acceptodds
Withdrawn from ICLR 2027

PhysCogSafe: Diagnosing Physical Safety Cognition in Generalist Robot Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generalist robot policies, particularly Vision-Language-Action (VLA) models, are emerging as a central paradigm for robotic manipulation, making their safe operation in the physical world a critical concern. Existing benchmarks primarily characterize whether executions are safe, but provide limited evidence about whether observed safety reflects risk-sensitive physical cognition or incidental success. To address this gap, we introduce PhysCogSafe, a controlled diagnostic framework for safety-relevant physical cognition in robotic manipulation. Rather than treating safety as a binary endpoint, PhysCogSafe diagnoses whether a task-capable policy selectively adapts its behavior to physical risk along three dimensions: geometry-conditioned reasoning, property-conditioned reasoning, and multi-subgoal compositional reasoning. Each diagnostic probe holds the instruction and task objective fixed while forming a matched family comprising a benign baseline, a physical-risk condition, a control that retains comparable scene changes without the causal hazard, and an independently verified safe reference. Counterfactual contrasts and trajectory-level monitors jointly assess task completion and targeted harm. We instantiate PhysCogSafe in LIBERO, evaluating four representative VLA policies and one world-action model across 11 scenario families. Across 50 of 54 policy–scenario pairs, introducing a physical hazard increases targeted harm relative to an otherwise comparable scene in which the hazard is absent, supporting the intended risk-specific comparison. After controlling for underlying task capability, we observe that successful, harm-free completion falls from 69.5% without the hazard to 11.7% with it, while 66.2% of hazardous-scene outcomes involve harm. Failure modes also depend on the required physical reasoning: policies often complete tasks requiring property-conditioned reasoning unsafely, whereas tasks requiring multi-subgoal compositional reasoning more often end in both task failure and harm. Overall, none of the five evaluated policies reliably translates task competence into risk-aware physical behavior. These findings highlight safety-relevant physical cognition as a capability that should be explicitly addressed in the development and evaluation of generalist robot policies.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generalist Robot Policies, Cognition Diagnosing, Vision Language Action Model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_crRXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PhysCogSafe diagnoses risk-conditioned behavior using matched benign, hazardous, and control scenes plus an independently verified safe reference. It covers geometry, object properties, and multi-subgoal constraints across five robot policies and 11 scenario families. Among baseline-successful trials, macro-averaged safe completion falls from 69.5% in controls to 11.7% under risk, with distinct failure profiles across dimensions.

Critical strengths
  1. Matched controls, unchanged goals, and verified safe references address important confounds in robot safety evaluation and make the diagnostic framework useful.
  2. Capability conditioning and the four joint outcomes distinguish unsafe task completion from harmful failure; explicit exclusions and cell-level denominators improve transparency.
Critical weaknesses
  1. Diagnostic attribution: Success on the native task establishes ordinary manipulation competence, not competence at the safe alternative required by a probe. A safe-reference controller demonstrates feasibility but not that the evaluated policy possesses that skill. The results identify failed adaptation, without fully distinguishing risk cognition from perception or control limitations.
  2. Construction dependence: Trajectory-conditioned probes are selected so the policy's original actions become harmful under risk. This is a valid stress test, but its difficulty depends on the policy and chosen trajectory. Even with the stated within-policy interpretation, pooled failure rates need stability checks using separate construction trajectories and frozen/shared probes to establish how much is specific to this selection.
  3. Statistical evidence: Several cells have only two baseline-successful trials (Tables 10–12), yet receive full weight in cell-level macro-averages. Paired confidence intervals and sensitivity to sparse cells are missing. The large overall gap is plausible, but exact headline values and dimension comparisons are less secure than their point estimates suggest.
  4. Dimension interpretation: D1, D2, and D3 contain only four, four, and three scenario families, with different tasks and physical interventions. Larger deficits in D2/D3 may partly reflect probe difficulty or additional motor/sequencing demands. Matched-difficulty or factorial probes are needed before treating these differences as evidence about the relative cognitive demands of the dimensions.
  5. Incremental contribution: Existing robot-safety evaluations already establish that task success does not imply safe execution. The matched-control protocol is a useful refinement, but the paper does not demonstrate that its taxonomy reliably separates failure causes or yields diagnostic insights unavailable from simpler outcome-based evaluation. This limits the significance of the claimed cognitive diagnosis.
Questions for authors
  1. Can a risk-aware comparison policy with the same observations/action interface complete the safe alternatives, and do explicit risk cues separate recognition failures from execution limitations?
  2. How stable are the results on independently constructed trajectories or frozen shared probes, and what are paired intervals and sparse-cell sensitivity results?
  3. Do the D1-versus-D2/D3 differences persist after controlling task difficulty and additional motor/sequencing requirements?
1 comment

Q1. Can a risk-aware comparison policy with the same observations/action interface complete the safe alternatives, and do explicit risk cues separate recognition failures from execution limitations?

A1. We agree with the diagnostic-attribution boundary. Native-task success shows only ordinary manipulation competence; the safe reference shows that a safe solution exists under the same robot interface, not that the evaluated policy possesses that skill. Our estimand is therefore failed risk-conditioned adaptation on a feasible probe, not a process-level separation of risk cognition, perception, or control. A same-interface risk-aware comparison policy would strengthen the positive- control story, but it is not required to interpret the present negative result once Πsafe\Pi_{\mathrm{safe}} already establishes solvability. Explicit linguistic risk cues similarly change the construct from native-prompt vision-conditioned behavior to language-conditioned safety, so they cannot cleanly separate recognition from execution under our protocol. We will state both as valuable future diagnostics and remove language that implies a fully decomposed cognitive attribution.

Q2. How stable are the results on independently constructed trajectories or frozen shared probes, and what are paired intervals and sparse-cell sensitivity results?

A2. We agree that trajectory-conditioned probes are policy- and trajectory- dependent stress tests, so pooled rates must not be read as a shared-item score. Primary claims will be within-policy / within-cell directional contrasts (Q(ER)<Q(EC)Q(E_R)<Q(E_C), H(ER)>H(EC)H(E_R)>H(E_C)), with the aggregate 69.5 ⁣→ ⁣11.769.5\!\rightarrow\!11.7 demoted to a descriptive summary of heterogeneous cells. We will add paired bootstrap CIs and sparse-cell sensitivity (nB≥5/10n_B\ge 5/10), so that cells with nB=2n_B=2 remain in the ledger but do not drive headlines. Full stability on independently reconstructed trajectories and frozen shared cross-policy probes is important; we will report what the current artifacts support and mark broader reconstruction as a limitation / future check rather than claim it is already established.

Q3. Do the D1-versus-D2/D3 differences persist after controlling task difficulty and additional motor/sequencing requirements?

A3. No: under the present design we cannot treat D1 versus D2/D3 as evidence of relative cognitive demand. The dimensions contain only 4/4/34/4/3 families with different tasks, interventions, and motor/sequencing burdens, so larger ΔQ\Delta Q may partly reflect probe difficulty rather than dimensional cognition. What the data do support is a qualitative outcome-composition contrast: D2 more often shows harmful completion (p11p^{11}), D3 more often harmful failure (p01p^{01}), while D1 is more mixed. We will remove headline rankings such as 41.7/70.5/63.441.7/70.5/63.4 as cognitive- difficulty scores, report family-level results, and state that matched- difficulty or factorial probes are future work. Correspondingly, the contribution is the matched EBE_B/ERE_R/ECE_C plus safe- reference attribution protocol, not a taxonomy that already separates failure causes beyond simpler outcome evaluation.

Reviewer_ihXyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a benchmark to evaluate if generalist robot policies adapt their behavior safely to physical hazards placed in a scene in the LIBERO benchmark. The authors create several hazard types across 10 distinct LIBERO tasks and use this to evaluate several open source VLA models: OpenVLA-OFT, π0, π0.5, GR00T N1.7, and Cosmos. Overall they find that safe completion drops from 69.5% to 11.7% when a hazard is present.

Critical strengths

The paper has a strong motivation. Enumerating and organizing different safety violation types for VLA policies is important. Furthermore building benchmarks that can reliable test policies for safety is necessary for broader deployment of VLA models.

Critical weaknesses

Experimental results are limited to simulation and evaluated on models trained using LIBERO which is known to be highly overfit [LIBERO-PRO Xueyang Zhou 2025]. Since policies trained on LIBERO typically cannot even adapt to slight changes in scene or target object position, it is not surprising that they fail to adapt to other safety constraints. Due to this I feel in order for these results to be generalizable there must be hardware experiments on non-overfit generalist VLA policies, like the DROID fine-tunes of common policies. This is my largest issue with the paper. How do we know that the models are not simply overfit and in replaying the memorized trajectories ?

Overall contribution seems to be limited. Two papers cited in this paper HazardArena (Chen et al., 2026) and RedVLA (Zhang et al., 2026b) both tackle similarly motivated problems. Both of these papers also include hardware results on Franka DROID setups.

Paper is somewhat difficult to parse and dense. Reads as though it was mostly written by AI and unmodified.

Questions for authors

How can you determine that the LIBERO policies tested are not simply overfitting to the training demonstrations ? Can you measure the trajectory similarity between the baseline and unsafe rollouts ?

1 comment

Q: How can you determine that the LIBERO policies tested are not simply overfitting to the training demonstrations? Can you measure the trajectory similarity between the baseline and unsafe rollouts?

A: We agree that LIBERO overfitting is a real confound, and that our results are simulation diagnostics on suite-trained policies, not hardware evidence of generalist safety. PhysCogSafe does not claim that these policies would transfer to DROID or the real world; it asks a narrower question: under matched EBE_B/ERE_R/ECE_C interventions, does an otherwise task-capable policy change behavior when a causal risk relation is introduced. Pure memorization of training demos does not explain the main contrast: formal ERE_R/ECE_C are closed-loop rollouts on new risk/control observations, not open-loop replay of demonstrations, and H(ER)>H(EC)H(E_R)>H(E_C) in 50/5450/54 cells with no reversal while Q(EC)≫Q(ER)Q(E_C)\gg Q(E_R). If failure were only ``any scene change breaks a memorized demo,'' the matched control should collapse similarly; it does not. Construction-time open-loop replay of frozen EBE_B actions further shows that the same nominal sequence remains task-valid without the hazard in ECE_C and activates the preregistered event in ERE_R. We will add trajectory-similarity diagnostics between EBE_B and closed-loop ERE_R/ECE_C (action MSE / EEF-path DTW), reporting both risk-insensitive near-replays and cases that diverge yet still fail safely. Hardware evaluation on non-overfit DROID fine-tunes is important future work; we will state it as a scope limitation and clarify that HazardArena/RedVLA address complementary questions (semantic suppression; risk elicitation), whereas our contribution is the matched-triplet plus safe-reference attribution protocol.

Reviewer_eTAwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PhysCogSafe, a diagnostic framework that tests whether a generalist robot policy adapts its actions to a physical hazard while it still completes the task. For each task, the framework builds three matched scenes with the same instruction and goal: a benign scene (EB), a risk scene (ER) that adds one hazard, and a control scene (EC) that adds the same scene change without the hazard. A scripted safe controller must complete the task in ER, which shows that a safe solution exists. Task completion and a scenario-specific harm event are monitored over the full trajectory. The authors organize the probes into three dimensions: geometry-conditioned reasoning (D1), property-conditioned reasoning (D2), and multi-subgoal compositional reasoning (D3). The framework is instantiated in LIBERO on 11 scenario families. Four VLA policies (OpenVLA- OFT, π0, π0.5, GR00T N1.7) and one world-action model (Cosmos) are evaluated. Of 55 planned model–scenario cells, 54 pass construction and 53 enter the capability-conditioned analysis. Harm is more frequent in ER than in EC in 50 of 54 cells. Among trials that succeed in EB, safe completion falls from 69.5% in EC to 11.7% in ER, and 66.2% of ER outcomes involve harm. Property- conditioned tasks end mostly in harmful completion; multi-subgoal tasks end mostly in harmful failure. The claimed contributions are the behavioral formulation and taxonomy, the matched protocol with safe references, and the evaluation with released tasks and tools.

Critical strengths

S1. The matched-triplet protocol isolates risk-conditioned behavior. Prior safety bench-marks report whether an execution is safe. This paper asks whether the policy changes its behavior because of the hazard. The control scene separates risk sensitivity from scene novelty. The safe reference shows that the task is feasible in the risk scene. The capability filter removes trials where the policy cannot do the task at all. Together, these three elements are a clear methodological step over HazardArena, RedVLA, and LIBERO-Safety.

S2. The failure-mode decomposition is informative and fully reported. The joint out-come split shows two distinct failure patterns: unsafe persistence in D2 and coupled task–safety failure in D3. Appendix E gives every cell-level number, the two excluded cells are stated, and code and data are released. This makes the results checkable.

Critical weaknesses

W1. The paper does not describe its own hazards, controls, or harm events. For 8 of 11 families, the reader cannot tell what the hazard is, what the matched control changes, what event the oracle flags, or what the safe reference does. Without this, I cannot judge whether the probes are valid. I have these questions. First, LIBERO VLA policies output end-effector deltas and a binary gripper command. If the D2-A harm oracle uses contact force, the policy cannot avoid harm except by not grasping, and D2-A becomes a refusal test, not an adaptation test. Second, D2-B hazards (for example, a mug in a microwave) are predicates on object type and location, not simulated physics. This overlaps with semantic-safety benchmarks and weakens physical cognition framing. W2. The evaluation design does not support the dimension-level claims. Section 5.1 specifies 50 initial states per cell, one rollout per state, and equal-weight macro-averaging, with no minimum cell size and no uncertainty estimate. Because RQ2 and RQ3 condition on baseline success, a weak policy gets a small denominator, and the protocol has no rule for this: the OpenVLA-OFT row in Figure 4a averages cells with nB = 2, 2, 5, 8. Separately, the paper builds 11 families from 9 of the 40 tasks in the four LIBERO suites and does not state the selection criterion. Each dimension then rests on 3 or 4 families (D3-B on one), yet Section 5.5 generalizes to the dimension despite large within-dimension variation (D1-1versus D1-4). The claims should be scoped to the tested families, or the selection justified.

Questions for authors

Q1. For each of the 11 families, please give a table with: the hazard, the matched control, harm event the oracle flags, the behavior of the safe reference, and whether the family uses fixed-state or trajectory-conditioned construction. For D2-A, please also state how the harm oracle works and whether the policy can change gripper force in LIBERO. If it cannot, D2-A is a test of refusal, not of adaptation.

Q2. Please add bootstrap confidence intervals and n-weighted averages for Q, H, ∆Q, and the pab rates, with a minimum cell size. Does the D2-versus-D3 difference survive? Please also state how the 11 families were selected from the 40 LIBERO tasks, and why the control is harder than the risk scene in the 9 cells with S(EC) < S(ER).

Q3. Can you separate perception failure from adaptation failure? Two tests would help: (a) add a safety cue to the instruction (for example, “the mug is glass”) and report the change in Q(ER); (b) query the VLM backbone of π0.5 or OpenVLA-OFT with the ER image and a question about the hazard. Without this, the “physical cognition” label is not supported by the data.

1 comment

Q1: For each of the 11 families, please give hazard / matched control / harm event / safe reference / construction type; for D2-A, clarify the oracle and whether gripper force is controllable.

A1: We will add a per-family appendix table covering all 11 probes (D1-1--D1-4, D2-1--D2-4, D3-1--D3-3): hazard, matched control, oracle event, safe-reference behavior, and fixed-state versus trajectory-conditioned construction. LIBERO policies cannot set gripper force: the action is 6D end-effector deltas plus open/close under fixed-gain position control. D2-A therefore does not test force modulation. Its oracle is a composite rough-handling / damage-risk criterion (approach speed, force proxies, release drop, post-release dynamics, and categorical failures), and never contacting the target is not counted as harm. Thus D2-A measures handling-style adaptation, with refusal reported separately. For D2-B, D2-3 is a material--microwave compatibility / abstention probe, while D2-4 is a physical active-stove route-adaptation probe; we will not pool them as identical adaptation tests.

Q2: Please add bootstrap CIs, nn-weighted averages, and a minimum cell size; does D2 versus D3 survive; how were the 11 families selected; why is S(EC)<S(ER)S(E_C)<S(E_R) in 9 cells?

A2: We agree and will report bootstrap 95% CIs, both unweighted and nBn_B-weighted averages, and sensitivity under minimum cell sizes (e.g., nB≥5/10n_B\ge 5/10). Sparse cells such as OpenVLA-OFT with nB∈{2,5,8}n_B\in\{2,5,8\} will remain in the ledger but will not drive headlines. We expect the qualitative gap of D2/D3 versus D1, and the p11p^{11} versus p01p^{01} composition contrast, to remain; the precise D2--D3 ordering (70.570.5 versus 63.463.4) will not be claimed as stable. Families were selected constructively for taxonomy coverage, native prompt/goal/asset preservation, and constructible EBE_B/ERE_R/ECE_C with a verified safe reference---not by random sampling of the 40 LIBERO tasks. In the 9 cells with S(EC)<S(ER)S(E_C)<S(E_R), ECE_C retains a matched nuisance burden that can lower task success without activating the targeted harm; event attribution is unaffected, since H(ER)>H(EC)H(E_R)>H(E_C) in 50/5450/54 cells with no reversal.

Q3: Can you separate perception failure from adaptation failure via (a) an instruction cue and (b) VLM hazard questioning?

A3: We agree perception and adaptation are distinct, and our data do not fully localize the failure stage; we will restate the claim as failed risk-conditioned behavioral adaptation, not process-level physical cognition. The present EBE_B/ERE_R/ECE_C design already fixes the native prompt and paired state, varying only the preregistered risk cue, so the contrast is behavioral under matched conditions. Proposal (a) changes the construct to language-conditioned safety and exits the native-instruction distribution. Proposal (b) probes a VQA path that is not the formal action-generating interface. Neither is therefore a necessary condition for the current results; we will mark full perception--adaptation decomposition as a limitation and future diagnostic, not as a requirement for the reported behavioral findings.

Comments on the paper

1 comment

We thank all reviewers for their careful and constructive feedback. The comments on hazard/oracle specification, statistical aggregation, diagnostic attribution, construction dependence, and the scope of our claims have been highly useful; we will revise the paper accordingly to make the evaluation contracts explicit, strengthen uncertainty reporting, and more carefully bound the contribution to matched risk-conditioned behavioral diagnosis.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.