AI
Real2Gym: Building Gyms from Videos, Bringing Skills to Robots
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 03:34:12 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 03:18:25 UTC.
Briefings
Real2Gym is an agentic framework that converts human and robot video demonstrations into reusable robot skills via a Real2Sim2Real pipeline. It reconstructs editable scenes, validates actions through physics, and distills successful attempts into task procedures. Evaluations show it outperforms GPT-6 Astra Direct Mode by 16.7% in simulation success rate and 33.3% in physical robot execution success rate across four tasks on a real Franka robot.
Key points
- Real2Gym uses a Real2Sim2Real framework to reconstruct editable scenes and align objects and cameras with input videos.
- The framework validates demonstrated actions through native physics execution and generates executable code for manipulation stages.
Context and limitations
Evaluation metrics are limited to four tasks on a single real Franka robot, and the framework operates without updating underlying model weights.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 03:34:12 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidlh6y4ymfh6lsaiodn6e2n33hbxeae77e4e6kyi7t2vtxgcar43a