Introducing Instance: Evaluation for Robot Policies

Instance launched its success detector for robot policies, the first piece of what the company intends to grow into fully automated robot evaluation. Teams describe a task and upload or record a rollout video, and the verifier judges whether the robot actually completed it. The company was founded by Claire Mao and Lucy Cai, MIT-trained computer scientists who have been friends since middle school. Cai did humanoid robotics research in MIT's Learning and Intelligent Systems lab, wrote collision-avoidance software at SpaceX that handled the entire Starlink fleet, and built Amazon's first LLM-based unit test generation at AWS; Mao wrote planetary atmosphere simulation software at NASA JPL and did propulsion research at the MIT Media Lab. The idea came from Cai's robotics research, where she spent a year watching a humanoid attempt tasks and marking success by hand. The pair benchmarked their judge on more than 10,000 held-out, human-labeled episodes across 7 robot platforms and report it beats a frontier vision-language model on accuracy at a fraction of the latency, and they bought their own robot arm to train policies on.