Reaction
PRISM turns four real videos into 256 counterfactual variants for humanoid robot training
A PRISM team member says the project recovers robot-object trajectories and trains one policy that generalizes across objects, scales and layouts.
TLDR
A member of the Amazon FAR team introduced PRISM as a real-to-simulation-to-real approach to training humanoid robots to move and manipulate objects. They say it expands four real videos into 256 counterfactual variants, then recovers robot-object trajectories to train one policy. A separate post sharing the work argues that making many variants from a few good demonstrations helps address the challenge of finding suitable training videos at scale.
Combined views
17K
1 Source, first seen ago
