Evaluating under safety constraints reveals that they prioritize task completion over obstacle avoidance, even when prompted to avoid obstacles. The fault lies in planning, not perception or instruction.
SafeHarness addresses this by providing two explicit obstacle-aware harnesses: one for route planning that uses bounding boxes to plan and verify clear paths, and one for contact execution that selects contact positions to avoid obstacles.
On the SafeLIBERO benchmark, SafeHarness with GPT-6 achieves 71.9% task success and 87.5% collision avoidance, significantly outperforming previous state-of-the-art and the same agent without harnesses. This shows that explicit safety integration, rather than stronger planners or better prompts, is key.