HumanTracker

HumanTracker is a humanoid motion tracking benchmark for evaluating contact-rich, long-horizon whole-body tracking. It is designed to diagnose failures that are often missed by frame-wise kinematic metrics, such as unstable support, contact timing errors, and foot skating.

This repository provides HumanTracker motion clips prepared for public research use. The release preserves the four motion categories and train/test structure while using anonymized clip filenames of the form <category>_<index>.npz.

Dataset Summary

The dataset includes robot-space tracking references stored as NumPy .npz files, plus JSON manifests for train and test splits. Each manifest entry points to one motion clip and records its category and frame count.

The four categories are:

• Agile: highly dynamic movements such as jumps, kicks, acrobatics, and fast footwork.
• Daily: routine daily motions such as walking, turning, gestures, and steady locomotion.
• Ground: low-posture and multi-contact transitions such as kneeling, sitting, rolling, and recovery motions.
• Interaction: human/object interaction motions requiring coordinated end-effector timing and whole-body stabilization.

Directory Structure

.
+-- Agile/
|   +-- Agile_1.npz
|   +-- ...
+-- Daily/
|   +-- Daily_1.npz
|   +-- ...
+-- Ground/
|   +-- Ground_1.npz
|   +-- ...
+-- Interaction/
|   +-- Interaction_1.npz
|   +-- ...
+-- PreferencePair/
|   +-- hf_records_idx_000000-000199.parquet
|   +-- ...
|   +-- hf_records.index.json
+-- train.json
+-- test.json

Example manifest row:

{
  "path": "Daily/Daily_1.npz",
  "category": "Daily",
  "frames": 1234
}

Data Format

Each .npz file contains time-series arrays for humanoid motion tracking. The keys can vary by clip, and commonly include:

• qpos: generalized coordinates / robot-space reference trajectory.
• qvel: generalized velocities.
• joint_names, jnt_type, njnt: joint metadata when available.
• frequency: capture or retargeted sequence frequency when available.
• split_points: segment boundary metadata when available.
• kpt_npose, kpt_cvel, navi_pose, navi_vel, foot_contact: additional keypoint, navigation, and contact signals for clips where these annotations are available.

Users should inspect np.load(path).files for the exact keys available in a given clip.

Loading Example

import json
import os
import numpy as np

root = "/path/to/HumanTracker"

with open(os.path.join(root, "train.json"), "r", encoding="utf-8") as f:
    train_manifest = json.load(f)

sample = train_manifest[0]
clip = np.load(os.path.join(root, sample["path"]), allow_pickle=False)

qpos = clip["qpos"]
qvel = clip["qvel"]

print(sample["category"], sample["frames"], qpos.shape, qvel.shape)
print("available keys:", clip.files)

The JSON manifests can also be loaded with Hugging Face Datasets:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={"train": "train.json", "test": "test.json"},
)

The loaded Hugging Face rows contain manifest metadata. Use the path field to load the corresponding .npz array file.

PreferencePair

HumanTracker also includes a 12K pairwise human preference set for preference-aligned motion tracking evaluation. Each pair compares two synchronized humanoid tracking rollouts for the same reference segment and stores the human preference label.

The preference data is stored under PreferencePair/ in RobotVisAnalyse-compatible Parquet files. Each Parquet row contains the following top-level columns:

• record_id: anonymized record identifier.
• timestamp: release timestamp placeholder.
• winner: one of left, right, similar, or bad_traj.
• invalid: whether the pair is marked invalid.
• left_label and right_label: anonymized side identifiers.
• record_json: the full pair record, including left/right trajectory arrays, metadata, flags, and comparison settings.

The release includes four mirror variants for each annotated comparison:

• original pair
• left trajectory mirrored
• right trajectory mirrored
• both trajectories mirrored

This augmentation preserves the pairwise preference label while improving left/right symmetry coverage for reward-model training and evaluation.

The trajectory payloads follow the same field structure used by RobotVisAnalyse preference records, including robot state, reference state, action, contact, keypoint, navigation, and qpos signals where available. Original capture filenames, annotator names, source paths, and policy names are not included in the public PreferencePair metadata.

The side-level policy fields are intentionally left empty in the public preference records.

Intended Uses

This dataset is intended for:

• Research on humanoid motion tracking and whole-body imitation.
• Benchmark prototyping and data-loader development.
• Category-aware analysis of motion tracking performance across daily, agile, ground-level, and interaction motions.
• Reproducible examples for robot-space reference trajectory loading.

Out-of-Scope Uses

This dataset should not be used to identify, profile, or re-identify performers. It is not a dataset of RGB videos, audio, biometric identity labels, or human-subject identity annotations.

Privacy and Anonymization Notes

The release uses anonymized filenames of the form <category>_<index>.npz. Original capture filenames, performer names, and capture-session names are not included in the manifest paths.

Because the dataset contains human motion trajectories, users should still treat the data as human-subject motion data and follow the applicable license, consent, and institutional requirements for their use case.

Licensing

Use of this dataset is governed by the license terms included in this repository.

Citation

If you use this dataset, please cite the HumanTracker project:

@misc{liu2026humantracker,
  title        = {HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark},
  author       = {HumanTracker Team},
  year         = {2026},
  note         = {Project page}
}