Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation
The tested coding agent could talk about the obstacle, but still hit it in most runs.
The paper says the failure was not perception or wording: the prompt already banned contact, and the model mentioned the obstacle in its reasoning. The authors trace the problem to planning, where safety never becomes a priority during route choice or contact execution. Their SafeHarness adds obstacle-aware route planning and contact selection, reaching 71.9% task success and 87.5% collision avoidance. ArXiv · AI/CL/LG's note
The paper says the failure was not perception or wording: the prompt already banned contact, and the model mentioned the obstacle in its reasoning. The authors trace the problem to planning, where safety never becomes a priority during route choice or contact execution. Their SafeHarness adds obstacle-aware route planning and contact selection, reaching 71.9% task success and 87.5% collision avoidance. ArXiv · AI/CL/LG's note
score 5