Instance Wants to Grade Every Robot - So Nobody Has to Watch

Instance is a San Francisco company in Y Combinator's Summer 2026 batch building the autonomous evaluation rig for robot learning. Its first product is a success detector that watches a robot's rollout video, reads the task description, and returns a verdict on whether the task was completed, with a subtask-by-subtask breakdown of where the run went right and where it fell apart. The company reports that its purpose-built detector beats frontier vision-language models on both accuracy and latency across more than 10,000 human-labeled episodes, seven robot platforms, and eight benchmarks, a narrowness it treats as the point: a model answering only whether a robot finished a task can be tuned to do that cheaply and quickly, at the volume robot teams run evaluations. Founded by MIT classmates and childhood friends Claire Mao, who built planetary-atmosphere simulation software at NASA's Jet Propulsion Laboratory, and Lucy Cai, who worked on satellite software at SpaceX, LLM testing at Amazon, and brain-computer interfaces at Blackrock Neurotech, the company positions its verification layer as the picks-and-shovels infrastructure beneath the current robotics buildout, with a public demo where teams submit a rollout video and watch a verdict returned live.