Real2Gym Turns Videos Into Interactive Robot Gyms, Beating GPT-6 Direct Mode by 33% on Real Robots
Shanghai-AI-Laboratory · hf · 2026-09-30
Shanghai AI Laboratory's Real2Gym is an agentic Real2Sim2Real framework that converts human and robot manipulation demos into interactive simulation gyms and transfers learned skills to real robots. It reconstructs editable scenes, validates actions via physics execution, and distills successes and failures into reusable procedures and recovery strategies—without weight updates. It outperforms GPT-6 Astra Direct Mode by 16.7% in reconstruction success with 74.9% fewer tokens, and by 33.3% on real Franka execution across four tasks.
More from Embodied
- Agile Robots ships humanoid robots daily from German factory, partners with DeepMind — CyberRobooo · 2026-09-30
- CrossBFM Distills a Shared Behavior Space Across Humanoid Robots in Under One GPU-Hour — Tan-Dzung Do · 2026-09-30
- Survey Maps Four Families of In-Context Learning for Robots, Toward Physical Recursive Self-Improvement — Knowin · 2026-09-30
- RobotGym launches free public MCP for Physical AI simulation with hosted GPUs — rocky_mountain12 · 2026-09-30
- USC's CLAM learns robot policies from unlabeled videos, 2-3x success over baselines — ebiyik_ · 2026-09-30
- NVIDIA's Robo Olympics uses Codex and natural language to train robot skills in physics simulation — RevLebaredian · 2026-09-30