Source-linked AI summary
UniTracker: Learning Universal Whole-Body Motion Tracker for Humanoid Robots
Kangning Yin, Weishuai Zeng, Ke Fan, Minyue Dai, Zirui Wang, Qiang Zhang, Zheng Tian, Jingbo Wang, Jiangmiao Pang, Weinan Zhang
TL;DR
Humanoid robots need expressive, robust whole-body control for real-world environments, but partial observations and unseen motions challenge existing policies. UniTracker addresses these challenges with a three-stage teacher–CVAE student–adaptation framework, successfully tracking over 8,100 motion sequences on a Unitree G1 while improving accuracy, generalization, and robustness.
Problem
Whole-body control for humanoid robots must remain expressive and robust in real-world settings despite deployment under partial observations and challenges from unseen motions.
Method
UniTracker combines a privileged teacher, a CVAE-based universal student that incorporates global context, and fast residual adaptation for difficult motion sequences.
Results
UniTracker successfully tracks over 8,100 motion sequences with one policy on a 29-DoF Unitree G1, outperforming teacher-student baselines and prior methods in accuracy, generalization, and robustness.
Takeaways & Limitations
Structured latent motion modeling supports generalization to unseen behaviors and stronger robustness under observation noise, while residual adaptation extends tracking to challenging motions.
Abstract
from arXiv · showhide
Achieving expressive and generalizable whole-body motion control is essential for deploying humanoid robots in real-world environments. In this work, we propose UniTracker, a three-stage training framework that enables robust and scalable motion tracking across a wide range of human behaviors. In the first stage, we train a teacher policy with privileged observations to generate high-quality actions. In the second stage, we introduce a Conditional Variational Autoencoder (CVAE) to model a universal student policy that can be deployed directly on real hardware. The CVAE structure allows the policy to learn a global latent representation of motion, enhancing generalization to unseen behaviors and addressing the limitations of standard MLP-based policies under partial observations. Unlike pure MLPs that suffer from drift in global attributes like orientation, our CVAE-student policy incorporates global intent during training by aligning a partial-observation prior to the full-observation encoder. In the third stage, we introduce a fast adaptation module that fine-tunes the universal policy on harder motion sequences that are difficult to track directly. This adaptation can be performed both for single sequences and in batch mode, further showcasing the flexibility and scalability of our approach. We evaluate UniTracker in both simulation and real-world settings using a Unitree G1 humanoid, demonstrating strong performance in motion diversity, tracking accuracy, and deployment robustness.
I. INTRODUCTION
UniTracker addresses the challenge of expressive, generalizable whole-body motion tracking for humanoid robots under partial observations. Its three-stage framework combines teacher supervision, CVAE-based student modeling, and fast adaptation, with evaluations showing broad tracking capability and improved robustness.
- Whole-body motion tracking coordinates multiple joints and limbs to support stable, expressive, and adaptable humanoid behavior in real-world environments.
- Existing teacher-student policies can lose reference-motion diversity, generalize poorly to unseen sequences, and exhibit orientation drift under partial observations.
- UniTracker uses a CVAE student conditioned on future motion references to model diverse behaviors and improve generalization across unseen motion patterns.
- The CVAE aligns a globally informed encoder with a partial-observation prior through KL divergence, incorporating global context during training for more coherent deployment behavior.
- Fast adaptation fine-tunes the universal policy for rare, highly dynamic, or out-of-distribution motions through single-sequence or batch-mode specialization.
- UniTracker tracks over 8,100 diverse motion sequences on a 29-DoF Unitree G1 using one policy and outperforms teacher-student baselines in accuracy, robustness, and generalization.
II. METHOD
The method formulates humanoid whole-body motion tracking as goal-conditioned reinforcement learning. Proprioceptive state and whole-body target state define the task, while dense rewards and position-controlled actions support PPO-based optimization.
- Whole-body motion tracking is formulated as a goal-conditioned reinforcement-learning task in which a policy tracks reference motions.
- The state combines robot proprioception with a goal specifying target states for all body parts.
- A reward based on proprioception and the goal state provides dense signals for policy optimization.
- The 29-DoF Unitree G1 uses fixed wrist joints, reducing the action space to 23 dimensions, with target joint positions executed by a PD controller.
- Proximal Policy Optimization is used to optimize the tracking policy.
B. Humanoid Motion Dataset Curation
UniTracker builds its motion-training pipeline around filtered human motion data, retargeting, oracle-state policy training, and reward mechanisms designed for stable tracking.
- Dataset construction: 8179 human motions remain after excluding human-object interactions, removing clips shorter than 10 frames, and applying PHC filtering.The motions are represented with SMPL parameters.
- Motion retargeting: Two-stage retargeting aligns SMPL motions with humanoid robots by optimizing robot shape and then minimizing selected-link distances over motion sequences.The procedure uses 16 corresponding body links and optimizes root translation, orientation, and joint positions.
- Reward design: The reward combines motion-tracking rewards, regularization, and penalties, with curriculum learning gradually introducing the latter terms.This weighting is intended to keep optimization focused on tracking while encouraging reasonable behavior.
- Training stabilization: Training terminates episodes when projected gravity exceeds 0.8 on the x or y axis or average link distance exceeds 0.5m.These conditions remove data from falls and severely diverged tracking states.
- Oracle policy training: Reference State Initialization randomly samples reference-motion starting points so the policy can learn multiple motion phases in parallel.The robot state is initialized from the corresponding reference pose, including root and joint kinematics.
D. Hierarchical Controller via Online Distillation
The deployable student uses partial robot observations and future references to infer a latent motion representation, then decodes actions through online distillation from the oracle policy.
- Deployable state space: The deployable proprioceptive input stacks joint positions, joint velocities, root angular velocity, gravity, and previous actions over 25 past steps.These histories provide the partial observations available during real-world deployment.
- Deployable state space: The deployable goal state stacks reference and current motion quantities over five future steps, with coordinate-frame rotations defined for different terms.The representation includes root orientations, velocities, rigid-body positions, and root positions.
- CVAE student: A CVAE models motion diversity by using a privileged-observation encoder, a deployable-state prior, and a decoder that produces actions.The latent representation addresses ambiguity caused by partial deployable observations.
- Online distillation: The training objective combines oracle-action imitation with a KL-divergence term aligning the prior and encoder distributions.The prior, encoder, and decoder are modeled as diagonal Gaussian distributions, with the encoder represented as a residual to the prior.
E. Fast Adaption on top of a Universal Policy
The fast adaptation stage targets rare, highly dynamic motions beyond the training distribution by adding a lightweight residual decoder to the universal tracker. It freezes the universal policy and fine-tunes only the residual correction for rapid motion-specific specialization.
- Rare, highly dynamic motions outside the training distribution motivate the fast adaptation stage.
- A lightweight residual decoder is appended to the universal tracker to adapt it to challenging motions.
- Only the residual decoder is fine-tuned with reinforcement learning while the universal policy parameters remain frozen.
- The deployed action combines the frozen universal policy action with the residual correction.
A. Experiment Setup
Experiments span simulation and real-world environments, with simulation training using Isaac Gym, domain randomization, and AMASS motions filtered by PHC. Evaluation includes sim-to-sim transfer to MuJoCo and four motion-tracking metrics.
- UniTracker is evaluated in simulated and real-world environments, with simulation training using 8192 parallel Isaac Gym environments.
- The simulation pipeline uses domain randomization and motion data derived from AMASS and filtered by PHC.
- Trained policies are transferred from Isaac Gym to MuJoCo for sim-to-sim validation alongside ablation and comparative studies.
- Performance is assessed with Success Rate, Mean Per Keypoint Position Error, Velocity Distance, and Acceleration Distance.
B. Baselines
UniTracker is compared with internal universal-policy and DAgger teacher-student baselines, as well as OmniH2O and Exbody2. It consistently outperforms these baselines and prior methods across all evaluation metrics.
- The study compares UniTracker with a universal policy trained from scratch and a DAgger student without CVAE modeling.
- OmniH2O and Exbody2 are included as recent state-of-the-art humanoid motion-tracking methods.
- UniTracker consistently outperforms all baselines and prior methods across every evaluation metric.
C. Ablation Studies
Ablations examine architecture, latent modeling, noise robustness, global consistency, downstream applications, and key CVAE hyperparameters. The reported results favor direct latent guidance without reference-input dominance, with stronger generalization and robustness for UniTracker.
- Architecture: Directly providing reference motion to the actor causes latent guidance to vanish, making behavior resemble a standard DAgger setup.
- Generalization: UniTracker successfully tracks an unseen crescent kick, whereas the variant without CVAE modeling fails to reproduce it effectively.
- Robustness: Under increasing observation noise, UniTracker outperforms the non-CVAE variant in Success Rate and Mean Per Keypoint Position Error.
- Robustness: UniTracker degrades more slowly as observation noise increases, indicating stronger robustness to observation perturbations.
- Global consistency: With CVAE modeling, the robot follows a straight walking trajectory, while the non-CVAE framework gradually deviates from the reference path.
- Hyperparameters: A latent dimension of 64, a five-frame future window, and a KL-loss weight of 0.1 provide the best reported hyperparameter settings.
D. Downstream Applications
UniTracker is evaluated on text-conditioned motion generation, video-based motion estimation, and fast adaptation, demonstrating tracking and rapid convergence across challenging settings.
- Text-conditioned SMPL motions are retargeted to the G1, and UniTracker accurately tracks them in both MuJoCo simulation and real-world experiments.
- Fast adaptation converges substantially faster than training from scratch on a challenging motion and the AMASS test dataset.The comparison uses identical observations and reward functions for both approaches.
- The fast-adaptation comparison evaluates convergence using episode length and cumulative reward.
IV. RELATED WORK
Prior humanoid whole-body control research spans optimization and reinforcement-learning approaches using teleoperation, motion datasets, video, and sparse task commands. UniTracker positions its three-stage framework as a unified, scalable alternative for motion tracking.
- Traditional humanoid whole-body controllers explicitly model the robot and environment before performing real-time optimization.These approaches are described as often struggling to adapt to environments.
- Reinforcement-learning controllers are categorized by control-signal sources including teleoperation, offline motion datasets, video-based estimation, and task commands.
- Offline motion methods use pre-collected sequences for imitation, while Exbody and Exbody2 decouple upper- and lower-body motions to balance stability and expressiveness.
- UniTracker combines a privileged teacher, CVAE-based student, and residual adaptation module in a unified and scalable whole-body tracking framework.The framework targets partial-observation deployment and adaptation to highly dynamic or out-of-distribution motions.