Convergent Robotics

News

Benjamin Alt ·

What does RoboHarm tell us about physical AI misalignment?

Robocurve’s RoboHarm benchmark showed frontier LLM agents carrying out harmful instructions on a real robot. We gave the same models the same instructions in plain text, and then with the robot’s own camera image. They refused. Preliminary results from a small run, and what they do and do not tell us about whether safety training transfers from conversation to physical AI.

Read the post →

Benjamin Alt, Laurie Burns-Mill, & Andrew Nelson ·

Building safe physical AI: Open questions, risks, and a call for collaboration

Physical AI could remove the limits that the cost and availability of human labor place on what we can build. The capabilities that make that possible also create distinct and pressing safety problems. We lay out our intuitions about those risks, the questions the field has not yet answered, and an invitation to work on them with us.

Read the post →