Event

GE-Sim 2.0 Closed-Loop Simulation Evaluation of the π0.5 Policy

1 reports · 1 sources

The Xbotics Embodied Intelligence Lab built a closed-loop simulation pipeline using the Zhiyuan GE-Sim 2.0 world model and the OpenPI π0.5 policy. In measured tests, the world model generated about 4.33 frames per second, the pure-policy closed loop ran at about 3.20 frames per second, and after integrating MiMo online scoring, the rate dropped to about 0.88 frames per second. The lab states that on a single A800 GPU, even with a 4-step distilled 2B world model, its generation rate is still about 3.7 times slower than the 16 fps real-world control rate. Due to data, compute, and time constraints, the experiment only performed reward-weighted behavior cloning and iterative fine-tuning, and did not conduct true reinforcement learning training. The π0.5 weights used in the test had previously been fine-tuned only on single-arm basic tasks and had never learned the dual-arm collaborative cardboard-box carrying task; MiMo scored based only on head-camera images from a limited number of frames, so the World Judge results for demo_001 and demo_002 differ from the retained full-process videos.