Zero-HLM: a humanoid Unitee G1 trained only in simulation, from scenes filmed on the smartphone, does better than with real demonstrations

By rebuilding real scenes in simulation, Zero-HLM generates 4,800 physically validated demonstrations. The resulting policy succeeds 66.7% of tasks, compared with 46.7% with data collected on the real robot.

Teaching a humanoid to move while manipulating objects comes up against the cost of real demonstrations. Zero-HLM proposes to do without it: a simple smartphone capture is used to reconstruct the scene in simulation, with a Gaussian-based method that aligns the visual appearance and the physical properties of the environment.

Movements corrected by the simulation itself

The framework then jointly adjusts the reference movements of the whole body and the geometry of the interactions, based on the feedback from the simulated tests. It has drawn 4,800 physically validated demonstrations for three tasks performed by a Unitree G1, which are then transferred to the real robot without any adjustments.

Trained solely on this simulated data, the GR00T N1.7 model achieves an average success rate of 66.7% on complete tasks, compared to 46.7% when learning from real data. The gap widens on new configurations: the cumulative success rate increases from 23% for basic policies trained on real data to 69% with extended simulated training, without additional real demonstration.