InterMimicGen: Scaling Humanoid Loco-Manipulation through a Self-Evolving Motion-Imitation Data Flywheel

Forum topic · 小凯 · 2026-10-07

Summary

InterMimicGen (arXiv:2610.06850) is a self-evolving motion-imitation framework that converts captured human-object interaction data into executable skills for humanoid robots with dexterous hands. The framework first consolidates mocap human-object interaction datasets and retargets them into humanoid reference motions while preserving whole-body coordination and hand-object relationships. It then trains a physics-based generalist tracker in simulation. Finally, a closed-loop data flywheel applies small, task-preserving modifications to where interactions occur and how bodies execute them, fine-tunes the tracker, and retains only variants that successfully complete tasks in simulation as seeds for the next round. Over successive rounds, these small modifications compound into broad coverage of sparse original demonstrations. Experiments demonstrate contact-preserving retargeting across robot configurations, wide tracking coverage with a single generalist policy, continual growth of executable motions across augmentation rounds, and successful real-robot transfer.

Paper Overview

Research area: Robotics
Authors: Yucheng Zhang, Sirui Xu, Jinhong Li, Liuyu Bian, Anatulya Nandi, Derek Zhang, Xiangchen Liu, Xueting Li, Umar Iqbal, Yu-Xiong Wang, Liang-Yan Gui
Published: 2026-10-05
arXiv: 2610.06850

Abstract (Translated)

Captured human-object interaction data provides a rich supervision signal for humanoid loco-manipulation, but such data is sparse, heterogeneous, and not directly executable by robots. The authors propose InterMimicGen, a self-evolving motion-imitation framework in which robot motion data and the tracking policy mutually improve each other.

The framework operates in three stages:

1. Data consolidation and retargeting: Human-object interaction mocap datasets are consolidated and retargeted into humanoid robot references, preserving whole-body coordination and dexterous hand-object relationships. 2. Physics-based generalist tracker: A generalist tracker is trained in simulation to execute these references on a humanoid robot equipped with dexterous hands. 3. Closed-loop data flywheel: Each augmentation round applies small, task-preserving modifications to where interactions occur and how the body executes them, fine-tunes the tracker, and keeps only simulation-verified variants that complete the task as seeds for the next round. Across more iterations, these small modifications compound into broader coverage of the sparse original demonstrations.

Key Results

Contact-preserving retargeting across robot configurations
Broad motion tracking coverage with a single generalist policy
Continual growth of executable motions across augmentation rounds
Successful transfer to real-robot hardware

Original Abstract

> InterMimicGen is a self-evolving motion-imitation framework where robot motion data and tracking policy improve each other. It consolidates human-object interaction datasets, retargets them to humanoid references, trains a physics-based generalist tracker, and closes a data flywheel through task-preserving augmentation rounds.

Paper: arxiv.org/abs/2610.06850

*Auto-collected on 2026-10-07*