AIAny
Icon for item

HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark

Provides a large-scale benchmark and a human-aligned metric for humanoid whole-body motion tracking — about 153 hours of optical mocap from professional performers plus HumanScore trained on 12K human-labeled preference pairs to reveal contact, timing, and stability failures.

Introduction

Evaluation metrics for humanoid motion tracking routinely disagree with human perception: frame-wise kinematic errors average pose differences but often miss the physical artifacts viewers notice most, such as foot skating, mistimed touch-downs, and unstable support. This work argues that scalable, perception-aligned evaluation — not just more data or lower per-frame error — is the right lever for diagnosing and improving contact-rich, long-horizon tracking.

Key Findings
  • A new benchmark dataset: roughly 153 hours of high-fidelity optical motion trajectories recorded from 24 professional performers and organized into four motion families (Daily, Highly Dynamic, Interaction, Ground) to stress different failure modes.
  • Preference-aligned metric: HumanScore, a trajectory reward model trained on 12K human-labeled pairwise comparisons (24K trajectories), predicts human preferences better than standard kinematic metrics and highlights contact/stability artifacts.
  • Reveals real failure modes: across representative state-of-the-art trackers, HumanScore exposes contact timing errors, unstable support, and foot skating that per-frame joint-error metrics often miss.
  • Practical dataset packaging: motions retargeted to a benchmark humanoid and released with robot-space .npz references and JSON manifests to enable scalable evaluation and per-family diagnosis.
What Sets It Apart
  • Perception-first evaluation: prioritizes human preference signals via a learned reward rather than relying solely on geometric errors, so it better matches what observers consider realistic or acceptable.
  • Contact- and stability-focused organization: the four motion families are chosen to elicit specific failure regimes (steady-state drift, impacts and fast phase errors, hand-body coordination, low-posture multi-contact transitions), enabling targeted diagnosis.
  • Scale and diversity: substantially larger and more contact-rich than many prior test suites, making it useful for long-horizon controller and teleoperation benchmarks.
Who It's For and Tradeoffs

Great fit if you develop whole-body humanoid trackers, teleoperation controllers, or imitation systems and need evaluation that correlates with human judgments of realism and stability. Look elsewhere if your goal is low-level per-frame pose refinement only or if you cannot perform retargeting to the provided humanoid — the benchmark assumes retargeted references and evaluation in robot-space. HumanScore requires human-labeled comparisons for alignment, so reproducing or extending the metric needs additional annotation effort.

Brief Method Notes

HumanScore is trained from synchronized tracking rollouts where annotators choose which of two tracker outputs better matches the reference. The benchmark emphasizes per-family metrics to pinpoint failure modes and includes prepared clips and manifests to ease benchmarking across trackers.

Information

  • Websitearxiv.org
  • AuthorsDairu Liu, Zekun Qi, Jiayu Zeng, Ruixi Yu, Yu Guan, Yintianrun Zhang, Xuchuan Chen, Sikai Liang, Zekai Li, Chenghuai Lin
  • Published date2026/08/13

More Items

Externalizes persistent scene state into a camera-indexed world bank and designs a long-horizon teacher whose sparse-attention supervision is distilled into a three-step student, enabling responsive, low-latency interactive long-horizon video generation with bounded denoiser context.

Predicts future video frames conditioned on an observed frame, a language instruction, and a sequence of end-effector poses and gripper states for robot manipulation. Uses per-arm SE(3) geometric encoding (PRoPE-style), a lightweight depth branch, SAM3 masks with a frozen V-JEPA teacher, and distribution-matching distillation for efficient, consistent action-conditioned rollouts.

Builds an editable, persistent 3D world state to drive iterative previsualization for film, games, and design — enabling local edits and recombinations instead of one-shot video regeneration. Uses separate stages for state construction, state evolution, and state access, with render-feedback camera refinement.