Learning human motion tracking policies through reinforcement learning (RL) has emerged as a core capability that underpins a majority of recent progress in humanoid whole-body control. However, achieving generalization to unseen motions at high fidelity requires jointly scaling training to thousands of GPU and dataset hours. This work hypothesizes that motion tracking is inherently compute-bound because training on a diverse set of motions introduces learning interference, similar to multi-task RL. To tackle this, we propose Quantile Tracking Optimization (QTOP), which improves sample efficiency by introducing a distributional advantage estimate conditioned on state and reference pairs. QTOP outperforms on and off-policy RL baselines both in simulation and real-world tracking scenarios on multiple datasets by up to 10%. Furthermore, we establish a scaling law for QTOP and find that, aside from scaling predictably, dataset size has minimal impact on performance for moderately sized GPU hour regimes.
Figures & tables
Fig. 2: Per-motion actor and critic gradient cosine similarities during training. At each point, gradients are computed for individual reference motions at a common parameter snapshot. Shaded bands denote the 5th–95th percentile over motion pairs, and dotted lines denote the median. Critic gradients exhibit substantially greater variation, including a stronger negative tail, indicating more severe disagreement in value learning.
Method
MPKPE (mm) ↓
R-MPKPE (mm) ↓
Global EE Pos. (cm) ↓
EE Ori. (degrees) ↓
MOSAIC (optical mocap, 3.1h)
QTOP ( α=0.5 )
41.84 ±0.90
38.12 ±0.90
36.66 ±1.82
11.44 ±0.00
QTOP ( α=0 )
43.17 ±0.22
39.53 ±0.27
32.71 ±0.87
13.37 ±1.44
PPO (Muon)
44.99 ±0.14
41.70 ±0.08
35.75 ±0.87
11.86 ±0.49
PPO (Adam)
45.68 ±0.56
41.96 ±0.55
41.88 ±0.19
13.12 ±0.11
SAC
54.56 ±0.89
51.48 ±0.73
77.90 ±8.47
14.64 ±0.39
TABLE I: Algorithm comparison on two datasets. Position errors are reported in millimeters or centimeters and orientation errors in degrees. Lower is better for all metrics.
Fig. 3: Left: Fitted scaling law surface to log-spaced compute points for three different dataset sizes. Right: R-MPKPE difference heatmap of doubling compute versus doubling dataset hour budgets. For reasonable compute budgets up to 700 GPU hours, dataset size has a marginal impact on performance.
Fig. 4: Extrapolating scaling law predictions to each dataset size yields reliable predictions of tracking error. Bottom right: Predicted versus observed R-MPKPE for each held-out dataset size when fit to the remaining two.
System
Intervention
C (GPUh)
D (hours)
Error
Gain/doubling
Ours
Compute
13.7–1303
25
.055 → .035
5.59%
Ours
Data
1303
5–25
.037 → .035
3.31%
SONIC
Compute
2k–21k
556
.027 → .024
3.10%
SONIC
Data
21k
22–556
.024 → .024
0.53%
TABLE II: Relative error reduction per budget doubling when scaling compute or data. Frames are converted to hours at 50 Hz.
Fig. 5: QTOP balances policy updates across motions. Using identical rollouts and frozen policy checkpoints, we compare the across-motion coefficient of variation of advantage magnitudes and actor-gradient norms.
Fig. 6: Real world tracking results for converged PPO and QTOP checkpoints in terms of R-MPKPE and joint MAE. QTOP outperforms PPO at the same compute and dataset budgets in unseen motion replay.
Fig. 7: QTOP faithfully replays challenging out-of-distribution motions such as martial arts, whereas PPO relies on multiple stabilization steps.
High-fidelity motion tracking serves as the ultimate litmus test for generalizable, human-level motor skills. However, current policies often hit a "generality barrier": as motion libraries scale in diversity, tracking fidelity inevitably collapses - especially for real-world deployment of high-dynamic motions. We identify this failure as the result of two compounding factors: the learning bottleneck in scaling multi-motion optimization and the physical executability constraints that arise in real-world actuation. To overcome these challenges, we introduce OmniXtreme, a scalable framework that decouples general motor skill learning from sim-to-real physical skill refinement. Our approach uses a flow-matching policy with high-capacity architectures to scale representation capacity without interference-intensive multi-motion RL optimization, followed by an actuation-aware refinement phase that ensures robust performance on physical hardware. Extensive experiments demonstrate that OmniXtreme maintains high-fidelity tracking across diverse, high-difficulty datasets. On real robots, the unified policy successfully executes multiple extreme motions, effectively breaking the long-standing fidelity-scalability trade-off in high-dynamic humanoid control.
Yunshen Wang, Shaohang Zhu, Peiyuan Zhi +7
State Key Laboratory of General Artificial Intelligence, Beijing Institute for General Artificial Intelligence (BIGAI) · Joint Laboratory of Embodied AI and Humanoid Robots, BIGAI & Unitree Robotics · Shanghai Jiao Tong University +4
Humanoid motion tracking is central to teleoperation and whole-body imitation, yet evaluation often disagrees with what people perceive in videos. Kinematic errors average per-frame pose differences but miss the physical artifacts that matter most, particularly unstable support and incorrect contacts such as foot skating and mistimed touch-downs. Meanwhile, widely used test suites are small and lack the diversity needed to stress contact-rich, long-horizon behaviors. We introduce HumanTracker to make humanoid tracking evaluation both perceptually aligned and scalable. The HumanTracker benchmark contains approximately 153 hours of optical motion trajectories from multiple professional performers, organized into four motion families with text labels for fine-grained diagnosis. We further propose HumanScore, a preference-aligned metric trained on 12K motion pairs containing 24K motions. Across representative state-of-the-art trackers, HumanScore better predicts human preferences and reveals contact and stability failures that kinematic metrics often miss.
Dairu Liu, Zekun Qi, Jiayu Zeng +11
1Nankai University · 3Galbot · 2Tsinghua University +3
We introduce Humanoid-GPT, a GPT-style Transformer with causal attention trained on a billion-scale motion corpus for whole-body control. Unlike prior shallow MLP trackers constrained by scarce data and an agility-generalization trade-off, Humanoid-GPT is pre-trained on a 2B-frame retargeted corpus that unifies all major mocap datasets with large-scale in-house recordings. Scaling both data and model capacity yields a single generative Transformer that tracks highly dynamic behaviors while achieving unprecedented zero-shot generalization to unseen motions and control tasks. Extensive experiments and scaling analyses show that our model establishes a new performance frontier, demonstrating robust zero-shot generalization to unseen tasks while simultaneously tracking highly dynamic and complex motions.
Zekun Qi, Xuchuan Chen, Dairu Liu +10
Tsinghua University · Galbot Inc. · Shanghai Jiao Tong University +2