Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation
LLM coding agents routinely collide with obstacles during robot manipulation; SafeHarness route and contact planning lifts collision avoidance to 87.5%.
Researchers evaluate coding-agent robot controllers under a safety constraint pairing each manipulation goal with an obstacle the robot must not touch, finding agents complete tasks but collide in most cases despite reasoning about the obstacle and receiving forbidding instructions. The failure is traced to planning: no notion of a clearing route, no replanning, and no awareness that contact execution is bounded by the constraint. The proposed SafeHarness adds obstacle-aware route planning with bounding boxes and waypoints plus obstacle-aware contact execution, reaching 71.9% task success and 87.5% collision avoidance, surpassing prior SOTA by 6.5% and 27.0% and beating the unharnessed agent by 2.3x and 1.5x.
- Coding agents neglect safety constraints even when prompts explicitly forbid obstacle contact
- Failure localized to planning: no clearing-route notion, no replanning, unconstrained contact execution
- SafeHarness grounds obstacles as bounding boxes and plans, verifies, then executes waypoint routes
- Obstacle-aware contact execution selects contact positions that avoid the obstacle
- AI models
- SafeHarness
From the original article
1 min read at arxiv.orgCoding agents have emerged as a promising paradigm for robot manipulation: a language model writes the robot controller as a program, and agents built in this way now operate robots without robot-specific training. Whether this paradigm is also safe, however, has not been asked. We evaluate coding agent under a safety constraint, where each task pairs a manipulation goal with an obstacle the robot must not touch.
By Bingxin Xu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.