Datasets:

RoboChallenge
/

icra_wbc

task_categories:
  - robotics
tags:
  - robot
  - manipulation
  - navigation
  - whole-body
  - teleoperation
  - bimanual
  - mobile-manipulation
  - G2
pretty_name: RoboChallenge ICRA Whole Body Control Dataset
size_categories:
  - 1K<n<10K

RoboChallenge ICRA Whole Body Control Dataset

This dataset contains 2,800 teleoperated episodes collected on the Agibot G2 humanoid robot. Each episode captures a full mobile manipulation sequence—navigating to a shelf, grasping drink(s), walking to a cart, and placing item(s) into it—recorded via multi-camera observations and full-body proprioception at 30 Hz.

Tasks

Task	Episodes	Description
`grasp_the_drink`	1,278	Navigate to the shelf, grasp one drink with one hand, carry it to the cart, and place it in.
`grasp_the_drink_with_both_hands`	1,522	Navigate to the shelf, grasp two drinks (one per hand), carry them to the cart, and place them in.

Drink Variants

Episodes cover 16 drink categories: ENONE Juice, Itoen Barley Tea, Schweppes Vitamin C, 7UP, Mirinda, Coconut Water, Frozen Pear Drink, JDB Herbal Tea, Pepsi, Huayang Soda, Weichuan Daily C, Aojiru, Tropicana, Minute Maid Juice, Coca Cola,ENONE Sparkling Water.

Directory Structure

{task_name}/
├── meta.json
└── data/
    └── episode_XXXXX/
        ├── meta/
        │   └── task_desc.json
        ├── states/
        │   └── aligned_joints.h5
        └── videos/
            ├── head_color.mp4
            ├── head_depth.mp4
            ├── head_stereo_left_color.mp4
            ├── head_stereo_right_color.mp4
            ├── hand_left_color.mp4
            └── hand_right_color.mp4

meta.json — Task-level metadata (sits alongside the data/ folder).
data/episode_XXXXX/ — One episode, zero-padded index (episode_00000, episode_00001, …).
- meta/task_desc.json — Episode-level task description, including natural-language prompts in English and Chinese, scoring rubric, and task tags.
- states/aligned_joints.h5 — Per-frame aligned proprioception, action commands, and camera timestamps (HDF5).
- videos/ — Six MP4 camera streams (see Cameras below).

Cameras

Each episode contains six video streams (fisheye cameras are excluded from this release):

File	View	Resolution	Codec	FPS
`head_color.mp4`	Head RGB	640 × 400	HEVC	30
`head_depth.mp4`	Head depth (encoded as video)	640 × 400	PNG	30
`head_stereo_left_color.mp4`	Head stereo left RGB	1920 × 1536	HEVC	30
`head_stereo_right_color.mp4`	Head stereo right RGB	1920 × 1536	HEVC	30
`hand_left_color.mp4`	Left wrist / hand RGB	1280 × 1056	HEVC	30
`hand_right_color.mp4`	Right wrist / hand RGB	1280 × 1056	HEVC	30

Synchronization

For each integer frame i in aligned_joints.h5, the group f"{i}"/timestamp/camera/<name> stores a uint64 array of length 1 representing the sensor timestamp. Use these timestamps to align decoded video frames to HDF5 rows. Available timestamp keys: hand_left_color, hand_right_color, head_color, head_depth, head_stereo_left, head_stereo_right.

`aligned_joints.h5` — Proprioception & Actions

Top Level

The HDF5 root contains one child group per aligned sample, named by string keys "0", "1", …, "N-1". Sort by integer value (not lexicographic order) when iterating.

Per-Frame Layout (`{i}/`)

Entry	Type	Description
`main_timestamp`	`uint64` scalar	Primary timeline stamp for the row
`action/`	Group	Commanded / policy-side targets
`state/`	Group	Measured robot state
`timestamp/`	Group	Per-sensor capture times (`timestamp/camera/…`)

`action/`

Path	Shape	dtype	Semantics
`action/joint/position`	`(14,)`	`float64`	Commanded arm joint positions (rad); 7 left + 7 right
`action/left_effector/position`	`(1,)`	`float64`	Commanded left gripper (rad)
`action/right_effector/position`	`(1,)`	`float64`	Commanded right gripper (rad)
`action/waist/position`	`(5,)`	`float64`	Commanded body/waist joint positions (rad)
`action/head/position`	`(3,)`	`float64`	Commanded head joint positions (rad)
`action/end/position`	`(2, 3)`	`float64`	Commanded end-effector positions (m); row 0=left, 1=right
`action/end/orientation`	`(2, 4)`	`float64`	Commanded EE orientations as quaternions `(x, y, z, w)`
`action/robot/velocity`	`(6,)`	`int64`	Chassis / mobile base command

`state/`

Path	Shape	dtype	Semantics
`state/joint/position`	`(14,)`	`float64`	Measured arm joints (rad)
`state/joint/velocity`	`(14,)`	`float64`	Arm joint rates
`state/joint/effort`	`(14,)`	`float64`	Arm joint torques
`state/joint/mode`	`(14,)`	`uint32`	Per-joint mode flags
`state/head/position`	`(3,)`	`float64`	Measured head joints (rad)
`state/head/velocity`	`(3,)`	`float64`	Head joint rates
`state/head/effort`	`(3,)`	`float64`	Head efforts
`state/head/mode`	`(3,)`	`uint32`	Head mode flags
`state/waist/position`	`(5,)`	`float64`	Measured waist joints (rad)
`state/waist/velocity`	`(5,)`	`float64`	Waist rates
`state/waist/effort`	`(5,)`	`float64`	Waist efforts
`state/waist/mode`	`(5,)`	`uint32`	Waist mode flags
`state/robot/position`	`(3,)`	`float64`	Measured base pose
`state/robot/orientation`	`(4,)`	`float64`	Base orientation (quaternion)
`state/end/position`	`(2, 3)`	`float64`	Measured EE positions
`state/end/orientation`	`(2, 4)`	`float64`	Measured EE quaternions
`state/end/arm_position`	`(2, 3)`	`float64`	Arm-frame positions
`state/end/arm_orientation`	`(2, 4)`	`float64`	Arm-frame orientations
`state/end/pose`	`(28,)`	`float64`	Flattened pose vector
`state/end/velocity`	`(24,)`	`float64`	End-state velocities
`state/end/wrench`	`(24,)`	`float64`	Wrench / F-T vector
`state/end/mode`	`(1,)`	`uint32`	EE controller mode
`state/end/errcode`	`(1,)`	`uint32`	Error code
`state/end/errmsg`	`(1,)`	`object`	Textual error (variable-length)
`state/left_effector/position`	`(1,)`	`float64`	Measured left gripper
`state/right_effector/position`	`(1,)`	`float64`	Measured right gripper

`task_desc.json`

Each episode's meta/task_desc.json contains a JSON object with:

Key	Description
`task_name`	Task identifier string (e.g. `grasp_the_drink`)
`prompt`	Object with a `text` template and variable slots (e.g. `{drink}`, `{drink1}`, `{drink2}`) plus a `drink` list of possible values
`description`	Short natural-language task summary
`scoring`	Rubric string describing scoring breakdown
`task_tags`	List of tags (`grasping`, `manipulation`, `navigation`, `whole_body`, `G2`)

Joint Limits (radians)

Hardware / URDF joint limits for the G2A platform:

Body (waist) — 5 joints

Joint	Min	Max
`idx01_body_joint1`	−1.0821	0.0002
`idx02_body_joint2`	−0.0002	2.6529
`idx03_body_joint3`	−1.9199	1.5710
`idx04_body_joint4`	−0.4363	0.4363
`idx05_body_joint5`	−3.0456	3.0456

Head — 3 joints

Joint	Min	Max
`idx11_head_joint1`	−1.5710	1.5710
`idx12_head_joint2`	−0.3492	0.3492
`idx13_head_joint3`	−0.5348	0.5348

Left arm — 7 joints

Joint	Min	Max
`idx21_arm_l_joint1`	−3.0718	3.0718
`idx22_arm_l_joint2`	−2.0595	2.0595
`idx23_arm_l_joint3`	−3.0718	3.0718
`idx24_arm_l_joint4`	−2.4958	1.0123
`idx25_arm_l_joint5`	−3.0718	3.0718
`idx26_arm_l_joint6`	−1.0123	1.0123
`idx27_arm_l_joint7`	−1.5359	1.5359

Right arm — 7 joints

Joint	Min	Max
`idx61_arm_r_joint1`	−3.0718	3.0718
`idx62_arm_r_joint2`	−2.0595	2.0595
`idx63_arm_r_joint3`	−3.0718	3.0718
`idx64_arm_r_joint4`	−2.4958	1.0123
`idx65_arm_r_joint5`	−3.0718	3.0718
`idx66_arm_r_joint6`	−1.0123	1.0123
`idx67_arm_r_joint7`	−1.5359	1.5359

End effectors (grippers)

Joint	Min	Max
`idx31_gripper_l_inner_joint1`	−0.91	0
`idx71_gripper_r_inner_joint1`	−0.91	0

Robot Platform

Model: Agibot G2A — a full-size humanoid with mobile base, 5-DOF waist, 3-DOF head, dual 7-DOF arms, and parallel-jaw grippers.
End Effectors: Listed per-episode in meta_info.json → ee_list (e.g. zhiyuan_gripper_omnipicker).
Collection: Teleoperation (task mode TDC).
Frequency: 30 Hz aligned proprioception; video streams at native sensor rate, synchronized via timestamps.

Quick Start

import h5py
import json
from pathlib import Path

task_root = Path("grasp_the_drink")
episode_dir = task_root / "data" / "episode_00000"

# 1. Read task description
with open(episode_dir / "meta" / "task_desc.json") as f:
    task_desc = json.load(f)
print(task_desc["grasp_the_drink"]["description"])

# 2. Load proprioception
with h5py.File(episode_dir / "states" / "aligned_joints.h5", "r") as f:
    frame_keys = sorted(f.keys(), key=int)
    print(f"Number of frames: {len(frame_keys)}")

    # Read frame 0
    frame = f[frame_keys[0]]
    action_joints = frame["action/joint/position"][:]      # (14,)
    state_joints  = frame["state/joint/position"][:]       # (14,)
    left_gripper  = frame["state/left_effector/position"][:] # (1,)
    print(f"Action joints shape: {action_joints.shape}")
    print(f"State joints shape:  {state_joints.shape}")

# 3. Video files
video_dir = episode_dir / "videos"
for mp4 in sorted(video_dir.glob("*.mp4")):
    print(mp4.name)

Citation

If you use this dataset, please cite:

@misc{dexmal_g2_icra_dataset,
  title   = {RoboChallenge ICRA Whole Body Control Dataset},
  author  = {Dexmal},
  year    = {2026},
  url     = {https://huggingface.co/datasets/RoboChallenge/icra_wbc}
}