TEXEDO Dataset

Test-Time Scaling for Controller-Aware Language-Conditioned Humanoid Motion Generation

TEXEDO is a text-motion dataset for the Unitree G1 humanoid, prepared from two sources: AMASS and CLAW. It keeps the MotionGPT-style split files while adding JSONL index files that are easier to load from the Hugging Face Hub. The dataset accompanies the TEXEDO paper, a text-to-motion pipeline that performs test-time scaling — sampling multiple candidate motions from a language prompt and selecting the best with controller-aware dynamic and semantic verifiers.

• 🌐 Project page: https://jianuocao.github.io/TEXEDO/
• 💻 Code: https://github.com/JianuoCao/TEXEDO
• 📄 Paper: https://arxiv.org/abs/2606.22998
• 📦 Checkpoints: https://huggingface.co/JianuoCao/TEXEDO-Checkpoint

Dataset Structure

TEXEDO_dataset/
  README.md
  train.txt
  val.txt
  test.txt
  data/
    train.jsonl
    validation.jsonl
    test.jsonl
    all.jsonl
  motions/
    amass/{id_prefix}/{id}.npy
    claw/{id_prefix}/{id}.npy
  texts/
    amass/{id_prefix}/{id}.txt
    claw/{id_prefix}/{id}.txt
  metadata/
    dataset_summary.json
  prepare_texedo_dataset.py

Each sample has one motion file and one text annotation file. Motion files are NumPy arrays with shape (num_frames, 36).

Motion Format

Each .npy motion file stores a float array with shape (T, 36), where T is the number of frames. The 36 dimensions are organized as:

Feature slice	Size	Description
motion[:, 0:3]	3	root position, (x, y, z)
motion[:, 3:7]	4	root quaternion, (w, x, y, z)
motion[:, 7:36]	29	joint positions / joint angles in the order below

Joint order for motion[:, 7:36]:

Joint index	Feature dim	Joint name
0	7	left_hip_pitch_joint
1	8	right_hip_pitch_joint
2	9	waist_yaw_joint
3	10	left_hip_roll_joint
4	11	right_hip_roll_joint
5	12	waist_roll_joint
6	13	left_hip_yaw_joint
7	14	right_hip_yaw_joint
8	15	waist_pitch_joint
9	16	left_knee_joint
10	17	right_knee_joint
11	18	left_shoulder_pitch_joint
12	19	right_shoulder_pitch_joint
13	20	left_ankle_pitch_joint
14	21	right_ankle_pitch_joint
15	22	left_shoulder_roll_joint
16	23	right_shoulder_roll_joint
17	24	left_ankle_roll_joint
18	25	right_ankle_roll_joint
19	26	left_shoulder_yaw_joint
20	27	right_shoulder_yaw_joint
21	28	left_elbow_joint
22	29	right_elbow_joint
23	30	left_wrist_roll_joint
24	31	right_wrist_roll_joint
25	32	left_wrist_pitch_joint
26	33	right_wrist_pitch_joint
27	34	left_wrist_yaw_joint
28	35	right_wrist_yaw_joint

Splits

Split	Samples
train	18,590
validation	2,324
test	2,325
total	23,239

Sources

Source	Samples
AMASS	9,245
CLAW	13,994

The original textseedo source label from the preparation metadata is normalized to claw in this release.

JSONL Fields

Each row in data/*.jsonl contains:

• id: six-digit sample id
• split: train, validation, or test
• source: amass or claw
• motion_path: relative path to the .npy motion file
• text_path: relative path to the original text annotation file
• num_frames: number of motion frames
• motion_dim: motion feature dimension, currently 36
• num_texts: number of captions in the text file
• captions: parsed caption entries with caption, tokens, start_time, and end_time

This release intentionally does not include raw_source or original_npz fields.

Loading

from datasets import load_dataset
from huggingface_hub import snapshot_download
import numpy as np
from pathlib import Path

repo_id = "JianuoCao/TEXEDO"
repo_root = Path(snapshot_download(repo_id, repo_type="dataset"))

ds = load_dataset(
    "json",
    data_files={
        "train": str(repo_root / "data/train.jsonl"),
        "validation": str(repo_root / "data/validation.jsonl"),
        "test": str(repo_root / "data/test.jsonl"),
    },
)

sample = ds["train"][0]
motion = np.load(repo_root / sample["motion_path"])
captions = sample["captions"]

Citation

@misc{cao2026texedotesttime,
  title={TEXEDO: Test-Time Scaling for Controller-Aware Language-Conditioned Humanoid Motion Generation},
  author={Jianuo Cao and Yuxin Chen and Yuzhen Song and Masayoshi Tomizuka and Chenran Li and Thomas Tian},
  year={2026},
  eprint={2606.22998},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2606.22998},
}

License & Terms

This release is distributed under the other license. The motions are derived from AMASS and CLAW; their original redistribution terms and licenses apply. Please review and comply with the source datasets' terms before use or redistribution.