ImpedanceMimic learns per-joint impedance as action space, beating PHC and SimPoE on force-accurate motion imitation
tweetsatpreet · x · 2026-11-18
A new paper backed by Harvard's Kempner Institute, ImpedanceMimic, proposes learned impedance as the action space for force-accurate motion imitation. Fixed PD gains can track a motion yet miss its contact forces, so the method instead learns per-joint stiffness and damping, imitating walking, running and jumping across 129 human subjects.
Key results:
- Outperforms other action spaces (fixed-PD, torque, muscle) and SOTA methods (PHC, SimPoE) on both motion and force accuracy
- Transfers without retraining to 22 unseen people and a simulated Unitree G1 humanoid
- Interpretable learned parameters: stiffness scales with load, damping anticipates unloading
Paper and project page are publicly available.
More from Embodied
- Google building new robotics model Astra with 1M-token context, config leaks — lyraxana · 2026-11-18
- GPT-6.2 Astra tops robot-code test: 2.5B net solving 79% of LIBERO, Python code hits 96.8% — burny_tech · 2026-11-18
- Bittensor subnet's zels decision model 'Jev' integrated into Even Realities glasses — markjeffrey · 2026-11-18
- Zhejiang & Unitree open-source Movens-5M: 5M sim-verified robot motions from web video — 机器之心 · 2026-11-18
- Waymo Temporarily Pulls Its Robotaxis From All U.S. Freeways — mikeocool · 2026-11-18
- EXOperation: a wearable robot hand makes human demos for robot learning intuitive and replayable — ybisk · 2026-11-18