RoboSafe: Safeguarding Embodied Agents via Executable Safety Logic

TL;DR

RoboSafe addresses the vulnerability of vision-language model (VLM) powered embodied agents to hazardous instructions by introducing a runtime guardrail that relies on executable safety logic. It fuses forward predictive reasoning over a long-term safety memory with backward reflective reasoning over short-term trajectories to detect contextual and temporal risks, outputting verifiable safety actions that block unsafe steps or trigger replanning. The framework builds a long-term safety knowledge base and a short-term trajectory memory, deriving safety predicates and executable logic that are interpretable and verifiable at runtime. Across multi-agent simulations and a real-world robotic arm, RoboSafe significantly reduces hazardous actions (e.g., up to a 36.8% reduction in risk occurrence) while preserving task performance and showing resilience to jailbreak attacks, highlighting its practical impact for safe embodied AI.

Abstract

Embodied agents powered by vision-language models (VLMs) are increasingly capable of executing complex real-world tasks, yet they remain vulnerable to hazardous instructions that may trigger unsafe behaviors. Runtime safety guardrails, which intercept hazardous actions during task execution, offer a promising solution due to their flexibility. However, existing defenses often rely on static rule filters or prompt-level control, which struggle to address implicit risks arising in dynamic, temporally dependent, and context-rich environments. To address this, we propose RoboSafe, a hybrid reasoning runtime safeguard for embodied agents through executable predicate-based safety logic. RoboSafe integrates two complementary reasoning processes on a Hybrid Long-Short Safety Memory. We first propose a Backward Reflective Reasoning module that continuously revisits recent trajectories in short-term memory to infer temporal safety predicates and proactively triggers replanning when violations are detected. We then propose a Forward Predictive Reasoning module that anticipates upcoming risks by generating context-aware safety predicates from the long-term safety memory and the agent's multimodal observations. Together, these components form an adaptive, verifiable safety logic that is both interpretable and executable as code. Extensive experiments across multiple agents demonstrate that RoboSafe substantially reduces hazardous actions (-36.8% risk occurrence) compared with leading baselines, while maintaining near-original task performance. Real-world evaluations on physical robotic arms further confirm its practicality. Code will be released upon acceptance.

Figures (6)

• Figure 1: Illustration of RoboSafe, where runtime safety guardrail generates executable safety logic to eliminate implicit temporal hazards and prevent contextual risks under dynamic scenarios.
• Figure 2: Overall runtime guardrail framework. RoboSafe introduces both Forward Predictive Reasoning for preventing contextual risks and Backward Reflective Reasoning for mitigating temporal risks, ensuring defense effectiveness in dynamic, temporally dependent scenarios.
• Figure 3: Case studies on contextual unsafe instructions. Red parts represent the final hazardous actions, which are precisely blocked by RoboSafe and actually not executed.
• Figure 4: Case studies on temporal unsafe instructions. Blue parts represent the replanning actions, which proactively mitigate the temporal hazards on long-horizon task process.
• Figure 5: (a) Ablation study on different guardrail VLMs (%). (b) Ablation study on different $\lambda$ (%).