G1-Dex1-4tasks

339 teleoperated episodes of four tabletop manipulation tasks on a real Unitree G1 humanoid (29-DoF) with two Dex1-1 parallel grippers: 181,403 frames at 30 fps, about 1.7 hours. Only the two arms and the grippers are recorded. Each frame has three camera images, the 16-dim joint state, the 16-dim action and a natural-language instruction. The format is LeRobot v2.1.

Tasks

task_index	Task	Instruction	Episodes	episode_index	Frames	Mean length
0	Drawer	Use the arm nearer to the drawer to pull it open, then use the other arm to pick up the purple cube and place it inside the drawer, and finally close the drawer.	100	0–99	47,741	15.9 s
1	Cups to coasters	Pick up the blue cup with the nearer arm and place it on the blue coaster, then pick up the green cup with the nearer arm and place it on the now-empty green coaster.	49	100–148	19,943	13.6 s
2	Cube to bin	Pick up the pink cube with the nearer arm and place it into the green bin.	91	149–239	18,463	6.8 s
3	Stack by color	Stack the blocks by color: put the red block in the center, then stack the blue block on the red block, then stack the yellow block on the blue block.	99	240–338	95,256	32.1 s

Each episode has one task, and the instruction is the same in every episode of a task.

Data

Feature	Shape	Content
observation.images.cam_left_high	480 × 640 × 3	head camera
observation.images.cam_left_wrist	480 × 640 × 3	left wrist camera
observation.images.cam_right_wrist	480 × 640 × 3	right wrist camera
observation.state	16, float32	measured joint and gripper positions
action	16, float32	commanded joint and gripper positions, same layout as the state

• Videos are AV1, 30 fps, one MP4 file per episode and camera.
• State and action layout:
  • 0–6: left arm, as shoulder pitch, shoulder roll, shoulder yaw, elbow, wrist roll, wrist pitch, wrist yaw;
  • 7–13: right arm, same order;
  • 14: left gripper;
  • 15: right gripper.
• Arm values are joint angles in rad. They map to the {left,right}_{shoulder_pitch, shoulder_roll, shoulder_yaw, elbow, wrist_roll, wrist_pitch, wrist_yaw}_joint joints of the official G1 29-DoF URDFs (unitreerobotics/unitree_ros, robots/g1_description/, for example g1_29dof_mode_15_with_dex1_1.urdf). All arm values lie within those joints' limits.
• Gripper values are Dex1-1 motor positions in rad. They are about 5.4 when open and smaller when closed. A grasp stops at about 0.6 on the drawer handle, 2.8–3.6 on cubes, and up to 4.6 on cup rims.
• meta/info.json robot_type is Unitree_G1_Dex1_Sim. This is the name of the 3-camera configuration in unitree_lerobot. Despite the name, all episodes were recorded on the physical robot.

Source and Changes

The episodes come from four public datasets: LGG100/drawer, LGG100/biomanual, LGG100/simple-pick and LGG100/Stack-the-cubes. This dataset changes them as follows:

• Merged into one dataset with four tasks. Episodes are renumbered in the order above, and each source keeps its own order. meta/episode_sources.jsonl maps every episode to its source repository and source episode index.

• 11 episodes removed, out of 350:

  Source	Source episode	Reason
  biomanual	24	8 frames, no task motion
  simple-pick	22	the cube is in the bin for the whole episode
  simple-pick	42, 43, 44, 45, 47, 48	a person resets the scene while the robot stays still
  simple-pick	46	several attempts, including a person moving the cube by hand
  simple-pick	49	the cube starts in the bin, and the head camera moves
  Stack-the-cubes	51	ends before the yellow block is stacked

• New instructions for drawer, biomanual and simple-pick. In the sources, all three read pick up cube. Stack-the-cubes keeps its source instruction.

Apart from the renumbered episode_index, index and task_index, all state and action values are unchanged. Every video is a byte-identical copy of its source file.

Known Characteristics

• Repeated frames. The cameras capture below 30 fps, so the videos repeat frames. A median of 3–7.5% of frames per camera is identical to the previous frame. The longest run of repeats is 21 frames (0.7 s).
• Timing.
  • The action leads the state by 3 frames (100 ms). Most Stack-the-cubes episodes 240–260 lead by 4 frames instead.
  • On 4 checked episodes, the video lags the state by about 3 frames.
• Isolated action jumps. Fourteen episodes have single action frames that change an arm joint by more than 0.15 rad. The largest jump is 0.226 rad. The episodes are 50, 53, 77, 80, 246, 260, 269, 284, 291, 292, 298, 328, 332 and 334. The state never changes by more than 0.126 rad per frame.
• Scene placement. Stack-the-cubes was recorded with a different robot placement relative to the table.
  • Its head camera sees only the tabletop, while the other tasks also show the wall.
  • Its episodes start with straighter elbows.
• Starting state. Episodes 240 and 241 start with the red block already in the center.

Usage

With a LeRobot release that reads format v2.1:

from lerobot.datasets.lerobot_dataset import LeRobotDataset  # older releases: lerobot.common.datasets.lerobot_dataset

dataset = LeRobotDataset("JingwuLuo/G1-Dex1-4tasks")
frame = dataset[0]  # images (3, 480, 640) in [0, 1], state, action, task

LeRobot releases that read format v3.0 need the dataset converted first, with LeRobot's v2.1-to-v3.0 conversion script. The parquet files can also be read directly; the default config of this card loads them with datasets.load_dataset("JingwuLuo/G1-Dex1-4tasks"), without the videos.

Related

JingwuLuo/ImageWAM3backbones holds three ImageWAM world-action models trained on this data with the same instructions. They trained on 335 episodes and held out episodes 8, 108, 157 and 248 for validation.