PDF-HR: Pose Distance Fields for Humanoid Robots

TL;DR

PDF-HR introduces Pose Distance Fields for Humanoid Robots, a differentiable pose prior that encodes pose plausibility as a distance to a manifold of valid configurations. Grounded in Riemannian geometry, it learns a continuous function $f_\phi(\mathbf{q})$ over the humanoid's $29$-DoF configuration space and supports gradient-based projection toward the pose manifold. The prior serves as a plug-and-play reward, regularizer, or planner objective across motion tracking and retargeting, yielding improved sample efficiency and motion quality on tasks such as single-trajectory tracking, style-based mimicry, general tracking, and AMASS retargeting. While generally beneficial, limitations include occasional accuracy gaps, mode collapse risks in some styles, and runtime concerns, with future work aimed at richer kinematic signals and higher-quality pose data.

Abstract

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.

Figures (17)

• Figure 1: We present PDF-HR, which learns the manifold of plausible G1 poses as a zero-level set. Left: The $f_\phi$ is trained to approximate the unsigned pose distance field. Given a query pose $\mathbf{q}$, we compute its distance $d$ to the nearest dataset sample and optimize the network to regress this value. Right: The learned prior provides quantitative scores for arbitrary poses, where a larger predicted value $f_\phi(\mathbf{q})$ indicates a significant deviation from the manifold, corresponding to an unnatural pose. This learned prior effectively benefits downstream tasks such as motion tracking and motion retargeting.
• Figure 2: Visualization of joint orientation distributions of Sideflip at early training stage. The visualization maps the directional vectors of the robot's links onto unit spheres centered at their respective joints. The color gradient corresponds to the probability density of the visited states.
• Figure 3: Visual comparisons of motion tracking performance on dynamic skills between our method and ADD zhang2025physics. 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.
• Figure 4: Visual comparison of retargeted motions produced by GMR araujo2025retargeting (red) and our method (green). GMR artifacts are highlighted with red markers.
• Figure 5: Real-world deployment. (a) Deploying GMR-retargeted Seq. 4. Artifacts introduced by GMR lead to poor alignment with the reference motion and frequent self-collisions. (b) Deploying our retargeted Seq. 4. Our method produces smooth, physically plausible motions that closely match the reference sequence. (c) Deploying our retargeted Seq. 2 for highly dynamic skills. Our method robustly executes challenging motions, including dance and spinning jumps.