Event

Technical Report Released on Graph Dynamics Model Policy for Prediction-Guided Action Selection

1 reports · 1 sources

TianLiang Ma released a technical report introducing the LY-GWM-RoboCasa-Human300 policy: at inference time it samples 4 action chunks, uses a graph dynamics model to predict subsequent states, and then a binary state reward scorer selects the highest-scoring candidate for execution. The policy combines a GR00T-N1.5 action generator fine-tuned on Human300; all three components are frozen during evaluation, and the prediction horizon is 16 steps. Local evaluation was conducted on 50 pretrain kitchen tasks in RoboCasa 1.0.1, with 50 episodes per task, achieving 207/2,500 successes (8.28%), higher than the historical self-trained GR00T-N1.5 baseline of 181/2,500 (7.24%). The RoboCasa maintainers separately re-evaluated this submission and reported 354/1,500 (23.6%), with Atomic-Seen, Composite-Seen, and Composite-Unseen at 48.3%, 16.3%, and 3.1%, respectively. The report states that the two sets of results are presented separately and that their difference is not attributed to model improvements.