# RoboCasa

A robot policy no longer executes generated actions directly; it first predicts the future state of each candidate and runs the one with the highest predicted reward. LY-GWM-RoboCasa-Human300 combines a Human300-fine-tuned GR00T-N1.5 action generator, an action-conditioned graph dynamics model, and a learned binary state-reward scorer, sampling four action chunks at inference. On 50 RoboCasa 1.0.1 pretrain-kitchen tasks with 50 episodes each, local success rose from 7.24% in the historical self-trained baseline to 8.28%, 26 more successes.

A robot policy no longer executes generated actions directly; it first predicts the future state of each candidate and runs the one with the highest predicted reward. LY-GWM-RoboCasa-Human300 combines a Human300-fine-tuned GR00T-N1.5 action generator, an action-conditioned graph dynamics model, and a learned binary state-reward scorer, sampling four action chunks at inference. On 50 RoboCasa 1.0.1 pretrain-kitchen tasks with 50 episodes each, local success rose from 7.24% in the historical self-trained baseline to 8.28%, 26 more successes.