Five Top Universities Launch First Benchmark for Robotic Three-View World Models

The TriWorldBench Challenge, jointly initiated by Peking University, Tsinghua University, Beihang University, Shanghai Jiao Tong University, and the University of Science and Technology of China, has released its first weekly leaderboard. It is the first evaluation benchmark for three-view world models in robotics, aiming to establish a more comprehensive evaluation system for embodied world models and push them from 'visual generation' to 'world understanding'. The benchmark evaluates models on multi-view consistency, task execution, and physical space understanding, and is open to global research teams. In the first week, models WoVR_Plus from CASIA-DRL, BetaBWM from TONGJI Spatial Intelligence Team, and Fysiverse-Video from Fysics AI took the top three spots. The benchmark focuses on head view, left-wrist view, and right-wrist view video generation and understanding. It uses 500 three-view synchronous episodes covering 50 robot manipulation tasks, and aggregates 19 evaluation signals across six dimensions into a total score called TWB-Score. The system also constructs STATE annotations from reference trajectories to evaluate motion and stillness. The leaderboard is updated regularly, and the organizers encourage global participation through the official website and GitHub repository.