RoboChallenge Dataset

Tasks and Embodiments

The dataset includes 30 diverse manipulation tasks (Table30) across 4 embodiments:

Available Tasks

arrange_flowers
arrange_fruits_in_basket
arrange_paper_cups
clean_dining_table
fold_dishcloth
hang_toothbrush_cup
make_vegetarian_sandwich
move_objects_into_box
open_the_drawer
place_shoes_on_rack
plug_in_network_cable
pour_fries_into_plate
press_three_buttons
put_cup_on_coaster
put_opener_in_drawer
put_pen_into_pencil_case
scan_QR_code
search_green_boxes
set_the_plates
shred_scrap_paper
sort_books
sort_electronic_products
stack_bowls
stack_color_blocks
stick_tape_to_box
sweep_the_rubbish
turn_on_faucet
turn_on_light_switch
water_potted_plant
wipe_the_table

Embodiments

ARX5 - Single-arm with triple camera setup (wrist + global + right-side views)
UR5 - Single-arm with dual camera setup (wrist + global views)
FRANKA - Single-arm with triple perspective setup (wrist + main + side views)
ALOHA - Dual-arm with triple wrist camera setup (left wrist + right wrist + global views)

Dataset Structure

Hierarchy

The dataset is organized by tasks, with each task containing multiple demonstration episodes:

.
├── <task_name>/                    # e.g., arrange_flowers, fold_dishcloth
│   ├── task_desc.json              # Task description
│   ├── meta/                       # Task-level metadata
│   │   ├── task_info.json         
│   └── data/                       # Episode data
│       ├── episode_000000/         # Individual episode
│       │   ├── meta/
│       │   │   └── episode_meta.json    # Episode metadata
│       │   ├── states/
│       │   │   # for single-arm (ARX5, UR5, Franka)
│       │   │   ├── states.jsonl         # Single-arm robot states
│       │   │   # for dual-arm (ALOHA)
│       │   │   ├── left_states.jsonl    # Left arm states
│       │   │   └── right_states.jsonl   # Right arm states
│       │   └── videos/
│       │       # Video configurations vary by robot model:
│       │       # ARX5
│       │       ├── arm_realsense_rgb.mp4       # Wrist view 
│       │       ├── global_realsense_rgb.mp4    # Global view 
│       │       └── right_realsense_rgb.mp4     # Side view
│       │       # UR5
│       │       ├── global_realsense_rgb.mp4    # Global view 
│       │       └── handeye_realsense_rgb.mp4   # Wrist view
│       │       # Franka
│       │       ├── handeye_realsense_rgb.mp4   # Wrist view
│       │       ├── main_realsense_rgb.mp4      # Global view 
│       │       └── side_realsense_rgb.mp4      # Side view 
│       │       # ALOHA
│       │       ├── cam_high_rgb.mp4            # Global view 
│       │       ├── cam_wrist_left_rgb.mp4      # Left wrist view
│       │       └── cam_wrist_right_rgb.mp4     # Right wrist view
│       ├── episode_000001/
│       └── ...
├── convert_to_lerobot.py           # Conversion script
└── README.md

Metadata Schema

task_info.json

{
    "robot_id": "arx5_1",                    // Robot model identifier
    "task_desc": {
        "task_name": "arrange_flowers",      // Task identifier
        "prompt": "insert the three flowers on the table into the vase one by one",
        "scoring": "...",                    // Scoring criteria
        "task_tag": [                        // Task characteristics
            "repeated",
            "single-arm", 
            "ARX5",
            "precise3d"
        ]
    },
    "video_info": {
        "fps": 30,                           // Video frame rate
        "ext": "mp4",                        // Video format
        "encoding": {
            "vcodec": "libx264",             // Video codec
            "pix_fmt": "yuv420p"             // Pixel format
        }
    }
}

episode_meta.json

{
    "episode_index": 0,                      // Episode number
    "start_time": 1750405586.3430033,       // Unix timestamp (start)
    "end_time": 1750405642.5247612,         // Unix timestamp (end)
    "frames": 1672                          // Total video frames
}

Robot States Schema

Each episode contains states data stored in JSONL format. Depending on the embodiment, the structure differs slightly:

Single-arm robots (ARX5, UR5, Franka) → states.jsonl
Dual-arm robots (ALOHA) → left_states.jsonl and right_states.jsonl

Each file records the robot’s proprioceptive signals per frame, including joint angles, end-effector poses, gripper states, and timestamps. The exact field definitions and coordinate conventions vary by platform, as summarized below.

ARX5

Data Name	Data Key	Shape	Semantics
Joint control	joint_positions	(6,)	Joint angle (in radians) from the base to the end effector.
Pose control	ee_positions	(6,)	End effector pose (tx, ty, tz, roll, pitch, yaw), where (roll, pitch, yaw) is relative euler angles from the arm base coordinate. X : back to front; Y: right to left; Z: down to up.
Gripper control	gripper	(1,)	Actual gripper width measurement in meter.
Time stamp	timestamp	(1,)	Floating point timestamp (in milliseconds) of each frame.

UR5

Data Name	Data Key	Shape	Semantics
Joint control	joint_positions	(6,)	Joint angle (in radians) from the base to the end effector.
Pose control	ee_positions	(7,)	End effector pose (tx, ty, tz, rx, ry, rz, rw), where (tx, ty, tz) is relative position from the arm base coordinate , (rx, ry, rz, rw) is quaternion rotation. X : front to back; Y: left to right; Z: down to up.
Gripper control	gripper	(1,)	Gripper closing angle, 0 for fully open, 255 for fully closed.
Time stamp	timestamp	(1,)	Floating point timestamp (in milliseconds) of each frame.

Franka

Data Name	Data Key	Shape	Semantics
Joint control	joint_positions	(7,)	Joint angle (in radians) from the base to the end effector.
Pose control	ee_positions	(7,)	End effector pose (tx, ty, tz, rx, ry, rz, rw), where (tx, ty, tz) is relative position from the arm base coordinate , (rx, ry, rz, rw) is quaternion rotation. X : back to front; Y: right to left; Z: down to up.
Gripper control	gripper	(2,)	Gripper trigger signals in the (close_button, open_button) order.
Gripper width	gripper_width	(1,)	Actual gripper width measurement
Time stamp	timestamp	(1,)	Floating point timestamp (in milliseconds) of each frame.

ALOHA

Data Name	Data Key	Shape	Semantics
Master joint control	joint_positions	(6,)	Maste joint angle (in radians) from the base to the end effector.
Joint velocity	joint_vel	(7,)	Speed of 6 joint and gripper
Puppet joint control	qpos	(6,)	Puppet joint angle (in radians) from the base to the end effector.
Puppet pose control	ee_pose_quaternion	(7,)	End effector pose (tx, ty, tz, rx, ry, rz, rw), where (tx, ty, tz) is relative position from the arm base coordinate , (rx, ry, rz, rw) is quaternion rotation. X : back to front; Y: right to left ; Z: down to up.
Puppet pose control	ee_pose_rpy	(6,)	End effector pose (tx, ty, tz, rr, rp, ry), where (tx, ty, tz) is relative position from the arm base coordinate , (rr, rp, ry) is euler (in radians). X : back to front; Y: right to left ; Z: down to up.
Gripper control	gripper	(1,)	Actual gripper width measurement in meter.
Time stamp	timestamp	(1,)	Floating point timestamp (in mileseconds) of each frame.

Convert to LeRobot

While you can implement a custom Dataset class to read RoboChallenge data directly, we strongly recommend converting to LeRobot format to take advantage of LeRobot's comprehensive data processing and loading utilities.

The example script convert_to_lerobot.py converts ARX5 data to the LeRobot dataset as a example. For other robot embodiments (UR5, Franka, ALOHA), you can adapt the script accordingly.

Prerequisites

Python 3.9+ with the following packages:
- lerobot==0.1.0
- opencv-python
- numpy
Configure $LEROBOT_HOME (defaults to ~/.lerobot if unset).

pip install lerobot==0.1.0 opencv-python numpy
export LEROBOT_HOME="/path/to/lerobot_home"

Usage

Run the converter from the repository root (or provide an absolute path):

python convert_to_lerobot.py \
  --repo-name example_repo \
  --raw-dataset /path/to/example_dataset \
  --frame-interval 1

Output

Frames and metadata are saved to $LEROBOT_HOME/<repo-name>.
At the end, the script calls dataset.consolidate(run_compute_stats=False). If you require aggregated statistics, run it with run_compute_stats=True or execute a separate stats job.