HumanoidMimicGen: Data Generation for Loco-Manipulation via Whole-Body Planning

Researchers introduce HumanoidMimicGen, a method for automatically generating training data for humanoid robots performing complex locomotion and manipulation tasks. The approach enables imitation learning at scale without labor-intensive teleoperation, achieving 20% performance improvements over models trained solely on real-world demonstrations.

HumanoidMimicGen addresses a fundamental bottleneck in humanoid robotics: the scarcity of training data for loco-manipulation tasks. Humanoid robots present unique challenges compared to traditional manipulators because their action spaces encompass coordinated control of arms, legs, and torsos simultaneously. This complexity has historically limited data-generation approaches, making large-scale imitation learning impractical. The method tackles this by adapting source demonstrations to new states through contact-rich whole-body planning, effectively synthesizing diverse training scenarios automatically. This breakthrough matters because data generation remains one of the costliest aspects of deploying embodied AI systems. The introduction of a nine-task loco-manipulation benchmark provides a standardized evaluation framework, enabling systematic analysis of how data quantity and quality affect policy learning outcomes. The 20% performance improvement when combining synthetic and real data suggests that synthetic data meaningfully complements limited real-world demonstrations rather than replacing them. This finding has broader implications for robotics development timelines and costs. As humanoid robots transition from research environments to real-world deployment, the ability to generate diverse training data algorithmically accelerates progress. The work also demonstrates that whole-body visuomotor policies benefit from co-training approaches, informing future architecture designs. For the robotics industry, this research reduces barriers to entry for organizations developing humanoid systems, potentially accelerating commercialization of mobile manipulation platforms. The methodology may generalize to other high-dimensional embodied systems facing similar data scarcity challenges.

• →HumanoidMimicGen automatically generates large training datasets for humanoid robot loco-manipulation without labor-intensive teleoperation.
• →The method adapts source demonstrations across diverse scenes and layouts while maintaining stability and collision-free trajectories.
• →Policies trained on combined synthetic and real-world data outperform those trained on real data alone by 20%.
• →A new nine-task benchmark enables systematic evaluation of how data generation and policy learning decisions impact humanoid performance.
• →The approach addresses the fundamental challenge of high-dimensional action spaces in humanoid robots involving coordinated arm, leg, and torso control.