Most Robotic

Instance’s evaluation system receives video from a physical robot rollout together with a description of the intended task and its subtasks. A learned success detector identifies whether the robot completed the overall objective and which intermediate steps succeeded or failed. The resulting labels attach automatically to repeated policy evaluations, eliminating manual video scoring for every trial.

The software sits inside the development loop between policy deployment and model improvement. Teams run a policy on their own robot, send the recorded episode through Instance and use the structured result to compare checkpoints or identify failure modes. The company’s longer-term roadmap includes a second robot that can restore the test scene between trials, extending the platform from automated scoring toward continuous physical evaluation.