Reproduce “ImDy: Human Inverse Dynamics from Imitated Observations” on OpenResearch

About this paper

Xinpeng Liu, Junxuan Liang, Zili Lin, Haowen Hou, Yong-Lu Li, Cewu Lu

Researchers developed IMDY, a scalable method for generating a synthetic human inverse dynamics dataset through motion imitation within a physics simulator. This large dataset is used to train IMDYS, a data-driven solver that accurately estimates full-body joint torques and ground reaction forces from kinematics, improving prediction accuracy on real-world data and outperforming baselines even in zero-shot scenarios.

Key insights

• Motion imitation learning, by successfully replicating human motion in a physics simulator, implicitly learns and provides access to the underlying dynamic signals (torques, GRFs) necessary for movement execution.
• Large-scale synthetic data generated via imitation can effectively overcome the data scarcity problem in inverse dynamics, enabling the training of more robust and generalizable deep learning models.
• A pre-training strategy using synthetic dynamics data significantly improves the performance and convergence speed of inverse dynamics models when fine-tuned on limited real-world datasets, demonstrating effective sim2real transfer.

Summary by alphaXiv.