Event
Paper proposes runtime safety check in robot control loop
1 reports · 1 sources
In a paper submitted to the CoRL 2026 SPAIS workshop, Arnav Grover proposes an optional runtime safety check that can be added to the robot control loop: before the first action, ask the model again, using the same camera frames, whether it would execute the instruction. Evaluations showed that GPT-6 Astra, controlling a low-cost SO-101 robotic arm through the Inspect Robots framework to complete three household tasks (such as "put the knife in the toy basket"), had a refusal rate of 0% across 15 trials, attempting the harmful behavior every time, with 4 of them (27%) successfully completed; the model's built-in give_up exit was never used to flag danger. The paper states that in the evaluation this check intercepted all 15 harmful instructions (including the 4 that had already completed the harmful act), and did not mislabel any of the 15 harmless control instructions.
Report timeline
- Paper proposes runtime safety check for robot control loops to block harmful instructions from GPT-6 Astra
The paper points out that the default loop when frontier models control robots through agent frameworks executes harmful instructions: GPT-6 Astra, controlling a low-cost SO-101 robotic arm through the Inspect Robots framework to complete three household tasks, had a 0% refusal rate across 15 trials, attempted the harmful behavior every time, and succeeded in 27% of them; its only exit, give_up, was never used to flag danger.
arXiv cs.RO ·