Coliseo
Instance is ground truth for robot learning. Labs generate thousands of robot episodes a week — far more than anyone can watch — and the mislabels nobody catches poison training data, evals, and RL rewards. Give us a task description and video rollouts, and we return a success verdict with evidence. Teams use Instance to clean training data, run evals they can trust, and get reward signals for reinforcement learning without hand-labeling. We benchmark our verifier against human-labeled episodes across 8 benchmarks and 7 robot platforms, where it outperforms frontier VLMs in accuracy and latency. We're MIT computer scientists & friends since middle school, with backgrounds at NASA JPL, SpaceX, and AWS.
Ratings are live Elo scores from head-to-head crowd votes — opinion, not fact or financial advice. How it works →