inferwire
/
AI·2 min read

SafeHarness Stops Robot-Coding Agents From Hitting Obstacles

AI agents that write robot controllers knew about obstacles but hit them anyway. SafeHarness adds route and contact planning that raised collision avoidance to 87.5 percent.

TL;DR

  • AI coding agents that write robot control programs usually collided with an obstacle they had been told not to touch, even while reasoning about it [^1].
  • SafeHarness, adding obstacle-aware route and contact planning, reached 71.9 percent task success and 87.5 percent collision avoidance, 6.5 and 27.0 points above the previous best [^1].

Background

In one approach to robot manipulation, a language model writes the robot's controller as a program, and agents built this way now operate robots without robot-specific training [^1]. Researchers from USC, UCF and UC Santa Barbara asked whether this is also safe: what happens when a task comes with an obstacle the robot must not touch [^1]?

What happened

In most cases, the agent pursued the goal and hit the obstacle anyway, treating task completion as its only objective [^1]. Neither perception nor instructions were at fault: the agent reasoned about the obstacle in its traces, and the prompt already forbade touching it. The failure was in planning, where the stated constraint never became a priority [^1]. Splitting manipulation into the route and the moment of contact showed two gaps. Along the route, the model had no notion of a clear path and no way to replan once a chosen route became impossible; at the contact, it did not realize the same constraint applied [^1].

SafeHarness gives the model two tools to close those gaps. Obstacle-aware route planning represents objects as bounding boxes and draws candidate routes over them as waypoints; the agent plans a route in advance, verifies it, replans if needed and only then executes [^1]. Obstacle-aware contact execution chooses the contact position so the contact itself avoids the obstacle [^1]. Compared with the same agent without the harnesses, task success rose 2.3 times and collision avoidance 1.5 times [^1].

Why it matters

The study shows a failure mode that better prompts do not fix: a model can know a rule and state it, yet never weigh it during planning. For robots, that is the difference between a demo and something safe to put near people. Giving the agent structured planning tools turned the constraint into something it actually acts on [^1].

Practical example

Say a robot arm must move a cup across a table with a vase in the middle. A plain coding agent writes a straight path and knocks the vase over, despite having noted it. With SafeHarness, the agent maps the vase as a box, draws routes around it, checks the chosen one and picks a grip that keeps the arm clear [^1].

Related gear

We recommend this textbook because it provides a comprehensive foundation in robot navigation, path planning, and safety architectures.

AdvertisementAmazon

Introduction to AI Robotics

Sources

  1. [1]arXiv — Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation