Soham Dan has built RoboDx, which generates repair tasks for robot navigation software in simulation, each with faults deliberately planted in it, according to a paper on arXiv. Before a repair is counted as verified, the robot has to drive correctly in test scenarios that the agent is never shown. In a second set of tasks, the agent cannot see the code and has to choose the cause from a list of possible explanations.

On a set of 100 tasks, almost every repair made by Claude Opus 5.5 with standard coding tools was verified. On 50 tasks with two faults, though, only 17 of its 49 verified repairs removed both planted edits. Most faults left behind were configuration constants whose correct values the documentation does not give, and without seeing the code, Opus named both faults in 10 of those 50 tasks.

Given the same measurements of each faulty robot, four current models identified the cause in no more than 38 of 94 tasks, but they were right on almost all of them when they also got the simulator's output for every hypothesis. On 97 tasks taken from real bug fixes, Claude Haiku 5.5 scored lowest.

Source: arXiv Robotics (cs.RO)

Background

Soham Dan

Robotics researcher

Robotics researcher.