AI 研究arXiv cs.LG8 Oct, 12:00 pm SGTSIGNAL B76E5D

RT-SAFE: Benchmarking Agent Safety in Real-Time Embodied Environments

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

Existing agent safety evaluations mostly focus on digital environments. This work introduces RT-SAFE, a simulated urban benchmark in which the world keeps evolving during inference and action execution, bringing decision latency into embodied safety assessment. Its value lies in showing that high task completion can mask safety failures and in offering an environment for offline RL training to reduce collisions.

01 ABSTRACT

RT-SAFE is a simulated urban benchmark for evaluating embodied-agent safety under real-time constraints, combining navigation tasks with moving actors, environmental hazards, and traffic rules while allowing the world to evolve throughout inference and action execution. Across eight VLMs, task completion is high but safe completion is extremely rare: in the hardest setting, only 0.7% of episodes finish without a safety event, and matched static versus real-time evaluations yield task completion rates of 91.3% and 94.1% while real-time execution increases collisions by about 12.3x. The authors argue that standard task success can mask substantial safety failures and that decision latency itself can become a source of physical risk, and they show the benchmark can support offline RL training to substantially reduce collision rates.

02 KEY FINDINGS

RT-SAFE incorporates real-time constraints into embodied-agent safety evaluation: the world does not pause while the agent reasons, so an action that appears safe at observation time may become unsafe before execution.
Across eight VLMs, task completion is high but safe completion is extremely low, with only 0.7% of episodes finishing without a safety event in the hardest setting.
Matched static and real-time evaluations yield similar task completion rates (91.3% and 94.1%), but real-time execution increases collisions by about 12.3x, indicating that latency can translate into physical risk.

03 RELEVANCE

As agents move from digital environments into the physical world, failures can cause human injury and hardware damage. This benchmark suggests that measuring embodied agents only by task success systematically underestimates safety risk and that real-time latency should be part of safety metrics. It can also support offline RL training, providing an evaluable path to reduce collisions under real-time constraints.

04 EVIDENCE BOUNDARY

The work is based on a simulated urban environment, and its conclusions cannot be directly extrapolated to real roads or real robotic systems. It evaluates eight VLMs and does not represent all embodied-agent architectures. Safety events and collisions are defined and measured by the simulator, leaving real-world sensor noise, actuator errors, and complex human-agent interaction uncovered. The offline RL training demonstration shows improvement within this benchmark, not a guarantee of safety in real scenarios.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG

Read original ↗