What robot AI needs isn't faster reactions. It's the ability to know when it has failed.
At ZenO, we ran about 3,000 experiments in simulation using the robot AI model pi0.5.
When we moved an object while the robot was reaching for it and told it right away, it made no difference. The model kept looking at the scene and caught up on its own.
When the robot dropped an object it was carrying, things were different. Once we told it exactly what happened ("You dropped it, start over"), the failure rate fell from 14.9% to 8.1%, nearly cut in half.
So the real challenge is knowing precisely what went wrong. That requires data showing what kind of failure occurred, what the situation looked like right after, and how a person recovered from it. Most robot data today is just clean footage of success.
ZenO is collecting diverse failure and recovery data to help build robot AI that can notice its own mistakes and fix them.