Event

Fysics AI and Fudan University Propose Unified World Action Model Fysiverse-UniWAM

1 reports · 1 sources

Fysics AI, together with the Cognition and Intelligent Technology Laboratory (CITLab) of Fudan University, proposed the unified world action model Fysiverse-UniWAM, which uses a single model to integrate three types of capabilities: autonomous navigation, mobile manipulation, and fixed-workstation dexterous manipulation. The model does not start from large-scale robot pretraining; instead, it starts from a pretrained video world model, jointly learns navigation and manipulation through a shared backbone network, and proposes manipulation anchor poses (MAP), building the MAP-Data dataset with more than 1.5 million trajectories and a cumulative total of approximately 7,500 hours. The team deployed the model on 4 robot platforms and conducted 24 real-robot tests covering the above three types of tasks; among them, UniWAM-Sim was trained only with MAP-Data, without adding real robot demonstration data, and was then directly deployed to two real robot platforms. The team states that the model ranked first in both navigation and mobile manipulation among the 24 real-robot tests.