The paper presents LY-GWM-RoboCasa-Human300, a policy for RoboCasa kitchen tasks. At inference it samples four action chunks with a Human300-fine-tuned GR00T-N1.5 generator, uses an action-conditioned graph dynamics model to predict future latent features and robot state for each chunk, and executes the chunk with the highest predicted-state reward logit from a learned binary scorer. All three components are kept frozen during evaluation, and the system uses a 16-step prediction horizon.
In the authors' local evaluation on 50 RoboCasa 1.0.1 tasks in the pretrain kitchen split, with 50 episodes per task, the system recorded 207/2,500 successes (8.28%) versus 181/2,500 (7.24%) for the historical self-trained GR00T-N1.5 baseline, a difference of 26 successes or 1.04 percentage points. The atomic task group improved from 141/900 to 150/900. Local audits recorded 212,700 changed environment-input chunks, including 28,223 in atomic tasks.
The RoboCasa maintainers separately re-evaluated the submission and reported 354/1,500 successes (23.6%) across 50 tasks, with leaderboard rates of 48.3%, 16.3%, and 3.1% for Atomic-Seen, Composite-Seen, and Composite-Unseen. The authors report the local and official results separately and explicitly do not attribute the difference to a model improvement. The technical report describes the training stages, implemented inference mechanism, evaluation provenance, and public reproduction entrypoints. Since there is only one source, no cross-source agreement or disagreement is assessed.