Towards Adaptable Humanoid Control Via Adaptive Motion Tracking
higher jumping height, (b) extended movement in badminton hitting, (c) extended movement in tennis hitting, and (d) longer jumping distance.
AI summary
Problem
Existing motion-prior methods sacrifice imitation accuracy for adaptability, while motion-tracking methods require large datasets and test-time reference motions, leaving a gap in achieving both accurate imitation and broad adaptability from minimal data.
Approach
The method sparsifies a single reference motion into keyframes, edits them with minimal physical assumptions, and trains a two-stage reinforcement learning policy with phase and tracking adapters to dynamically warp time and adjust low-level actions for flexible adaptation.
Key results
- Significantly higher imitation accuracy and adaptability than baselines in simulation
- Successful real-world deployment on a Unitree G1 humanoid robot
- Robust performance across wide adaptation ranges in agile tasks
- Outperforms large-data baselines despite using only a single reference motion
Why it matters
It provides a practical, data-efficient pathway for humanoid robots to learn and adapt to diverse real-world tasks from minimal human demonstrations, bridging the gap between simulation and hardware deployment.
Abstract
Humanoid robots are envisioned to adapt demon- strated motions to diverse real-world conditions while accu- rately preserving motion patterns. Existing motion prior ap- proaches enable well adaptability with a few motions but often sacrifice imitation accuracy, whereas motion-tracking methods achieve accurate imitation yet require many training motions and a test-time target motion to adapt. To combine their strengths, we introduce AdaMimic, a novel motion tracking algorithm that enables adaptable humanoid control from a sin- gle reference motion. To reduce data dependence while ensuring adaptability, our method first creates an augmented dataset by sparsifying the single reference motion into keyframes and applying light editing with minimal physical assumptions. A policy is then initialized by tracking these sparse keyframes to generate dense intermediate motions, and adapters are subse- quently trained to adjust tracking speed and refine low-level actions based on the adjustment, enabling flexible time warping that further improves imitation accuracy and adaptability. We validate these significant improvements in our approach in both simulation and the real-world Unitree G1 humanoid robot in multiple tasks across a wide range of adaptation conditions (Fig. 1). Videos and code are available on our project page.