The author tested GPT-6 Astra controlling a low-cost SO-101 robot arm through the Inspect Robots harness.
Robot safety check catches harmful commands missed by control loop
Asking the controlling model to judge an instruction before moving identified all 15 hazardous commands in a small robot-arm study.
Academic
Arnav Grover
Purdue University
Research Digest··3 min read
Grover shows that a model which rejects harmful actions when explicitly asked about them may still pursue those actions inside a robot control loop.
§