Safe Embodied AI for Long-horizon Tasks: A Cross-layer Analysis of Robotic Manipulation

A comprehensive survey examines safety mechanisms for embodied AI systems performing long-horizon robotic manipulation tasks, identifying critical gaps in current research across planning, policy design, and execution phases. The analysis reveals that while safety receives attention, evidence remains fragmented with limited formal guarantees, particularly for contact-rich manipulation scenarios in real-world deployment.

This arXiv survey addresses a fundamental challenge in autonomous robotics: ensuring safe operation of AI systems that must reason and act over extended time periods in physical environments where failures directly harm people or damage infrastructure. The research comes amid accelerating development of embodied AI capabilities, where systems increasingly transition from controlled laboratory settings to real-world deployment scenarios. The fragmentation identified across planning-time, policy-time, and execution-time safety mechanisms reveals that the robotics community lacks unified frameworks for comprehensive safety assurance.

The survey's emphasis on long-horizon robotic manipulation exposes how multiple failure modes compound within single systems—semantic misunderstanding of tasks, cascading errors across subtasks, execution drift from planned trajectories, and unpredictable contact dynamics all interact to create safety challenges that single-layer defenses cannot adequately address. This multi-layered risk profile distinguishes embodied AI safety from software-only systems where failures typically carry lower physical consequences.

For developers and organizations deploying robotic systems, the identified gaps present both challenges and opportunities. The lack of manipulation-specific safety benchmarks hampers validation of safety claims, creating regulatory and liability uncertainties. The weakness of formal guarantees for contact-rich tasks suggests current systems rely on empirical heuristics rather than provable safety properties—a significant concern for high-stakes applications like surgical robotics or hazardous environment operations.

The research direction toward cross-layer assurance indicates the field recognizes that safety cannot be bolted on as an afterthought but requires integrated design from initial planning stages through deployment. Organizations advancing robotic solutions should prioritize formal verification methods and comprehensive evaluation protocols aligned with deployment contexts.

• →Current safety research in embodied AI is fragmented across three intervention points—planning, policy design, and execution—without unified assurance frameworks.
• →Contact-rich long-horizon manipulation tasks lack sufficient formal safety guarantees, relying instead on empirical heuristics with unclear reliability boundaries.
• →Policy-time safety mechanisms remain underdeveloped compared to planning and execution-time approaches, creating a critical research gap.
• →Manipulation-specific safety benchmarks are immature, hindering validation of safety claims for real-world robotic deployment scenarios.
• →Cross-layer safety design integrating multiple intervention levels from initial planning through deployment represents the emerging consensus for robust embodied AI systems.