood and Narodytska, Nina and Singh, Gagandeep , title =. Proc. ACM Program. Lang. , month =. 2022 , issue_date =. doi:10.1145/3563325 , abstract =

    work page doi:10.1145/3563325 2022

  44. [44]

    arXiv preprint arXiv:2511.07899 , year=

    Statistically Assuring Safety of Control Systems using Ensembles of Safety Filters and Conformal Prediction , author=. arXiv preprint arXiv:2511.07899 , year=

    arXiv

  45. [45]

    arXiv preprint arXiv:2505.00908 , year=

    Learning Conservative Neural Control Barrier Functions from Offline Data , author=. arXiv preprint arXiv:2505.00908 , year=

    arXiv

  46. [46]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Planning with Learned Object Importance in Large Problem Instances using Graph Neural Networks , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2021 , month=. doi:10.1609/aaai.v35i13.17421 , abstractNote=

    work page doi:10.1609/aaai.v35i13.17421 2021

  47. [47]

    Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems , pages =

    Li, Qingbiao and Gama, Fernando and Ribeiro, Alejandro and Prorok, Amanda , title =. Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems , pages =. 2020 , isbn =

    2020

  48. [48]

    6th Annual Conference on Robot Learning , year=

    Learning Multi-Object Dynamics with Compositional Neural Radiance Fields , author=. 6th Annual Conference on Robot Learning , year=

  49. [49]

    Graph Neural Networks: Adversarial Robustness

    G \"u nnemann, Stephan. Graph Neural Networks: Adversarial Robustness. Graph Neural Networks: Foundations, Frontiers, and Applications. 2022. doi:10.1007/978-981-16-6054-2_8

    work page doi:10.1007/978-981-16-6054-2_8 2022

  50. [50]

    International Conference on Learning Representations , year=

    Fundamental Limits in Formal Verification of Message-Passing Neural Networks , author=. International Conference on Learning Representations , year=

  51. [51]

    Advances in Neural Information Processing Systems , editor=

    Randomized Message-Interception Smoothing: Gray-box Certificates for Graph Neural Networks , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    2022

  52. [52]

    Message-Aware Graph Attention Networks for Large-Scale Multi-Robot Path Planning , year=

    Li, Qingbiao and Lin, Weizhe and Liu, Zhe and Prorok, Amanda , journal=. Message-Aware Graph Attention Networks for Large-Scale Multi-Robot Path Planning , year=

  53. [53]

    Learning-based Motion Planning in Dynamic Environments Using

    Ruipeng Zhang and Chenning Yu and Jingkai Chen and Chuchu Fan and Sicun Gao , booktitle=. Learning-based Motion Planning in Dynamic Environments Using. 2022 , url=

    2022

  54. [54]

    Are socially-aware trajectory prediction models really socially-aware? , journal =

    Saeed Saadatnejad and Mohammadhossein Bahari and Pedram Khorsandi and Mohammad Saneian and Seyed-Mohsen Moosavi-Dezfooli and Alexandre Alahi , keywords =. Are socially-aware trajectory prediction models really socially-aware? , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.trc.2022.103705 , url =

    arXiv 2022

  55. [55]

    arXiv preprint arXiv:2206.03669 , year=

    Toward certified robustness against real-world distribution shifts , author=. arXiv preprint arXiv:2206.03669 , year=

    Pith/arXiv arXiv

  56. [56]

    , booktitle=

    Rober, Nicholas and Everett, Michael and How, Jonathan P. , booktitle=. Backward Reachability Analysis for Neural Feedback Loops , year=

  57. [57]

    and Xu, Xiangru and Grizzle, Jessy W

    Ames, Aaron D. and Xu, Xiangru and Grizzle, Jessy W. and Tabuada, Paulo , journal=. Control Barrier Function Based Quadratic Programs for Safety Critical Systems , year=

  58. [58]

    2018 Annual American Control Conference (ACC) , year=

    Permissive Barrier Certificates for Safe Stabilization Using Sum-of-squares , author=. 2018 Annual American Control Conference (ACC) , year=

    2018

  59. [59]

    5th Annual Conference on Robot Learning , year=

    Safe Nonlinear Control Using Robust Neural Lyapunov-Barrier Functions , author=. 5th Annual Conference on Robot Learning , year=

  60. [60]

    IEEE Control Systems Letters , year=

    Robust Control Barrier Functions With Sector-Bounded Uncertainties , author=. IEEE Control Systems Letters , year=

  61. [61]

    2017 IEEE International Conference on Robotics and Automation (ICRA) , year=

    Safe certificate-based maneuvers for teams of quadrotors using differential flatness , author=. 2017 IEEE International Conference on Robotics and Automation (ICRA) , year=

    2017

  62. [62]

    , journal=

    Prajna, Stephen and Jadbabaie, Ali and Pappas, George J. , journal=. A Framework for Worst-Case and Stochastic Safety Verification Using Barrier Certificates , year=

  63. [63]

    2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year=

    Comparative Analysis of Control Barrier Functions and Artificial Potential Fields for Obstacle Avoidance , author=. 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year=

    2021

  64. [64]

    BarrierNet: Differentiable Control Barrier Functions for Learning of Safe Robot Control , year=

    Xiao, Wei and Wang, Tsun-Hsuan and Hasani, Ramin and Chahine, Makram and Amini, Alexander and Li, Xiao and Rus, Daniela , journal=. BarrierNet: Differentiable Control Barrier Functions for Learning of Safe Robot Control , year=

  65. [65]

    and Ames, Aaron D

    Mehra, Aakar and Ma, Wen-Loong and Berg, Forrest and Tabuada, Paulo and Grizzle, Jessy W. and Ames, Aaron D. , booktitle=. Adaptive cruise control: Experimental validation of advanced controllers on scale-model cars , year=

  66. [66]

    2023 , eprint=

    Verification and Synthesis of Robust Control Barrier Functions: Multilevel Polynomial Optimization and Semidefinite Relaxation , author=. 2023 , eprint=

    2023

  67. [67]

    A Distributed Simplex Architecture for Multi-Agent Systems , author=. J. Syst. Archit. , year=

  68. [68]

    ArXiv , year=

    Control Barrier Function Contracts for Vehicular Mission Planning Under Signal Temporal Logic Specifications , author=. ArXiv , year=

  69. [69]

    2018 IEEE Conference on Decision and Control (CDC) , year=

    Control of Multi-Agent Systems with Finite Time Control Barrier Certificates and Temporal Logic , author=. 2018 IEEE Conference on Decision and Control (CDC) , year=

    2018

  70. [70]

    , journal=

    Chen, Yuxiao and Singletary, Andrew and Ames, Aaron D. , journal=. Guaranteed Obstacle Avoidance for Multi-Robot Operations With Limited Actuation: A Control Barrier Function Approach , year=

  71. [71]

    , journal=

    Tan, Xiao and Dimarogonas, Dimos V. , journal=. Distributed Implementation of Control Barrier Functions for Multi-agent Systems , year=

  72. [72]

    International Conference on Learning Representations , year=

    Learning Safe Multi-agent Control with Decentralized Neural Barrier Certificates , author=. International Conference on Learning Representations , year=

  73. [73]

    Proceedings of the 23rd International Conference on Hybrid Systems: Computation and Control , articleno =

    Jagtap, Pushpak and Swikir, Abdalla and Zamani, Majid , title =. Proceedings of the 23rd International Conference on Hybrid Systems: Computation and Control , articleno =. 2020 , isbn =. doi:10.1145/3365365.3382196 , abstract =

    arXiv 2020

  74. [74]

    , journal=

    Lindemann, Lars and Dimarogonas, Dimos V. , journal=. Control Barrier Functions for Multi-Agent Systems Under Conflicting Local Signal Temporal Logic Tasks , year=

  75. [75]

    ArXiv , year=

    Learning Environment-Aware Control Barrier Functions for Safe and Feasible Multi-Robot Navigation , author=. ArXiv , year=

  76. [76]

    2023 , eprint=

    Efficient LQR-CBF-RRT*: Safe and Optimal Motion Planning , author=. 2023 , eprint=

    2023

  77. [77]

    Implicit Graph Neural Networks , volume =

    Gu, Fangda and Chang, Heng and Zhu, Wenwu and Sojoudi, Somayeh and El Ghaoui, Laurent , booktitle =. Implicit Graph Neural Networks , volume =

  78. [78]

    IEEE Robotics and Automation Letters , volume=

    Learning safe, generalizable perception-based hybrid control with certificates , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  79. [79]

    Enforcing safety for vision-based controllers via Control Barrier Functions and Neural Radiance Fields , year=

    Tong, Mukun and Dawson, Charles and Fan, Chuchu , booktitle=. Enforcing safety for vision-based controllers via Control Barrier Functions and Neural Radiance Fields , year=

  80. [80]

    2022 , eprint=

    Control-oriented meta-learning , author=. 2022 , eprint=

    2022

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.