Learning World Models from Egocentric Video for Home Robots
In 30 seconds
A video world model trained on robot egocentric footage predicts outcomes of actions for evaluation and planning.
Research question
Can world models replace physical evaluation?
Problem
Evaluating home robots is slow and unsafe.
Previous approach
Physical rollouts or hand-built simulators.
New approach
Action-conditioned video prediction.
Results
Correlates with real success rates on household tasks.
Limitations
Hallucinated physics; single robot.
Industry impact
Positions video generation vendors as robotics infrastructure.