Pose and motion priors play a crucial role in humanoid robotics. Although such priors have been widely studied in human motion recovery (HMR) domain with a range of models, their adoption for humanoid robots remains limited, largely due to the scarcity of high-quality humanoid motion data. In this work, we introduce Pose Distance Fields for Humanoid Robots (PDF-HR), a lightweight prior that represents the robot pose distribution as a continuous and differentiable manifold. Given an arbitrary pose, PDF-HR predicts its distance to a large corpus of retargeted robot poses, yielding a smooth measure of pose plausibility that is well suited for optimization and control. PDF-HR can be integrated as a reward shaping term, a regularizer, or a standalone plausibility scorer across diverse pipelines. We evaluate PDF-HR on various humanoid tasks, including single-trajectory motion tracking, general motion tracking, style-based motion mimicry, and general motion retargeting. Experiments show that this plug-and-play prior consistently and substantially strengthens strong baselines. Code and models will be released.
As shown in pipeline, we learn the manifold of plausible G1 poses as a zero-level set. Left: The network $f_\phi$ approximates the unsigned pose distance field. For a query pose $\mathbf{q}$, it regresses the distance $d$ to the nearest dataset sample. Right: The prior quantifies plausibility, where higher $f_\phi(\mathbf{q})$ values indicate significant deviation from the manifold (unnatural poses). This metric effectively guides optimization in downstream motion tasks.
Applications:
Visual comparisons of motion tracking performance on dynamic skills between our method and ADD.
The number of training samples is annotated on the left of each strip. Our method successfully masters these complex skills with remarkably fewer samples, whereas the baseline frequently suffers from falls or collisions even after extensive training.
Trackability in simulation
We benchmark our method against GMR by evaluating tracking performance in simulation using ADD. While GMR-retargeted motions frequently exhibit implausible artifacts that cause tracking failures, our approach ensures high trackability and naturalness.
Deployability in real world
We further evaluate deployability using BeyondMimic. We select 10 motion sequences and compare policies trained on GMR-retargeted data versus our retargeted data. Our method successfully deploys on 9/10 sequences, whereas GMR succeeds on only 6/10.
Our method denoises unrealistic poses by projecting them onto the learned valid pose manifold. Drag the sliders below to observe the projection process.
Noisy
Projected
Noisy
Projected
We set the right-wrist end-effector to track an $\infty$-shaped trajectory. Compared to QP-based IK (blue, left), our PDF-HR-regularized HL-IK (gray, right) yields more natural and human-like whole-arm configurations.Drag the sliders below to observe the process.
@article{gu2026pdfhr,
title={PDF-HR: Pose Distance Fields for Humanoid Robots},
author={Yi Gu and Yukang Gao and Yangchen Zhou and Xingyu Chen and Yixiao Feng and Mingle Zhao and Yunyang Mo and Zhaorui Wang and Lixin Xu and Renjing Xu},
year={2026},
journal={arXiv preprint arXiv:2602.04851}
}