Zero-HLM: 45 sim demos become 4,800, humanoid zero-shot sim2real hits 69% vs 23% baseline

机器之心 · wechat · 2026-11-13

BIGAI, BIT, Tsinghua and Deta Intelligence introduced Zero-HLM, a Real2Sim2Real pipeline that trains humanoid whole-body loco-manipulation policies entirely on simulation data with zero real-robot demos:

Phone-captured scenes are reconstructed with Gaussian splatting; GPT-Astra aligns geometry and configures collision/joints to build environments that look real and support physical interaction.
Starting from 45 teleoperated source demos, the system generates whole-body motions for new configurations, validates them with the deployment-matching whole-body controller, and co-adjusts scene and reference motions when execution drifts — only verified success trajectories enter training. 45 demos expand automatically to 4,800.
On a real Unitree G1 (GR00T N1.7 + SONIC controller) across three long-horizon tasks (basketing a banana, crouching to move a hat, two-arm basket carry + foot door-close): 800 sim demos per task reach 66.7% average success, 20 points above the 100-real-demo baseline.
Across five generalization settings (new starts, grasp/place positions, object-goal combos), 100 real tests yield 69% aggregate success vs 23% for the real-data baseline.
A GPT-Astra data-generation baseline largely fails (20%/0%/n.a.), showing physical execution feedback is essential.

The team also released Zero-HLM-Bench (3 tasks, 9 skill stages); under a unified protocol GR00T N1.7 scores 69.7% full-task success vs DiT4DiT 64.0% and π₀.₅ 59.0%. Visual domain gap remains the main sim2real bottleneck.

Original post →

Zero-HLM: 45 sim demos become 4,800, humanoid zero-shot sim2real hits 69% vs 23% baseline

More from Embodied