Research map / 3D vision / Human pose estimation

PDF-HR: Pose Distance Fields for Humanoid Robots

NeurIPS 2026 · 3D vision › Human pose estimation · approach: Physics and sensor-based motion capture · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Learns a lightweight MLP pose distance field over retargeted Unitree G1 poses and uses it as a plug-and-play pose-prior reward for RL motion tracking/mimicry and as a regularizer for GMR retargeting.

Problem

Humanoid robot pose and motion priors are limited by scarce high-quality humanoid motion data, and existing HMR priors are not directly reusable for humanoid control, tracking, and retargeting.

Approach

Model the plausible G1 pose manifold implicitly as the zero level set of a continuous unsigned distance field, trained by regression to nearest-neighbor joint-space distances over a large retargeted corpus with on-manifold, near-manifold and interpolation samples; use its value as a differentiable reward shaping term or optimization regularizer.

Key results

• Single-trajectory tracking: samples to reach 80% success reduced from 336.5M to 166.2M on Sideflip and ADD failed on Backflip and Double Kong while PDF-HR reached 183.6M and 288.5M.
• Retargeting on DanceDB: the ADD policy tracked 101.6 frames on average with PDF-HR versus 95.7 for GMR, and body rotation error fell from 0.197 to 0.162 rad.
• BeyondMimic real-robot deployment succeeded on 9/10 sequences with PDF-HR-retargeted data versus 6/10 with GMR.

Surprising findings

• Final tracking accuracy is occasionally lower than the baseline with the pose-prior reward, despite faster convergence.
• The retargeting regularizer makes the current implementation slower than GMR and can produce more severe jitter.

Limitations

• Final tracking accuracy occasionally lower than baseline
• Mode collapse in style-conditioned tracking similar to AMP
• Retargeting runtime slower than GMR with occasional jitter; training data may be physically infeasible

Experimental setup

Datasets: PHUMA LaFAN1 AMASS DanceDB

Benchmarks: single-trajectory motion tracking (9 skills) style-based motion mimicry (6 skills) general motion tracking (LaFAN1 subset) DanceDB retargeting BeyondMimic deployment

Metrics: sample efficiency (samples to 80% success) position error rotation error tracking duration undesired contacts

Baselines: ADD AMP GMR BeyondMimic TextOp

Builds on: Pose-NDF NRDF AMP ADD GMR

Closest papers by meaning

• NRDF: Neural Riemannian Distance Fields for Learning Articulated Pose Priors (CVPR 2024 · similarity 0.86)
• Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · similarity 0.85)
• Humanoid Horizon: Extending Task Horizon in Whole-Body Loco-Manipulation via Parallel Training, Dynamic Starting, and Reward Gating (NeurIPS 2026 · similarity 0.85)
• H-GAP: Humanoid Control with a Generalist Planner (ICLR 2024 · similarity 0.85)
• Normalizing Flows on the Product Space of SO(3) Manifolds for Probabilistic Human Pose Modeling (CVPR 2024 · similarity 0.84)
• ManiPose: Manifold-Constrained Multi-Hypothesis 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.84)