1clarkj/mimicvla-panda-usb

MimicVLA: Panda USB insertion from human phone recordings

72 episodes of a simulated Franka Panda inserting a USB plug into a wall port in MuJoCo. Every episode is a human demonstration: a person's plug motion, tracked with ArUco markers in a handheld iPhone video, was retargeted to the Panda and replayed. The images, joint states and actions were recorded during that replay.

• Code: github.com/1clarkj/MimicVLA
• Source videos and tracking: 1clarkj/mimicvla-human-usb-raw. Its lerobot_episode_index column maps each recording to an episode here.

Format

LeRobot v3.0, 20 Hz, 8,089 frames, one task: "Insert the USB stick into the USB port."

Feature	Shape	Meaning
observation.images.camera1	256 × 256 × 3	Fixed front camera looking at the port
observation.images.camera2	256 × 256 × 3	Wrist camera on the Panda hand
observation.state	9	7 arm joint angles (rad) and 2 finger positions (m), measured before the action
action	6	[dx, dy, dz, rx, ry, rz]: world-frame end-effector translation (m) and rotation vector (rad) per step

The action is applied to the end-effector target; differential IK (Mink) and the Panda's position actuators track it. The plug is rigidly attached to the hand.

Task and selection

The simulated port is a 15 × 7 mm rectangular opening in a flush wall; the plug is 12 × 4.5 mm. That is slightly larger than the real port (13.4 × 5.9 mm, 0.7 mm clearance per side) to absorb marker-tracking error and error in the measured marker-to-port and marker-to-plug transforms. Success means the plug-tip centre is inside the opening and at least 12 mm deep.

Episodes start after a short simulator alignment to the first tracked pose and run to the end of the human motion. Each was accepted in manual review, and its replay reaches 7 mm depth during the human motion, finishes at least 12 mm deep, and has no deep collision or joint-limit fault. capture_manifest.json records each episode's source clip, frame range and peak contact force.

Splits

The policies in the project were trained on 64 episodes and held out these 8 (stratified by recording batch, seed 2026):

[0, 6, 23, 28, 47, 55, 67, 70]

Evaluation start states

Two sets of start poses, for evaluating a policy closed loop in MuJoCo:

• The 72 episodes' own starts. capture_manifest.json lists each episode's source clip and its alignment_actions: the few actions that move the arm from its reset pose to the episode's first pose.
• 55 unseen starts (eval_starts/excluded55/, the eval_starts subset in the viewer). Start poses of recordings that are not in this dataset: 32 failed the replay gate, 7 were rejected in review, 16 were never reviewed. No policy trained on them; they are the main generalisation test in the project.

A start state is defined by its alignment actions alone, so there are no images or episodes for the 55: the evaluator resets the simulator, executes the alignment actions, and hands control to the policy, which renders its own camera images. starts.jsonl has one row per start; capture_manifest.json in the same folder is the file the code reads.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("1clarkj/mimicvla-panda-usb")
frame = dataset[0]

With the code repository, uv run mimicvla download fetches this dataset into data/. mimicvla build-dataset rebuilds it from the raw recordings' tracking and review list (see the raw dataset's card): it selects the same 72 episodes with identical actions, and joint states agree to within 0.003 rad. source_track in capture_manifest.json is the tracking file in the raw dataset.

Limitations

Simulation only. Images are rendered, with fixed lighting. One person, one port pose. Contact forces in the replays are often high (peak per episode is in capture_manifest.json).