acceptodds
Desk rejected at ICLR 2027

PhysSense: Benchmarking Sensor-Grounded Physical Safety Reasoning in Embodied Agents

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied agents increasingly rely on vision-language and vision-language-action models to reason about and interact with the physical world, yet many safety-critical states cannot be reliably inferred from visual observations alone. Properties such as temperature, contact force, payload weight, electrical state, gas concentration, obstacle clearance, and robot internal conditions often require dedicated physical sensing, creating a fundamental partial-observability challenge for embodied safety. We introduce **PhysSense**, a standards-grounded benchmark for evaluating whether embodied models can identify safety-relevant physical variables, recognize when critical evidence is missing, and make sensor-grounded safety decisions. The benchmark contains 580 instances derived from 200 semantic task template across ten operational physical-risk categories, and formulates safety reasoning as a three-way decision problem: Safe, Unsafe, or Ask, where Ask explicitly represents the need for additional physical sensing rather than generic uncertainty. To further study efficient physical-safety reasoning, we propose **PhysReason**, a training-free Big-Brain/Small-Brain framework in which a stronger model externalizes reusable physical-safety knowledge into an explicit RuleBank offline, while lightweight models retrieve and apply relevant rules during online inference. Across eight closed- and open-source models, PhysReason improves average decision accuracy from 67.74% to 85.19% and increases recall on missing-information cases from 83.96% to 97.36%. Our results reveal substantial limitations in current models' ability to reason about unobserved physical states and suggest that explicit sensor-grounded knowledge can improve embodied safety reasoning without requiring parameter updates or repeated online use of a large foundation model. The source code is available at https://anonymous.4open.science/r/PhysSense-FB77

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Embodied AI,Safety Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_h6caOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

PhysSense introduces a benchmark for sensor-grounded safety decisions in embodied-agent scenarios. It contains 580 synthetic instances from 200 templates across ten categories, with the action labels Safe, Unsafe, and Ask. The inputs are textualized observations and sensor readings rather than raw visual perception or physical robot execution. The paper also proposes PhysReason, a training-free system that routes sensor fields to an explicit rule bank and uses an LLM to interpret the evidence and produce a decision. Across eight models, PhysReason raises mean accuracy from 67.74% to 85.19% and Ask recall from 83.96% to 97.36%, while some robot-operational-state measures regress substantially.

Critical strengths
  • The benchmark frames safety as a three-way decision, including abstention, rather than forcing every case into safe/unsafe classification. Counterbalanced safe/unsafe cases are a useful way to probe whether a model attends to sensor values.
  • The paper reports heterogeneous model outcomes and multiple error analyses, exposing that aggregate gains do not guarantee improvement on every safety-relevant subtask. The operational-state regression is important evidence for readers to consider.
  • PhysReason makes its rule bank and sensor-routing logic explicit, which is a practical and inspectable alternative to relying only on a long free-form prompt.
Critical weaknesses
  • The benchmark’s Ask label is structurally leaked. All 180 Ask instances have an empty sensor dictionary, while all Safe/Unsafe instances provide the required readings (Sec. 3.2.2; Appendix A.1–A.2). Thus, an empty-dictionary check identifies Ask without reasoning about which measurement is absent or whether remaining evidence suffices. There are no partial-missingness, conflicting-sensor, or sensor-present-but-uninformative cases. High Ask recall and the triplet score cannot establish robust evidence-sufficiency reasoning under this design.
  • The reference labels rely on hidden, template-specific numerical thresholds whose provenance and independent validation are not established clearly enough. The paper states that standards motivate hazard coverage but do not prescribe every numerical boundary (Appendix A.5–A.6); the release uses representative low/high values rather than dense sweeps over sensor ranges (Appendix A.2). Since evaluated models never see the thresholds, a disagreement may reflect a different reasonable operating limit, a point the authors themselves acknowledge. Without documented derivation and independent safety-expert validation, it is unclear what the measured error means beyond reproducing these particular annotations.
  • The main Direct-versus-PhysReason experiment is confounded and under-replicated. Direct is asked for a short JSON answer, while PhysReason gets a staged evidence-audit prompt and a larger output cap: 200 versus 700 tokens for default endpoints and 1,200 versus 2,400 for Claude. The instructions, response schema, and parsing path also differ (Sec. 4.1; Appendix B.3–B.6). The authors explicitly characterize the results as comparisons of full configurations, not an isolated RuleBank test. Every model-method pair has one run over 580 examples, with no across-run confidence intervals, and actual latency/token use is not measured. The +17.46-point mean gain therefore cannot be attributed to the RuleBank or judged for run-to-run stability.
  • PhysReason is a bundle of a 27-rule knowledge artifact, a router, a selected core-rule projection, and a longer reasoning prompt, but no ablation isolates these components or compares against a token-matched stronger direct prompt. The failure on robot operational state is especially revealing: the category contains motor temperature, battery temperature, and tilt, but the routed method selects one motion/tilt rule whose listed critical states omit the thermal signals (Table 9). Seven of eight models remain at or below 62.50% accuracy in this category; whether joint thermal-and-stability retrieval fixes the failure is untested (Sec. 5.2; Appendix C.1).
  • Aggregate accuracy obscures safety-critical regressions, and the evaluation scores only final labels—not the correctness of a requested sensor, intermediate evidence audit, or actual execution. Although mean hazardous approval falls from 12.75% to 8.44%, it rises for GPT-5.4, Claude Opus 5, and Qwen 3.5 35B; Qwen 3.5 35B increases from 18.50% to 22.50% (Table 2). Robot-state accuracy falls by 26.56 points, and all-variant correctness on the 20 robot-state pairs falls from 70.63% to 17.50% (Tables 3–4). The evidence does not support a general safety-improvement claim.
Questions for authors
  1. Can you compare Direct and PhysReason with identical prompts, token caps, response schemas, and parsers, then ablate the RuleBank, routing, rule projection, and evidence-audit steps separately? Please report actual token consumption and latency.
  2. Can you provide repeated runs or another justified uncertainty analysis for every model-method pair, and report uncertainty on hazardous approvals and safe rejections rather than only aggregate accuracy?
  3. How were the per-template thresholds derived and independently validated by robotics/safety experts? Can you test continuous or held-out sensor values, and jointly route the thermal and stability evidence in the robot-state category?

I am willing to discuss with the authors.

Reviewer_p4asOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces PhysSense, a synthetic text dataset of 580 instances (200 hand-designed templates, ten risk categories, 14 fields) for three-way classification of a scenario as Safe / Unsafe / Ask. Each instance is a text prompt containing a task, a textualized visual observation, an optional operator message, and named numeric sensor fields; the label is the three-way class. Matched variants keep the task and visual text fixed and change or remove sensor values. The paper also proposes PhysReason, a training-free retrieval-augmented prompting scheme: a stronger model writes 27 rules offline (ten core rules, one per benchmark category), a deterministic keyword router selects a rule from sensor-field names, and an online model applies it with an evidence-audit prompt. Validation is zero-shot over eight models on all 580 instances, one run per configuration, comparing PhysReason against a single Direct baseline.

Critical strengths
  • The matched-variant construction is the right idea in principle. Holding task text and visual description fixed while intervening only on sensor evidence is a sensible way to isolate evidence-sensitivity, and it enables the all-variant consistency metric. This is the paper's most defensible design choice.
  • Disconfirming results are reported rather than buried. The paper states that PhysReason raises hazardous approvals for three models and drops robot-operational-state accuracy from 85.00% to 58.44% (§5.2), and that invalid outputs increase. This candor is valuable.
Critical weaknesses
  • The benchmark does not test its own central construct, and the text-only setup does not support the embodied framing. Classification is a legitimate task — the problem is not that the benchmark asks for a label, but that its inputs and metrics do not measure the capabilities it claims. Every input is text: the "visual observation" is textualized and sensor readings are synthetic numbers, so there is no perception, action, execution, or environment, and the paper's "embodied agents" and "sensor-grounded" framing is not supported. The two claimed capabilities are supplied by the input rather than measured: (a) the safety-relevant variable is named in the field itself (liquid temperature c, grip force n), so "identifying relevant physical variables" is given, not tested; (b) Ask instances "remove all required measurements," and PhysReason's prompt explicitly instructs a missing-evidence audit (§3.3) that Direct does not receive, so Ask largely reduces to a schema/missingness check. The paper's own metrics evaluate only the final label and explicitly do not evaluate "sensor-request content" (§4.2), so the headline notion of Ask as a request for additional sensing is never operationalized. The benchmark therefore cannot distinguish physical-safety reasoning from instruction-following over a fixed schema. This is not merely a matter of insufficient runs or unmatched budgets: even a budget-matched, multi-seed version of the same task would not measure the claimed capability, because the task never requires selecting the relevant variable or deciding what evidence to request.
  • There is no valid comparison supporting the method claim, and a leakage risk is left unaddressed. PhysReason differs from Direct in output-token budget (700 vs. 200), reasoning instructions, output format, and parser; the paper runs once per configuration, reports no confidence intervals, and provides no ablation (rules-only, audit-only, matched-prompt, matched-budget). §4.1 concedes the comparison "measures the full configurations rather than isolating the contribution of the RuleBank." The reported gains are therefore unattributable by construction. Compounding this, the RuleBank has exactly one core rule per benchmark category, aligned with the benchmark's own field names (Table 9, §B.2), and the paper does not document whether category definitions or annotation thresholds were visible when the rules were generated. This is a serious potential-leakage concern that the authors only partially acknowledge when they state the design does not isolate the RuleBank.
  • Scope and generalization are narrower than the framing, with no held-out evaluation. The dataset is a 580-instance, 200-template, author-authored set with author-defined numerical thresholds and no external annotation or human validation of labels. There is no template-level or category-level held-out split and no unseen-category evaluation, so nothing supports generalization beyond the templates that define the categories. Related work on selective prediction/abstention (Ask is an abstention class) and on prompt-sensitive benchmark evaluation is not engaged.
Questions for authors
  • Can you provide a matched-budget, matched-prompt comparison: Direct with the same output-token limit, output format, parser, and evidence-audit instruction, with and without the RuleBank? Without this, what evidence supports attributing any gain to rule guidance?
  • Were the benchmark categories, field names, or annotation thresholds visible to the model that generated the RuleBank? If so, how do you rule out that the ten core rules encode the label-construction logic? Please release the generation prompts.
  • Since sensor-field names map directly to category-specific rules, how does the benchmark test identification of relevant variables rather than schema-conditioned lookup? Do you have results with anonymized field names or on unseen categories?
  • Ask cases remove all required measurements. What is Direct's Ask recall when given the same missingness-audit instruction without rules? This isolates the audit prompt from the RuleBank.
  • Why is there no template- or category-level held-out split, and what evidence is there that results transfer beyond the 200 author templates?
  • Given one run at temperature 0 on hosted models, what is the run-to-run variance, and can you provide repeated runs or confidence intervals for the headline deltas?
Reviewer_kdUNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces PhysSense, a sensor-grounded physical safety reasoning benchmark for embodied agents, designed to evaluate their ability to make safety decisions under incomplete physical evidence. The benchmark consists of 580 synthetic instances derived from 200 task templates, covering ten categories of physical risks, including temperature, contact force, payload, and electrical state. By keeping task semantics and visual descriptions fixed while varying sensor readings or removing critical measurements, the benchmark formulates safety reasoning as a three-way decision problem: SAFE, UNSAFE, and ASK. Furthermore, the authors propose PhysReason, a training-free framework in which a Big Brain constructs an offline safety RuleBank, while a Small Brain retrieves and applies relevant rules during inference. Experiments across eight language models show that PhysReason improves average accuracy from 67.74% to 85.19% and ASK Recall from 83.96% to 97.36%. However, hazardous approval rates increase for some models, and performance declines significantly in the robot operational state category.

Critical strengths
  1. The paper extends embodied safety evaluation beyond conventional hazard recognition to explicitly assess the sufficiency of physical evidence for safety-critical decisions. It distinguishes among confirmed safety, confirmed danger, and insufficient evidence. In particular, the construction of matched counterfactual instances, where task semantics and visual descriptions remain unchanged while sensor evidence varies, provides a meaningful way to examine whether models adjust their safety decisions based on physical measurements rather than superficial semantic cues.

  2. PhysReason integrates offline safety knowledge construction, rule retrieval, and online evidence auditing, enabling models to incorporate explicit physical safety knowledge without parameter updates. The paper evaluates eight language models of different sizes and architectures and reports multiple metrics, including overall accuracy, hazardous approval rate, ASK Recall, and matched-set consistency. These evaluations provide a relatively comprehensive assessment of the models' safety decision-making behavior.

Critical weaknesses
  1. An obvious dataset shortcut in the ASK category undermines the validity of evidence-sufficiency evaluation. According to Section 3.2.2, all 180 ASK instances have empty sensor dictionaries, whereas all SAFE and UNSAFE instances contain relevant sensor readings. Consequently, a model could identify ASK cases simply by checking whether the sensor dictionary is empty, without understanding which physical measurements are actually required for the task. Therefore, the substantial improvement in ASK Recall does not necessarily demonstrate that the model has acquired the ability to identify missing safety-critical evidence. The authors are encouraged to introduce more challenging scenarios involving partially missing sensor information, irrelevant sensor readings, and multiple interacting physical variables to better evaluate task-dependent evidence-sufficiency reasoning.

  2. Insufficient experimental visualization and case analysis limit the clarity and persuasiveness of the results. The paper contains only one figure, Figure 1 on page 2, which presents the overall framework, while the remaining experimental results are primarily reported in tables. There are no dedicated experimental analysis figures, performance comparison plots, or representative visual case studies. Although Appendix D provides a limited number of text-based error analyses, it does not clearly illustrate how model decisions change under different sensor conditions or how successful and unsuccessful rule retrieval affects safety reasoning. For a benchmark emphasizing physical evidence variation and counterfactual safety reasoning, the lack of representative matched-scenario visualizations makes it difficult to intuitively understand the dataset construction and evaluation mechanisms. The authors are encouraged to include concrete SAFE/UNSAFE/ASK matched examples, performance comparison plots across risk categories, and representative success and failure cases of RuleBank retrieval and evidence auditing to improve the interpretability and presentation quality of the paper.

Questions for authors

Since all ASK instances lack sensor readings, while SAFE and UNSAFE instances contain relevant measurements, models may distinguish ASK cases simply by detecting empty sensor inputs. Could the authors provide additional experiments involving irrelevant sensor readings, partially missing critical evidence, and combinations of multiple sensors to demonstrate that the models can genuinely identify the physical information required for a given task?

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

Related papers

Discussion (0)

Sign in to comment.