Instance

building @tryinstance (YC S26) I like robots! @MIT_LISLab MIT

Instance is a verification layer for robotics teams that automatically judges whether a robot actually completed a task from its rollout footage. A user describes the task, drops in a dataset of episodes, and Instance returns a pass or fail verdict for each rollout along with subtask captions explaining what happened. The team positions it as an accuracy and latency improvement over prompting frontier vision language models to grade robot behavior, which matters because reinforcement learning, evaluation, and data curation pipelines are increasingly bottlenecked by the cost of humans watching video to label success. The company is aimed at robot learning teams that generate large volumes of episodes and need trustworthy labels before those episodes feed back into training. According to the company site, Instance works across different robots and camera angles, and the launch demo lets prospective users verify their own episodes through a hosted console. This launch positions the product as the first public entry point (a demo and API) into what has so far been an internal tool. Instance is part of Y Combinator's S26 batch and is led by Lucy Cai, an MIT computer science graduate who did robotics research in CSAIL's Learning and Intelligent Systems group under Leslie Pack Kaelbling, with prior stints on satellite software at SpaceX, automated testing at Amazon, and BCI pipelines at Blackrock Neurotech. Her cofounders, per the company site, are longtime friends from middle school with engineering backgrounds spanning SpaceX, AWS, and NASA JPL.

Tags

Comments (9)