Fast Motion Retargeting Pipeline for Scalable Humanoid Robot Learning

Key points

• While large-scale human motion capture datasets provide a rich source of demonstrations for humanoid robot learning, retargeting the diverse motions to different humanoid robots remains a nontrivial challenge.1
• Existing kinodynamic trajectory optimization pipelines are computationally expensive and often require manual adjustments per motion.1
• For motions requiring additional verification, a contact implicit iLQR controller is utilized to test the retargeted reference under full rigid body dynamics.1
• We evaluate retargeting quality by training motion tracking reinforcement learning policies using several state-of-the-art pipelines and comparing their tracking fidelity and success rates.1

Sources (1)

• [1]Fast Motion Retargeting Pipeline for Scalable Humanoid Robot Learning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 04:22 PM

  We present FRDV (Fast Retargeting with Dynamics Validation), a scalable framework consisting of rapid constrained kinematic retargeting and an additional dynamics validation pipeline.

  While large-scale human motion capture datasets provide a rich source of demonstrations for humanoid robot learning, retargeting the diverse motions to different humanoid robots remains a nontrivial challenge.

Extractive summary: sentences quoted from the sources.