acceptodds
Under review as a conference paper at ICLR 2027

RT-Safe: Benchmarking Agent Safety in Real-Time Embodied Environment

Abstract

Rapid progress in AI agents has brought growing attention to agent safety, with extensive evaluation focused on digital environments. As agents move into the physical world, embodied safety becomes increasingly important: failures can cause human injury and costly hardware damage. Beyond selecting safe actions, embodied agents must also operate under real-time constraints: the physical world does not pause while an agent reasons. As pedestrians move and vehicles approach during inference, an action that appears safe at observation time may become unsafe before execution. Real-time embodied safety therefore depends on both decision quality and decision latency. We introduce RT-SAFE, a simulated urban benchmark for evaluating embodied-agent safety under real-time constraints. RT-SAFE combines navigation tasks with moving actors, environmental hazards, and traffic rules, while allowing the world to evolve throughout inference and action execution. Across eight VLMs, agents achieve high task completion yet almost never complete safely: in the hardest setting, only 0.7% of episodes finish without a safety event. More strikingly, matched static and real-time evaluations yield task completion rates of 91.3% and 94.1%, respectively, while real-time execution increases collisions by . These results reveal that standard task success can mask substantial safety failures, and that decision latency itself can become a source of physical risk. Finally, we show that RT-SAFE can support offline RL training, where different reward designs produce distinct tradeoffs among task progress, collision avoidance, and successful completion.

Primary area:
datasets and benchmarks
Keywords:
Real-Time Safety, Embodied Agents, Safety Benchmarking

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_LDeuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces RT-SAFE, a benchmark for evaluating agent safety in real-time embodied environments. Built on SimWorld and Unreal Engine 5, it provides 36 navigation routes across five city maps, allowing the environment to evolve during both inference and action execution while separately recording Passive and Active Collisions. Evaluating eight VLMs, the authors find that task completion remains similar between Static and Real-time evaluation, while collision counts increase by 12.3× and Safe Success Rate (SafeSR) drops from 19.8% to 0.7%. The paper also investigates reasoning effort and demonstrates the benchmark's potential for Offline RL training.

Critical strengths
  1. RT-SAFE is a a practical benchmark for real-time embodied safety. It incorporates measured inference latency into a continuously evolving simulation and separately records safety events during inference and action execution, enabling a more comprehensive evaluation of real-time safety.

  2. It has many systematic experimental analysis. The paper evaluates eight VLMs and investigates the effects of environment difficulty, execution mode, reasoning effort, and navigation behavior on safety. It reveals a gap between task completion and safety, while also demonstrating the benchmark's potential for Offline RL training.

Critical weaknesses
  1. There are confounding factors in the Static versus Real-time comparison. Section 3.3 changes both simulator timing and prompt instructions. The Real-time prompt explicitly emphasizes continuous environment evolution and encourages conservative actions, potentially affecting action selection and decision frequency. This makes it difficult to attribute the observed safety differences solely to the real-time execution mechanism.

  2. Pedestrian behavior modeling and collision counting may affect evaluation reliability. Approximately 94% of collision counts in the Hard Real-time setting are Human Contacts, but the paper does not clarify whether scripted pedestrians can autonomously avoid the agent. Moreover, Appendix A.3 indicates that sustained contacts may generate repeated collision counts without general deduplication.

  3. The framework fails to adequately validate safe task feasibility under the restricted action space. RT-SAFE provides only 16 predefined movement, turning, and waiting actions, potentially limiting obstacle avoidance in dynamic environments. Without an Oracle Planner or Expert Policy under the same action space and real-time constraints, it remains unclear whether the extremely low SafeSR primarily reflects poor decision-making or limitations in task feasibility.

Questions for authors

Q1. Could the authors provide a controlled comparison using identical prompts while changing only simulator timing, to disentangle the effects of real-time execution and prompt instructions?

Q2. Do pedestrians autonomously avoid or react to the agent? Could the authors report deduplicated collision statistics based on independent contact events and analyze the sensitivity of the results to pedestrian behavior and counting rules?

Q3. Could the authors evaluate an Oracle Planner or Expert Policy with privileged simulator state under the same action space and decision-latency constraints, and report its SafeSR in the Hard Real-time setting?

Q4. Response latency varies substantially across VLMs, potentially confounding cross-model safety comparisons. Could the authors use matched decision delays or controlled artificial delays for the same model to better separate the effects of inference latency and decision quality?

3 comments

Thank you for your detailed review and for recognizing RT-SAFE’s practical value and systematic analyses. We address your concerns below, and additional experiments in our main response. W1 / Q1: Identical prompt comparison. Please see §2 of our main response. With the same neutral prompts, mean collisions increase from 4.06 in static to 18.22 in real time, and the shift toward shorter actions. These results strengthen our main finding that real time execution affects both safety and action selection, beyond differences in prompt instructions. W2 / Q2: Pedestrian behavior and collision counting. Please see §§1 and 4 of our main response. Pedestrians can make local adjustments, but they do not anticipate the agent’s future actions or compensate for its inference delay. RT-SAFE therefore evaluates whether the agent can navigate safely while its surroundings continue to evolve, without assuming that other actors will accommodate its reasoning time. Repeated contacts with the same actor are deduplicated within each recording window. When counting only whether an episode contains any collision, the proportion increases from 60.4% in static to 99.3% in real time. This shows that the safety gap extends across episodes, rather than arising solely from repeated collision counts in a few episodes.

W3 / Q3: Safe task feasibility. Please see §5 of our main response. Privileged trajectory search validated successful trajectories with zero recorded safety events using the same action space at both 0 second and 1 second decision delays for all tasks. These results demonstrate safe task feasibility, including under nonzero latency.

Q4: Reasoning quality versus inference latency. Please see §3 of our main response. At controlled delays of 0, 3, and 5 seconds, longer delays lower mean safety scores across all three effort settings, while higher effort does not consistently improve safety at matched delays. These results reinforce our central finding that more reasoning alone does not ensure safer behavior: reasoning effort and decision frequency must be considered together with the exposure time they have.

Thank you again for your detailed review. With around one day remaining in the rebuttal period, we would appreciate any follow up on whether our responses and additional experiments address your concerns.

We have further updated §5 of our general response: with the zero safety event trajectories validated on all 36 hard tasks at 0s and 1s delays, and UE-validated trajectories with zero collisions for all 36 easy tasks at the fix 3s delay, using the same 16 actions. This matches the difficulty and decision delay used in our RL experiments.

If you have any questions or concerns, we would be happy to address them during the remaining discussion period.

The additional experiments address most of my original soundness concerns. The restricted action space remains a limitation, but I no longer view it as sufficient to outweigh the paper’s contributions. I therefore raise my score from 2 to 3.

Reviewer_FF8ROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RT-SAFE evaluates VLM-based urban navigation while the environment continues running during inference. It compares static and real-time execution, studies reasoning effort, and explores offline RL, reporting substantial differences between task completion and safe completion.

Critical strengths

The paper addresses an important practical issue: inference latency can affect safety in a changing environment. The matched static/real-time evaluation provides a useful way to study this issue. Simulator-derived safety events and the passive/active collision split make the evaluation interpretable without relying on an external language-model judge.

Critical weaknesses

The collision counts would be easier to interpret with controls for stationary-agent exposure, actor avoidance, and repeated-contact counting. These details matter for understanding how much of the measured difference reflects avoidable risk. The seed analysis is helpful, but uncertainty estimates for the main comparisons would make the model rankings easier to assess. The RL results would also benefit from a clearer discussion of the trade-offs relative to behaviour cloning.

Questions for authors

Clarification: Do pedestrians avoid the agent, and how are repeated contacts counted? Could you clarify the release plan for the environment and evaluation records? Suggested additional evidence: Could you report a stationary-agent baseline and a latency-matched expert baseline, together with uncertainty estimates? Controlled-latency comparisons or an additional exposure-normalised metric would also help interpret the results.

3 comments

Thank you for your thoughtful review and suggestions. We appreciate your recognition of the practical importance of inference latency, the matched static vs real time comparison, and the interpretable safety event breakdown.

W1: Pedestrian behavior and collision counting. Please see §§1 and 4 of our main response. Pedestrians can make local adjustments, but they do not anticipate the agent’s future actions or accommodate for its inference delay. RT-SAFE therefore evaluates whether the agent can navigate safely while its surroundings continue to evolve, without assuming that other actors will accommodate its reasoning time. Repeated contacts are deduplicated within each recording window, and the safety gap persists when counting only whether each episode contains any collision.

W2: Uncertainty and model rankings. The paired bootstrap analysis shows 37.37 additional collisions per episode in real time compared with static, with a 95% confidence interval of [28.00, 48.96], supporting our central comparison. Most of the model level differences are also supported, and we will report these uncertainty estimates and qualify ranking claims accordingly.

W3: RL versus behavior cloning. Please see §6 of our main response. With the same expert data, RL reduces collisions but lowers task success. Adding policy generated training data restores success to 87.5% while reducing collisions per 100 m of route progress from 6.7 to 2.0.

Q1: Pedestrian avoidance, repeated contacts, and release. Please see W1 above for pedestrian behavior and collision counting, and §7 of our main response for release details.

Q2: Stationary-agent baseline. We kept the agent stationary at the starting positions of the same 36 hard tasks for 60 simulated seconds. Collisions occurred in 22/36 episodes (61.1%), averaging 0.94 per episode (95% bootstrap CI: [0.64, 1.28]). The rate was 0.94 collisions per simulated minute, compared with 3.17 during the neutral prompt real time navigation. Thus, remaining stationary does not guarantee safety, reinforcing our benchmark’s design goal: the agent should take responsibility for its own safety and avoid harm to surrounding people and objects, without relying on others to move out of its way.

Q3: Expert baseline. Please see §5 of our main response. Privileged trajectory search validated successful trajectories with zero recorded safety events using the same action space at both 0 second and 1 second decision delays. These results demonstrate safe task feasibility, including under nonzero latency,

Q4: Controlled latency and exposure normalization. Please see §3 of our main response for the controlled delay results and W2 above for uncertainty estimates. Longer assigned delays lower mean safety scores across all three effort settings, while higher effort does not consistently improve safety. We additionally normalized the original eight model results. The gap remains large after distance normalization. Average recorded episode duration increases from 1.12 to 22.78 minutes, showing that longer exposure accompanies more accumulated collisions.

EnvironmentCollisions per episode ↓Collisions per 100 m traveled ↓
Static3.313.38
Real-time40.6841.08
Reviewer_FF8ROpenReview ↗

Follow-up after author response

Thank you for the clarifications and additional experiments. The fixed-delay control, explicit collision-counting rules, and updated RL/BC comparison address my main concerns. The limited expert-delay coverage and incomplete environment release remain limitations. On balance, I am revising my recommendation to weak acceptance.

Thank you so much for your feedback, follow up, and revising your recommendation to weak acceptance. Your suggestion really help strengthen both the experiment evidence and reproducibility of RT-SAFE.

We have further updated § 5 and § 7 of our general response to address your remaining concerns. In §5, we extend the privileged trajectory search to a fixed 3 second decision delay on all 36 easy real time tasks, matching the difficulty and decision delay used in our RL experiments. Using the same 16 action space, we found and UE-validated a successful trajectory with zero collisions for every task.

In §7, we've updated the anonymous repository, including access to the packaged Linux runtime for all five maps, agent and action interfaces, task definitions, evaluation code, example rollout records, and support for reproducing the additional controls.

We truely appreciate your assessment and would be happy to address any further questions if you have.

Reviewer_2GrSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RT-SAFE evaluates embodied-agent safety by creating a simulated environment to change normally during model inference. RT-SAFE has evaluated eight VLMs where their task completion remains high but collisions increase under real-time execution. The work try to address a relevant problem about safety events and latency.

Critical strengths
  1. The paper focus on an important evaluation gap: embodied agents are supposed to remain safe while the environment changes during inference, which is more challenging and practical in real-world application.
  2. The benchmark provides useful information by evaluating eight VLMs and reporting different safety-event categories.
Critical weaknesses
  1. It is unclear whether pedestrians can avoid the agent during inference. If they simply follow fixed paths, they may collide with the agent while it is stationary. This could lead to that more collisions happen, especially since most collisions are reported to occur during inference.
  2. The static and real-time evaluations use different timing prompts, making it difficult to isolate the effect of latency. Also, prolonged contact may be counted as multiple collisions.
Questions for authors

Do pedestrians can actively avoid the agent while it is reasoning? Does the safety degradation persist when static and real-time evaluations use the same neutral prompts and inference delays are experimentally controlled?

3 comments

Thank you for your thoughtful review and for recognizing the importance of the real time evaluation gap and our safety evaluations. We address your concerns through the clarifications and experiments in our main response.

W1 / Q1: Pedestrian behavior. Please see §1 of our main response. Pedestrians can make local adjustments, but they do not anticipate the agent’s future actions or compensate for its inference delay. RT-SAFE therefore evaluates whether the agent can navigate safely while its surroundings continue to evolve, without assuming that other actors will accommodate its reasoning time.

W2: Prompt differences and collision counting. Please see §§2 and 4 of our main response. The safety gap between static and real time evaluation stays with the same neutral prompts and when counting only whether each episode contains any collision.

Q2: Controlled prompts and inference delays. Please see §§2–3 of our main response. With neutral prompts, mean collisions increase from 4.06 in static to 18.22 in real time. Under controlled delays, longer latency lowers mean safety scores across all three effort settings, while higher effort does not consistently improve safety.

Thank you again for your thoughtful review. With around one day remaining in the rebuttal period, we would appreciate any follow up on whether our responses and additional experiments address your concerns.

§§1~4 of our general response clarify pedestrian behavior and collision counting, and provide identical prompt and controlled delay comparisons. We have also updated §5 with additional expert trajectory results and §7 with the environment release details.

If you have any other questions or concerns, we would be happy to address them during the remaining discussion period.

Thank you for your responses and additional experiments. These have addressed my main concerns about your work. However, the pedestrian behavior still have limitations because of its gap of the real world. Therefore, I am revising my recommendation to weak acceptance.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response to the reviewers

We thank to the reviewers for their thoughtful feedback, which has helped us clarify and strengthen our work. We truly appreciate their recognition of RT-SAFE’s main contributions:

  • The important real time evaluation gap: RT-SAFE studies agent safety while the environment continues evolving during inference, addressing an important practical issue highlighted by 2GrS and FF8R to evaluate the agent in real time embodied environment.
  • A practical and interpretable benchmark: The static vs real time comparison and the simulator derived passive & active collision breakdown enable interpretable evaluation without an external language model judge (FF8R). Incorporating measured inference latency also supports a more comprehensive evaluation of real time safety (LDeu).
  • Useful empirical analyses: The evaluation of eight frontier VLMs provides information across safety event categories (2GrS). The systematic analyses of difficulty, execution mode, reasoning effort, and navigation behavior reveal the gap between task completion and safety, while demonstrating the benchmark’s potential for RL training (LDeu).

Below, we clarify the design choices raised by the reviewers and present additional experiments that strengthen the evidence for our findings.

1. Pedestrian Behavior (reviewers 2GrS, FF8R, LDeu): In our environment, the pedestrians would follow predefined routes and can make local adjustments, such as detouring around blockages. However, they do not explicitly anticipate the agent’s future actions or adjust their behavior to keep the agent safe while it is reasoning. This is a deliberate design choice: RT-SAFE evaluates whether an agent can navigate safely while the surrounding world continues to evolve, without assuming that others will accommodate the time it spends reasoning. Consequently, remaining stationary during inference is not automatically safe; the agent must account for changes that may occur before its next action begins. A related assumption appears in the “invisible robot” setting of Chen et al. (2019), where pedestrians respond to other pedestrians but not to the robot. We will clarify this distinction and the intended scope of the benchmark in the revised paper.

Chen, C., Liu, Y., Kreiss, S., & Alahi, A. (2019). Crowd-Robot Interaction: Crowd-aware Robot Navigation with Attention-based Deep Reinforcement Learning. IEEE International Conference on Robotics and Automation (ICRA).

2. Separate the prompt effect from environment mode (reviewers 2GrS, LDeu): Here are the results when evaluating Qwen3.8-27B-FP8 using an identical neutral prompt in both static and real time environments, on 36 hard tasks, using the same environment seed (0), and medium reasoning effort. Mean collisions increased from 4.06 to 18.22 per episode in real time, while 1 m moves became more frequent and 4 m moves became less frequent. These results strengthen our finding that the safety gap between evaluating in static enviornment, and the model's behavior in shifting toward shorter actions in real time envionrment. We will include this controlled comparison in the revised paper.

EnvironmentSR (%) ↑SPL ↑Collisions per episode ↓1 m action(%)4 m action(%)
Static88.90.8554.069.946.5
Real time97.20.95318.2216.038.8

Qwen Team. "Qwen3.8-Max: A New Bar for Coding and Cowork." Qwen, Aug. 2026, qwen.ai/blog?id=qwen3.8

AuthorsOpenReview ↗

General response to the reviewers (continue)

3. Separate reasoning quality from inference latency (reviewers 2GrS, FF8R, LDeu): We evaluated Qwen3.8-27B-FP8 on the same 36 hard tasks using Low, Medium, and Xhigh effort settings at fixed simulated decision delays of 0, 3, and 5 seconds. The simulator pauses during actual inference, then advances by the assigned delay before executing the action, ensuring equal per decision delays across effort settings. Safety scores use the same formula in Figure 5 (and Appendix B.5) in the paper. These results strengthen two findings from our paper: longer decision latency is associated with lower safety scores under all three effort settings, and higher reasoning effort does not necessarily improve overall safety, even at the same latency. Also, the cautious intention expressed by higher reasoning effort also does not necessarily translate into safer actions. For example, Xhigh explains “take a short step and reassess before moving farther” at 5 second, and it selects shorter movements and makes more decisions, accumulating more delay.

Delay per decisionEffortSR (%) ↑SPL ↑Safety score ↑
0 sLow91.70.888−10.64
0 sMedium83.30.816−7.69
0 sXhigh86.10.825−10.22
3 sLow97.20.952−13.86
3 sMedium91.70.900−18.03
3 sXhigh94.40.914−15.83
5 sLow91.70.899−20.61
5 sMedium94.40.921−21.78
5 sXhigh94.40.918−31.06

4. Collision Counting (reviewers 2GrS, FF8R, LDeu): In our simulator’s collision counter, repeated contacts with the same pedestrian or object are counted only once within each recording window. A new window begins when a new movement begins, so the same asset can be counted again in later windows. We will clarify these rules in the paper. To examine whether repeated counting drives the observed gap, we also counted each episode only once if it contained any collision. Across eight models and 36 tasks per mode, collisions occurred in 60.4% of static episodes (174/288) and 99.3% of real time episodes (286/288). This shows that the static vs real time safety gap extends across episodes, rather than arising solely from repeated counts in a few episodes.

5. Feasibility with a privileged expert using the same action space (reviewers FF8R, LDeu): To directly test whether the restricted action space permits safe completion, we did a privileged state trajectory search on all 36 tasks in the hard realtime setting, using the same 16 actions and original decision budgets. For every task, we found and UE-validated a successful trajectory with zero recorded safety events at each of two simulated decision delays: 0 s and 1 s.

We additionally extended this analysis to a fixed 3-second decision delay on easy real time tasks, which approximately the average action-generation latency and match the setup from the RL section in the paper. We found and UE validated successful trajectories with zero collisions for all tasks.

Validation used the original simulator and safety-counting rules, with the environment continuing to evolve throughout each assigned delay and action execution. These results establish that the action space permits safe completion across the tasks, including under nonzero decision latency. This shows that the near zero SafeSR does not result from the action space which makes safe completion impossible.

6. RL Results Update (reviewer FF8R): To clarify the safety success trade off relative to behavior cloning (BC), we updated our experiments using Qwen3.5-4B on 16 tasks from two maps across two environment seeds, with Easy difficulty and a fixed 3-second decision delay. With the same expert training data, RL reduces collisions per 100 m of route progress from 6.7 to 3.7 compared with sampled BC, while success decreases from 87.5% to 78.1%. And adding policy generated trajectories and retraining restores success to 87.5% and further reduces collisions to 2.0 per 100 m. These results support the benchmark’s potential for learning safer policies, while showing that the final improvement also involves additional training data. We will include the updated RL results in the paper.

MethodSR (%) ↑SPL ↑Collisions per episode ↓Collisions per 100 m of route progress ↓
Prompting25.00.2499.1617.8
BC (argmax)34.40.3437.1913.3
BC (sampled)87.50.8655.286.7
RL (expert data)78.10.7752.813.7
RL (expert + RL data)87.50.8701.592.0
AuthorsOpenReview ↗

Updated Code Release and Revised Manuscript

7. Release (reviewer FF8R): We have updated the anonymous repository and revised manuscript. The repo provides agent and action interfaces, prompts, task definitions, evaluation code, and example rollout records; it also provides the access to the packaged Linux runtime with all five maps, environment check, and configurations and execution support for the identical prompt, fixed delay, and stationary agent controls. We thank all the reviewers for their feedback, which helped improve the clarity and reproducibility of RT-SAFE.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.