A preprint from Hamburg, Hong Kong and Munich teams describes 'action-conditioned latent foresight' for humanoid robots that must walk and handle objects at once. A generator proposes several command sequences, a model predicts how close each would bring the scene to a goal taken from demonstrations, and the robot runs the first commands of the best one and replans.
On ten simulated tasks it reports 89.7% average success. On seven tasks with a real G1 humanoid robot it reports 85.7% under standard conditions and 65.7% with unseen object layouts, versus 70.0% and 48.6% for GR00T N1.7, a baseline the authors trained and tested themselves.
Results are the authors' own, on seven tasks, and the paper has not been peer reviewed.
Limits
Preprint; author-run comparison on seven real-robot tasks; not peer reviewed.