TrajWAM combines motion forecasting with one-step robot action prediction

Researchers report a method for dynamic robot manipulation that predicts actions in one inference step while retaining an explicit forecast of the target object's motion. Their arXiv preprint describes TrajWAM and reports evaluations in simulation and the real world. The system forecasts future object masks from a short history, then uses those forecasts to guide a video model's representation of future scenes. That representation conditions the robot's action predictions. A distillation process compresses the video and action models while preserving their connection. The authors report leading task performance, but provide no numerical speed or accuracy comparison in the abstract.