Towards Adaptable Humanoid Control via Adaptive Motion Tracking

TL;DR

AdaMimic is introduced, a novel motion tracking algorithm that enables adaptable humanoid control from a single reference motion that reduces data dependence while ensuring adaptability and validate significant improvements in the approach in both simulation and the real-world Unitree G1 humanoid robot in multiple tasks.

Abstract

Humanoid robots are envisioned to adapt demonstrated motions to diverse real-world conditions while accurately preserving motion patterns. Existing motion prior approaches 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 single 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 subsequently 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. Videos and code are available at https://taohuang13.github.io/adamimic.github.io/.

Figures (6)

• Figure 1: Method overview. (a) Human motions are reconstructed into SMPL motions via GVHMR Shen2024WorldGroundedHM and retargeted to the humanoid robot. Sparse keyframes are then selected and edited to form an augmented dataset for adaptive tracking. (b) Based on this dataset, AdaMimic first trains a tracking policy with fixed phase intervals and double critics for sparse global tracking and dense local tracking rewards, followed by phase and tracking adapters that enable effective time warping for improved tracking performance. (c) The resulting policies can be directly deployed on the real Unitree G1 robot.
• Figure 2: Motivations of keyframing and adapters. (Top) Global tracking errors at the landing moment of far jumping indicate that AdaMimic outperforms baselines—DeepMimic-Adapt, which augments motions with rules, and AdaMimic-Stage1, which tracks at a fixed speed—by incorporating sparse keyframes and adaptive time warping to improve physical plausibility. (Bottom) Local tracking errors further verify the improvements.
• Figure 3: (Top) Task visualization: Five representative motions used as input for humanoid retargeting. (Bottom) Keyframing and editing: Sparse keyframes are extracted from each motion, and few selected ones are further edited to enable adaptation. Colors denote adaptation difficulty relative to the original keyframe (gray): blue indicates easy adaptation cases, while red indicates hard adaptation cases. The adaptation ranges are presented in \ref{['tab:adaptaion_range']}.
• Figure 4: Baselines trained with large-scale motion data. Across five tasks, AdaMimic achieves better performance than UniTracker Yin2025UniTrackerLU and its variant adapted with the rule-based motions from DeepMimic-Adapt. Besides, the results indicate that additional motion data does not provide UniTracker with obvious gains in specialization over DeepMimic-Adapt.
• Figure 5: Effectiveness of adapters. (Top) The phase interval and tracking action adapt to different far-jump distances, where longer distances correspond to extended airtime and larger action compensation, while shorter distances lead to reduced adjustments. (Bottom) With adapters, the policy performs effective time warping, adjusting motion speed to improve adaptation.