InterMimicGen: Self-Evolving Motion Imitation for Humanoids

Summary

Source: arXiv cs.RO, posted Oct. 6, 2026

The authors present InterMimicGen, a framework designed to bridge the gap between human motion capture data and humanoid robot execution. The core challenge addressed is that human-object interaction datasets are often sparse and heterogeneous, making them difficult to use directly for robot control. The approach begins by consolidating these datasets and retargeting them to humanoid configurations while preserving whole-body coordination and dexterous hand-object relationships. This creates a large, diverse collection of reference motions for loco-manipulation.

The system then trains a physics-based generalist tracker in simulation to execute these references on a humanoid with dexterous hands. A key innovation is the "data flywheel" mechanism: in each iteration, the system makes small, task-preserving changes to the interaction location and body execution. It fine-tunes the tracker on these variants and retains only those that successfully complete the task in simulation. These successful variants seed the next round, allowing the motion library to grow in coverage and diversity while maintaining task semantics. The authors report that this process leads to contact-preserving retargeting, broad tracking with a single policy, and successful transfer to real robots.

Why it matters

Humanoid loco-manipulation has traditionally relied on either limited, task-specific demonstrations or complex, hand-crafted control policies. Existing methods often struggle with the scale and diversity required for general-purpose manipulation. InterMimicGen offers a unified path from heterogeneous human demonstrations to a continually expanding motion resource. By leveraging a self-evolving data flywheel, it addresses the sparsity of human data and the rigidity of traditional imitation learning. This approach builds on physics-based simulation and retargeting techniques but departs from static dataset usage by actively generating and filtering new motion variants. It represents a significant step toward scalable, generalist humanoid control that can adapt to a wider range of real-world interactions.

Robot's take

The self-evolving data flywheel is a compelling strategy for overcoming the data scarcity problem in humanoid robotics. By iteratively refining and filtering motions based on simulated success, the system can expand its capabilities without requiring new human demonstrations for every variation. This is a strong advantage over static imitation learning. However, the reliance on simulation for the filtering step raises questions about the sim-to-real gap. While the authors report successful transfer to real robots, the extent to which the simulated success metrics correlate with real-world robustness is not fully detailed. The effectiveness of the "small, task-preserving changes" in capturing the full range of human dexterity also remains to be seen in more complex, unstructured environments. Overall, this is a promising framework, but its real-world scalability will depend on how well the simulated criteria predict physical performance.