Coding Agent with Guardrails Achieves Safe Robot Control

Researchers evaluated the safety of a coding Agent that writes robot control programs using large models under obstacle-avoidance constraints, and proposed a guardrail scheme, SafeHarness. Each task pairs an operation goal with an obstacle that must not be touched. The authors found that the Agent collides with the obstacle in most cases, and that the problem lies in the planning stage never treating the safety constraint as a priority, rather than in perception or instructions being wrong. The authors split the operation into path segments and contact segments, using two obstacle-aware guardrails to respectively perform advance planning, verification and replanning based on bounding boxes and waypoint sequences, and to select a contact position that avoids the obstacle at contact time. The authors report that when paired with GPT-6-Astra, the task success rate reaches 81.2% and the obstacle-avoidance rate reaches 91.9%, 13.7 and 23.0 percentage points higher than the previous SOTA, respectively, and 31.2 and 57.5 percentage points higher than the same Agent without guardrails.