The research aims to address the data bottleneck in humanoid robot autonomy by developing a system that can learn from simulated environments rather than relying on human-labeled data.

The authors propose HALO, which uses a pipeline to generate vision-language-motion data across a thousand procedurally created homes, enabling the robot to learn a variety of physical skills.

The authors report that HALO successfully transfers its learned skills to real-world environments, achieving a high success rate across multiple trials without prior exposure to real homes.