Zero-HLM Framework Enables Efficient Sim-to-Real Training for Humanoid Robots
机器人仿真数据流水线来了，德塔智能联合发布Zero-HLM框架
The Zero-HLM framework, developed by Dianyi Yang, Mengya Liu, and Siyuan Huang's team at Deta Intelligence, combines real-world phone-captured data with physics-based simulation to generate large-scale training data for humanoid robots. Using only 45 simulated teleoperation demonstrations, the team generated 4,800 physically validated training samples, achieving a 1.8x speedup in data collection. When deployed on the Yuji G1 humanoid robot, the pure simulation-trained policy achieved an average success rate of 66.7% across three long-horizon tasks, with 100% success in object-target combination tests.
The Zero-HLM framework demonstrates a novel approach to sim-to-real transfer by integrating real-world phone-captured data with physics-based simulation, enabling efficient generation of training data for humanoid robots. The results show promising performance in simulated-to-real transfer, though real-world deployment and generalization remain areas for further exploration.
Source: Chinese robotics — Hardware and sensing · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
data_collection_speed_increase
1.8
View original evidence
结果显示， 仅用45条仿真遥操作示范数据 ，该团队就通过Zero-HLM生成了 4800条 经过物理验证的训练数据，数据采集 速度提升1.8倍 。
Open source S7Reported success rate
66.7%
data_collection_speed_increase
1.8
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
结果显示， 仅用45条仿真遥操作示范数据 ，该团队就通过Zero-HLM生成了 4800条 经过物理验证的训练数据，数据采集 速度提升1.8倍 。
Open source S7
并且基于 GR00T N1.7 模型训练的纯仿真策略，在三项长时序任务中的 平均 成功率达到66.7% 。
Open source S8
通过上述流程，该团队利用45条仿真遥操作示范，生成了4800条经过物理验证的成功轨迹，覆盖三项长时序任务及不同场景配置。
Open source S23
研究团队针对每项任务采集15条仿真遥操作示范，并将动作拆分为行走、操作和腿部交互等片段。
Open source S17
与真实世界遥操作采集相比，整体数据采集速度平均提高1.8倍。
Open source S24
Source：Chinese robotics — Hardware and sensing · news.qq.com