GE-Sim 2.0: A Roadmap Towards Comprehensive Closed-loop Video World Simulators for Robotic Manipulation

Introduction

Released Weights

• gesim_community_v2.0.1_g01op_distill_2B: The released GE-Sim 2.0 world simulator checkpoint. It is post-trained on Genie-01(G01) + OmniPicker(OP) data, building on internal pretrained weights, and distilled into a 2B-parameter few-step model for efficient closed-loop rollout.
• pi05_gesim_g01op_test: A pi05 policy checkpoint fine-tuned on demo tasks, provided for demo-task policy rollout and closed-loop evaluation with the released world simulator.

Model Highlights

• Action-conditioned multi-view world simulation: GE-Sim 2.0 generates robot rollouts from visual history and calibrated action trajectories, supporting head-view and wrist-view observations for embodied manipulation.
• Large-scale real-world training data: The simulator is trained with diverse robot episodes from teleoperation, real-world deployment, rich object interaction, successful executions, and failure trajectories, improving trajectory coverage and reducing hallucinated interactions.
• Proprioceptive state expert: A lightweight state expert decodes dual-arm joint angles and gripper states from video latents, providing downstream policy models with both visual and proprioceptive observations for next-chunk prediction.
• World Judge (coming soon): The paper introduces a VLM-based World Judge that scores generated rollouts against task instructions.
• Efficient rollout: The acceleration framework generates a 25-frame rollout in 2.3 seconds on a single H100 and supports up to 4x frame skipping for faster evaluation.
• WorldArena CVPR Challenge champion: GE-Sim 2.0 won the WorldArena CVPR Challenge with a 2B-parameter model, outperforming both robotic world models and closed-source general video generators reported in the paper.

Intended Use

Citation

License