hello3x3/zte_embodied_2026

ZTE Embodied 2026 Dataset (LeRobot v3.0)

Overview

This dataset is converted from the raw data of the 2026 16th ZTE Cup Global Elite Challenge - Algorithm Elite Challenge - Embodied Intelligence (Preliminary).

• Competition Link: ZTE Challenge - Embodied Intelligence
• Format: LeRobot v3.0
• Conversion Script: scripts/datasets/convert_zte_to_lerobot.py

Dataset Info

Property	Value
Codebase Version	v3.0
Robot Type	Dual-arm Dexterous (dual 7-DOF arms + dual 6-DOF hands)
Total Episodes	350
Total Frames	51,467
Total Tasks	15
FPS	30
Video Resolution	1280 x 720
Video Codec	av1 (libsvtav1)
Action Dimension	26
State Dimension	26

Splits

Split	Episodes	Index Range	Description
train	250	0:250	5 groups x 50 episodes
test	100	250:350	5 groups x 20 episodes

Features

observation.images.cam

• dtype: video
• shape: [720, 1280, 3]
• names: height, width, rgb
• info: codec=av1, pix_fmt=yuv420p, fps=30, channels=3

observation.state

Joint positions (26-dim float32) from joint.txt:

Index	Name	Description
0-6	left_arm_joint1~7	Left arm 7-DOF joint positions
7-13	right_arm_joint1~7	Right arm 7-DOF joint positions
14-19	left_thumb_0, left_thumb_1, left_index, left_middle, left_ring, left_pinky	Left hand 6-DOF finger positions
20-25	right_thumb_0, right_thumb_1, right_index, right_middle, right_ring, right_pinky	Right hand 6-DOF finger positions

action

Action commands (26-dim float32) from action.txt, same structure as observation.state.

Metadata Features

Feature	dtype	Description
timestamp	float32	Frame timestamp (frame_index / fps)
frame_index	int64	Index within episode
episode_index	int64	Episode index
index	int64	Global frame index
task_index	int64	Task description index

Tasks

Index	Description
0	Pick up the Cocacola and place it into the box.
1	Pick up the green tea and place it into the box.
2	Pick up the Sprite and place it into the box.
3	Pick up the mineral water from the box and place it on the table.
4	Pick up the Fanta from the box and place it on the table.
5	Pick up the Sprite from the box and place it on the table.
6	Pick up the apple from the basket and pass it to me.
7	Pick up the orange from the basket and place it on the table.
8	Pick up the banana from the basket and place it on the table.
9	Pick up the banana and place it into the box.
10	Pick up the apple and place it into the box.
11	Pick up the orange and place it into the box.
12	Pick up the doll and place it into the box.
13	Pick up the ball and place it into the box.
14	Pick up the Rubik's Cube and place it into the box.

Directory Structure

zte_embodied_2026/
├── data/
│   └── chunk-000/
│       └── file-000.parquet          # Frame-level data (action, state, metadata)
├── meta/
│   ├── info.json                     # Dataset metadata and feature definitions
│   ├── stats.json                    # Global statistics (mean, std, min, max)
│   ├── tasks.parquet                 # Task description index
│   └── episodes/
│       └── chunk-000/
│           └── file-000.parquet      # Per-episode metadata and statistics
├── videos/
│   └── observation.images.cam/
│       └── chunk-000/
│           ├── file-000.mp4          # Concatenated video segments
│           ├── ...
│           └── file-007.mp4
└── README.md

Usage

Load with LeRobotDataset

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("zte_embodied_2026", root="./datasets/zte_embodied_2026")
print(f"Episodes: {ds.num_episodes}, Frames: {ds.num_frames}")

# Access a sample
sample = ds[0]
# sample["observation.images.cam"]  -> torch.Tensor [3, 720, 1280]
# sample["observation.state"]       -> torch.Tensor [26]
# sample["action"]                  -> torch.Tensor [26]
# sample["task"]                    -> str

Load with UniVAM DataLoader

In the JSONL config file, add:

{"repo_id": "./datasets/zte_embodied_2026", "dataset": "zte"}

The CAMERA_KEYS and ACTION_KEYS mappings for "zte" are already configured in src/univam/utils/dataloaders/lerobot_.py.

Raw Data Format (Before Conversion)

The original data was organized as follows:

release/
├── train/          # 250 episodes (5 groups x 50)
│   ├── 1_1/
│   │   ├── action.txt        # CSV: 26 columns (14 arm + 12 finger joints)
│   │   ├── joint.txt         # CSV: same structure as action.txt
│   │   ├── instruction.txt   # Natural language task description
│   │   ├── instruction.pt    # Encoded instruction embedding [1, 50, 4096]
│   │   └── video.mp4         # 30 FPS, 1280x720
│   └── ...
├── test/           # 100 episodes (5 groups x 20), 16 frames each
└── sample_result/  # Example submission format

Key differences from the converted format:

• instruction.pt (embedding tensor) is not included in the lerobot dataset
• joint.txt maps to observation.state
• action.txt maps to action
• video.mp4 frames are re-encoded as av1 and concatenated per chunk

Citation

If you use this dataset, please cite the original competition:

2026年第十六届中兴捧月全球精英挑战赛 - 神算师算法精英挑战赛 - 具身智能（初赛）
https://zte.uchallenge.cn/challenge/69638677c8440b6a6e14563c