Responding to the call. I maintain an open-source red-team for robot VLA policies, and the hardest open question for me is what counts as evidence. In simulation, I can show an odd instruction takes a policy out of its safe region, with a benign control beside it. What I can't show yet is whether that transfers to a real arm. If anyone reading this runs learned policies on hardware and wants to compare notes on how you decided a checkpoint was safe enough, I'd like to talk.
Responding to the call. I maintain an open-source red-team for robot VLA policies, and the hardest open question for me is what counts as evidence. In simulation, I can show an odd instruction takes a policy out of its safe region, with a benign control beside it. What I can't show yet is whether that transfers to a real arm. If anyone reading this runs learned policies on hardware and wants to compare notes on how you decided a checkpoint was safe enough, I'd like to talk.