acceptodds
Under review as a conference paper at ICLR 2027

EmbodiedSafe: Benchmarking Context-Aware Runtime Hazard Handling in Embodied Agents

Abstract

Embodied agents driven by vision-language models (VLMs) integrate visual perception and language understanding to plan and execute multi-step tasks in interactive environments. However, deploying embodied agents in household tasks raises safety concerns, as unsafe behavior may cause property damage or physical injury. Existing benchmarks expose such behavior, but hazardous-execution rates alone do not capture appropriate responses across task contexts and stages. These responses depend on the current context, available corrective actions, and the timing of safety requirements. We present EmbodiedSafe, a benchmark for evaluating context-aware runtime hazard handling by embodied agents with and without runtime guardrails. EmbodiedSafe covers four evaluation levels: Hazard Rejection, Context Sensitivity, Hazard Resolution, and Hazard Prospection, assessing refusal, context-dependent decisions, corrective execution, and temporal safety. It contains 1,480 household evaluation samples covering 77 fine-grained hazard mechanisms and 20 temporal safety requirements. We evaluate five runtime guardrails across three embodied-agent workflows, with supplementary comparisons spanning four VLM backbones. Although guardrails reduce hazardous execution, failures persist across all four levels: the best context-pair accuracy remains below 58%, hazard-resolution success remains below 1%, and safe completion averages 0.22% under implicit temporal safety requirements.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Embodied Agents, Vision-Language Models, AI Safety Benchmark

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_2YisOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes EMBODIEDSAFE, a new benchmark evaluating context-aware runtime hazard handling by embodied agents with and without runtime guardrails. The benchmark covers four evaluation levels: Hazard Rejection, Context Sensitivity, Hazard Resolution, and Hazard Prospection, assessing refusal, context-dependent decisions, corrective execution, and temporal safety. The evaluation shows that three embodied-agent workflows fall short on five runtime guardrails.

Critical strengths
  • The safety risks considered in the paper are grounded in public safety guidelines and equipment manuals.

  • Empirical evaluation considers a broad suite of runtime guardrails.

Critical weaknesses
  • The intellectual merit of the work is limited. Compared with existing embodied benchmarks, the proposed benchmark does not have clear advantages in task settings. The evaluation metrics seem broader, but the underlying properties of the agents being evaluated are unclear.

  • The paper studies the safety awareness of embodied agents with and without runtime guardrails. However, mixing the agent's own safety awareness with the guardrail's power reduces the study's scientific value. Especially, some of these guardrails are deployed independently of the agent. If the purpose of testing these guardrails is to show their overly high false rejection rate, then the study should be clearly split into two parts: safety of embodied agents and capabilities of runtime guardrails.

Questions for authors

None

3 comments

(W1) Advantages in task settings and evaluated underlying properties

(1) Novelty. EmbodiedSafe goes beyond evaluating whether agents recognize or avoid hazards. Its key contribution is a unified evaluation of context-aware runtime hazard handling, progressively assessing whether agents refuse hazardous requests, distinguish safe from unsafe contexts, resolve existing hazards during task execution, and satisfy implicit safety requirements before their deadlines. These capabilities are evaluated through controlled task settings grounded in 77 household hazard mechanisms and 20 temporal safety requirements, as compared with prior benchmarks in Table 5.

(2) Evaluated capabilities and metrics. Section 3.3 explicitly defines each metric and the capability it measures. Hazard Rejection uses ESR and ARR to assess hazardous execution and accurate refusal. Context Sensitivity uses PCA, UPR, SBR, and DER to distinguish correct context-dependent decisions from different decision errors. Hazard Resolution uses GSR, RSR, RSSR, and NHR to evaluate task completion, complete and partial hazard resolution, and newly introduced hazards, while Accuracy, CWR, and WCR measure plan selection and guardrail-induced changes. Hazard Prospection uses IR, UCR, and SCR to distinguish task incompletion, unsafe completion, and timely satisfaction of safety requirements.

Together, these task settings and metrics provide a comprehensive evaluation of complementary safety capabilities, rather than simply reporting additional safety scores.

(W2) Separating agent and guardrail capabilities

(1) Separation of agent and guardrail effects. Our experimental design explicitly distinguishes agents' baseline safety behavior from the effects of runtime guardrails. As described in Section 4.1, we evaluate three agent workflows on the same tasks under six configurations: Original (without guardrails) and five runtime guardrails. Original measures agent behavior without intervention, while comparisons against Original reveal how each guardrail changes safety outcomes under otherwise identical settings.

(2) Beyond over-refusal. Our findings are not limited to excessive rejection of safe actions. For example, RoboSafe improves average Pairwise Context Accuracy (PCA) from 0% to 53.35%, but still permits both safe and unsafe operations in 41.59% of context pairs, compared with only 3.37% where it blocks both (Table 2). Moreover, hazard-resolution and temporal-safety failures persist despite runtime interventions (Tables 3–4). These results show that our evaluation captures both insufficient hazard prevention and unnecessary blocking, rather than focusing solely on false refusals. Our goal is to evaluate agent safety behavior both with and without runtime intervention, rather than attribute the observed failures solely to agents or guardrails.

Thanks for the reviewers' responses. I take the diversity of evaluation protocols as a contribution. However, my concerns remain. For example, the responses regarding the guardrails' capabilities confirm my concerns about the separation of agent and guardrail effects. It is still unclear how guardrails enhance the safety of the agent -- whether as an additional layer of defense or by somehow enhancing the model's safety awareness. Therefore, I tend to maintain my original rating.

Thank you for the follow-up. First, we examine whether embodied agents themselves exhibit these safety failures. This is measured by the Original configuration, which uses no runtime intervention. We evaluate three agent workflows and further examine different VLM backbones (Appendix D.5) and a stronger reasoning configuration (Table R1). Without guardrails, agents reach hazardous states in 60.37% of direct hazardous requests and achieve 0% pairwise context accuracy (Tables 1–2). These limitations also persist across different backbones, with RSR remaining at 0–1.95% and implicit SCR at 0–1.33% in the cross-backbone comparison.

Second, we examine how effectively existing runtime guardrails improve safety when deployed on these agents. These guardrails act as an additional layer of defense rather than enhancing the model's intrinsic safety awareness. They introduce safety constraints during planning or intervene before proposed actions are executed, without modifying the underlying model's parameters. Comparing each guardrail with Original under identical settings therefore measures its effect on system-level safety. Guardrails can reduce hazardous execution, lowering ESR from 60.37% to 22.41% and raising PCA from 0% to approximately 53%. However, their benefits remain limited when agents must proactively resolve hazards or infer implicit safety requirements: RSR remains below 1% in the main evaluation (Table 3), and implicit SCR stays near 0% even among completed tasks (Table R2).

We hope this clarification addresses your concern and would appreciate your reconsideration of the rating.

Reviewer_x4goOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces EmbodiedSafe, a benchmark for evaluating the safety capabilities of embodied agents. The benchmark comprises 1,480 evaluation samples covering 77 hazard mechanisms and 20 temporal safety requirements, organized into four dimensions: Hazard Rejection, Context Sensitivity, Hazard Resolution, and Hazard Prospection. The evaluation includes paired safe and unsafe contexts with identical instructions and target operations, tasks requiring agents to complete household goals while resolving existing hazards, and temporal safety assessments with explicit or implicit requirements and deadlines. The experiments evaluate three embodied-agent workflows under six configurations, including an unprotected baseline and five runtime guardrails, with additional evaluations of three VLM backbones on a shared subset. The evaluation combines simulator-based execution outcomes, trajectory analysis, and semantic judgments validated through human annotation. The results show that existing guardrails can reduce hazardous execution, but substantial limitations remain in context-sensitive decision-making, proactive hazard resolution, and satisfying implicit temporal safety requirements.

Critical strengths
  1. The paper addresses safety challenges in embodied AI, an increasingly important research area as embodied agents become more capable and widely deployed. The topic is timely and has substantial practical significance.
  2. The paper introduces Hazard Resolution as an evaluation dimension and designs meaningful controlled experiments. By reporting Goal Success Rate (GSR), Resolution Success Rate (RSR), Resolution State Satisfaction Rate (RSSR), and New Hazard Rate (NHR), the evaluation helps distinguish different failure modes.
Critical weaknesses
  1. The paper does not sufficiently analyze the model's reasoning capabilities. Although it evaluates four VLM backbones, it lacks a analysis of reasoning depth and inference-time compute budgets. Some evaluation tasks, particularly Hazard Resolution and Hazard Prospection, require complex state reasoning and long-horizon planning. Increasing the reasoning budget or employing stronger reasoning configurations may improve performance on these tasks. Therefore, the current experiments do not sufficiently distinguish failures arising from the underlying models' limited reasoning capabilities from those attributable to the limitations of the runtime guardrails themselves.

  2. The paper leaves the safety criteria and action semantics ambiguous. In the thermal exposure case, the reference plan and the runtime guardrail appear to apply different safety judgments to the find Candle action, but the paper does not adequately clarify the action's execution semantics or safety boundaries. Specifically, it remains unclear whether this action constitutes hazardous exposure under the benchmark's criteria, how the reference plan's safety label is established, and whether the relevant action information is accessible to the guardrail. Furthermore, reference plan A and the blocked plan C share the same action prefix. Would selecting plan A instead trigger the same guardrail intervention? Since the alternative plan is not executed, this case also provides insufficient evidence to assess closed-loop safety recovery.

  3. The paper evaluates long-horizon safety reasoning in Hazard Prospection only to a limited extent. The experimental design of Hazard Prospection does not sufficiently support conclusions about long-horizon safety reasoning. First, tasks contain only approximately 6.44 annotated actions on average, and the paper lacks a systematic analysis of how performance varies with task length, the number of intervening actions, and the complexity of safety constraints. Second, approximately 70% of executions remain incomplete, meaning that the Safe Completion Rate (SCR) reflects both general task execution capabilities and safety reasoning abilities. The authors are encouraged to evaluate tasks of varying lengths and report safety-constraint satisfaction rates conditioned on task completion to better distinguish these two sources of failure.

Questions for authors

Please see the weakness.

3 comments

(W1) Reasoning and Budget

(1) Inference-time budget. We inspect termination records for Hazard Resolution and Hazard Prospection. Hazard Resolution shows no runs reaching the configured turn limit or planner timeouts. In Hazard Prospection, 2% reach the turn limit in one configuration, while planner timeouts range from 4% to 24%. These findings suggest that budget-related termination is not the primary source of the observed failures.

(2) Stronger model and reasoning configuration. We compare GPT-5.6 Sol with high reasoning effort against GPT-5.6 Luna with low reasoning effort on matched subsets: 20 Hazard Resolution executions, 20 Plan Selection cases, and 10 Hazard Prospection task pairs per configuration, with tasks, evaluation criteria, and guardrails fixed.

Table R1. Stronger-configuration results (%).

Model / reasoningGuardrailHR RSR ↑HR RSSR ↑Plan Acc. ↑CWR ↓HP Exp. SCR ↑HP Imp. SCR ↑
Luna / lowOriginal0.0025.0070.000.0050.000.00
Sol / highOriginal5.0032.5085.000.0050.000.00
Luna / lowRoboSafe0.0025.0075.000.0030.000.00
Sol / highRoboSafe0.0026.6765.0015.0030.000.00

HR = Hazard Resolution; HP = Hazard Prospection; Exp./Imp. = explicit/implicit requirements.

First, the stronger configuration improves hazard resolution: under Original, RSR rises from 0% to 5%, RSSR from 25% to 32.50%, and plan accuracy from 70% to 85%. Second, safety failures persist: with RoboSafe, RSR remains at 0% and CWR reaches 15%. For Hazard Prospection, explicit SCR is unchanged and implicit SCR remains 0% in all four configurations (Table R1). Because model and reasoning effort both change, this is a stronger-configuration sensitivity analysis, not a reasoning-budget ablation. We will report these results and budget settings in the appendix.

(W2) Action Semantics and Plan Selection

(1) Action semantics and safety validation. Under our low-level action interface, find Candle only navigates to and observes the target without manipulating it, and does not violate the annotated safety criteria. RoboSafe receives the proposed action, current observations, and execution history. Blocking this action is therefore overly restrictive under our evaluation criteria. For reference-plan validation (Appendix B.2.3), we execute each plan in the simulator from a fresh reset and verify goal completion, resolution of all initial hazards, and absence of new hazards. Three annotators independently review the results, and only tasks satisfying all three criteria with unanimous agreement are retained.

(2) Evaluation protocol. Plan Selection evaluates how guardrail feedback affects plan choices. When the initially selected plan is blocked, the agent can select a replacement, which is scored but not executed. In the candle example, RoboSafe blocks plan C, after which the agent selects plan B, which omits the required hazard-resolution action. The reference plan shares the same initial action but is not separately tested under RoboSafe in this episode. Thus, this example illustrates guardrail-induced degradation in plan selection, rather than closed-loop recovery. Actual hazard resolution is evaluated separately in Safe Completion through executed trajectories (GSR, RSR, RSSR, and NHR).

(W3) Temporal Safety

(1) Temporal task complexity. The reported mean of 6.44 counts annotated high-level actions, excluding navigation, search, and retry actions. The 50 Hazard Prospection tasks span 4–12 actions (mean 6.44), 0–8 activation-to-deadline intervening actions (mean 1.72), and 39 single-constraint versus 11 dual-constraint tasks.

(2) Completion-conditioned safety. To separate task incompletion from temporal safety failures, we supplement SCR with:

P(Safe∣Completed)=SCRUCR+SCR.P(\mathrm{Safe}\mid\mathrm{Completed})=\frac{\mathrm{SCR}}{\mathrm{UCR}+\mathrm{SCR}}.

Table R2. Safety satisfaction conditioned on task completion (%).

MethodExp.t ↑Imp. ↑
Original95.651.47
ThinkSafe100.000.00
Poex94.201.67
AgentSpec94.740.00
GuardAgent100.000.00
RoboSafe95.120.00

Under explicit requirements, 94.20–100% of completed tasks satisfy the safety constraints, indicating that task incompletion largely explains the low overall SCR. However, under implicit requirements, this rate remains at only 0–1.67%. Thus, the difficulty of satisfying implicit temporal safety requirements persists even when task incompletion is excluded.

Thank you for the response with additional analysis. The new result of tab. 1R and tab.2R address most of my concerns.

Thank you for acknowledging the additional analysis. We’re glad it addressed most of your concerns, and we hope it helps in your re-evaluation of the paper.

Reviewer_8RJ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents EmbodiedSafe, a benchmark for context-aware runtime hazard handling in household embodied agents. It structures evaluation into four levels: Hazard Rejection (L1), Context Sensitivity (L2), Hazard Resolution (L3), and Hazard Prospection (L4). The dataset contains 1,480 evaluation samples derived from 77 hazard mechanisms and 20 temporal safety requirements. In AI2-THOR, the authors evaluate three agent workflows with and without five runtime guardrails, plus supplementary backbone comparisons. Results suggest that guardrails can reduce immediate hazardous execution but remain much less successful at context-dependent decisions, hazard resolution, and temporally constrained task completion.

Critical strengths
  1. Evaluation Criterion. The benchmark covers multiple dimensions of embodied safety. Distinguishing unsafe execution from unnecessarily blocking safe actions (L2) is valuable. Paired safe/unsafe contexts provide a concrete way to test whether an intervention uses the environment rather than just the instruction text.

  2. Coverage of guardrails and workflows. The benchmark compares five runtime guardrails across three agent workflows, using simulator-validated hazard mechanisms. The execution-level outcomes help expose failure cases that pure text-based refusal evaluation might overlook.

Critical weaknesses
  1. Limited contribution (major). Recent works, including IS-Bench (AAAI 2026), SafetyALFRED (ACL 2026), and EMBGuard (ICML 2026), have already explored interactive hazard mitigation, corrective planning, and context-dependent safety assessment in embodied agents. I don't think line 141-145 illustrates significant novel contribution over exsiting ones. The proposed evaluation dimensions substantially overlap with these studies, making the incremental novelty insufficiently justified.

  2. Experimental concerns over prompting sensitivity and reproduction (major). Appendix D.3.1 shows that explicit corrective instructions improve RSR from nearly 0% to 60%–85%, raising concerns about prompt sensitivity. Considering the sensitivity, the complete system prompts should be provided, and simple safety-oriented prompting baselines should be added.

Questions for authors
  1. What fundamentally distinguishes EmbodiedSafe from existing agentic safety benchmarks?
  2. How would simple safety-oriented prompts (instead of heavy guardrails) affect the reported results?
1 comment

(Q1) Novelty over prior work

Sorry for being unclear. What fundamentally distinguishes EmbodiedSafe is that it evaluates whether agents respond appropriately to hazards as the task context becomes more complex, rather than merely detecting hazards or avoiding unsafe actions. Grounded in household hazard mechanisms and temporal safety requirements drawn from public safety guidelines and equipment manuals, L1–L4 progressively evaluate instruction-level refusal, context-dependent execution, active hazard resolution, and timely responses to implicit safety requirements.

Existing work has four key limitations:

  1. Limited hazard coverage. Existing benchmarks typically focus on selected hazard types or predefined scenarios, leaving broader household risks and temporal safety requirements insufficiently covered. EmbodiedSafe systematically covers 77 hazard mechanisms across six household safety categories and 20 temporal safety requirements, grounded in real-world safety guidance.
  2. Limited evaluation of context-dependent execution. Identifying unsafe actions does not establish whether agents can make appropriate decisions during actual task execution. For example, EMBGuard (ICML 2026) evaluates the safety of individual observation–action pairs, rather than agents' execution behavior. EmbodiedSafe evaluates agents with and without runtime guardrails, using matched safe and unsafe cases that differ in only one risk-relevant condition while keeping the instruction and target operation fixed. Importantly, blocking both cases also counts as a failure.
  3. Lack of execution-level hazard resolution. Recognizing hazards or predicting corrective actions does not establish whether agents actually resolve hazards during execution. For example, SafetyALFRED (ACL 2026) evaluates corrective action prediction using pre-rendered trajectories, without verifying the outcomes of autonomous execution. EmbodiedSafe directly checks whether agents resolve an existing hazard, complete an independent household goal, and avoid introducing new hazards.
  4. Limited evaluation of implicit temporal requirements. Checking whether mitigation actions occur in the correct order does not establish whether agents can infer unstated safety requirements and their deadlines. For example, IS-Bench (AAAI 2026) evaluates whether mitigation occurs before or after risk-prone actions, but does not isolate this inference capability. EmbodiedSafe pairs explicit and implicit versions of the same task to assess whether agents can infer the required intervention and its deadline, track them across intervening actions, and act in time.

We appreciate the suggestion to better clarify our novelty over prior work. We hope these clarifications make the contributions of EmbodiedSafe clearer. We will also revise the Related Work section to highlight these distinctions, including adding EMBGuard to the comparison.

(Q2) Complete prompts and instruction-specificity ablation

(1) Safety-oriented prompting baseline. Thanks for the suggestion. Our evaluation already includes Poex, which adds safety constraints directly to the planner's prompt without using a separate action-level guard. Thus, Poex serves as a planner-side safety prompting baseline (Sections 2 and 4.1). Compared with Original, Poex improves mean Direct ARR from 0% to 16.48% (Table 1), but ReAct RSR remains at 0% (Table 3). This suggests that general simple safety prompting can improve hazard rejection but offers limited benefits for hazard resolution.

(2) Complete prompts and instruction specificity. The general task-instruction template is provided in Appendix B.2.3, while Appendix D.3.1 examines the effect of adding sample-specific recovery requirements. Providing these requirements increases RSR from 0% to 60–85% (Table 12). This suggests that agents can execute specified corrections but struggle to determine which corrections are needed. In household tasks, users may not explicitly identify existing hazards or specify how to resolve them. Our main evaluation therefore tests whether agents can identify and resolve hazards from the observed environment while completing the task.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.