| AGIBOT releases tau0-WM open source embodied world model with 17,800 hours of real-robot teleoperation data AGIBOT chief scientist Luo Jianlan, also an associate professor at the Shanghai Academy of Creative Intelligence, released tau0-WM, a 5-billion-parameter pre-trained embodied world model the team describes as the largest open source release of its kind globally, QbitAI reported. The pre-training corpus runs approximately 30,000 hours, of which 17,800 hours are real-robot teleoperation footage rather than synthetic or video data. The model layers test-time computation on the policy head, ranking candidate actions before execution and invoking a simulator for correction when candidate quality is insufficient. Reported average success rates exceed the pi0.5 and Fast-WAM baselines on four long-horizon manipulation tasks: Toolbox, School Bag, Badminton and Faucet. |