Mind the Refinement Gap: When Safe High-Level Robot Plans Produce Unsafe Executions

High-level robot plans can be unsafe even when approved by safety monitors: The paper found that plans deemed "safe" at an abstract level produced unsafe executions in practice, revealing a critical gap between what safety systems check and what robots actually do during execution.

The problem stems from missing details in plan abstraction: Safety monitors typically verify simplified, high-level action sequences, but they often miss navigation paths and implicit effects that occur during actual robot execution—a gap the researchers call the "refinement gap."

This vulnerability affects real-world systems: The study tested RoboGuard, a safety system used in language-enabled robots, and found that 12 out of 12 targeted test cases showed discrepancies between what the safety monitor approved and what the robot actually executed, demonstrating this isn't a theoretical concern.

Graph-based refinement offers a practical solution: The researchers propose checking safety against detailed execution traces (including navigation and implicit actions) rather than just abstract plans, which they show can catch unsafe behaviors that high-level monitors miss.

The issue is particularly relevant for AI systems following natural language instructions: End-to-end tests showed the problem occurs when robots generate plans from human language instructions, suggesting this gap could affect increasingly common voice-controlled or instruction-following robot systems.