Event

Zero-HLM: Humanoid Robots Trained on Simulated Data Outperform Real-Robot Data

1 reports · 1 sources

Dianyi Yang and other authors released Zero-HLM, a real-to-sim-to-real framework that converts smartphone-captured and simulation-source demonstrations into training data for humanoid mobile manipulation, without requiring real robot data. The authors claim to have generated 4,800 physically validated demonstrations and achieved zero-shot transfer on three tasks with the Unitree G1. The authors report that GR00T N1.7 trained in simulation achieved an average all-task success rate of 66.7%, versus 46.7% for real-data training; on generalization configurations, scaling up simulation training raised the aggregate success rate from 23% for the real-data baseline policy to 69%, without adding real demonstrations.