myx160/pap_1112

PAP 1112 — Unitree G1 / internal Dex1 teleoperation

LeRobot Dataset v2.1 conversion of pap/pap_1112, recorded on 2026-11-12. All 201 original episodes and 52,829 time steps are retained, including the four identified mild JPEG artifacts. The dataset contains 804 camera videos at 640 × 480 and a nominal 30 FPS, totaling approximately 29 minutes 21 seconds of synchronized multi-camera time steps.

Format and compatibility

The layout and feature schema follow the LeRobot revision pinned by the official Physical Intelligence openpi repository:

• LeRobot commit: 0cf864870cf29f4738d3ade893e6fd13fbd7cdb5 (CODEBASE_VERSION = "v2.1").
• The corresponding openpi lockfile uses datasets==3.6.0.
• One Parquet file and four MP4 files per episode; standard meta/info.json, episodes.jsonl, episodes_stats.jsonl, and tasks.jsonl.
• This is a dataset conversion, not a fine-tuned policy or a complete robot-specific openpi configuration. A G1/Dex1 input/output transform and training configuration are still needed. Do not apply ALOHA-specific joint sign or gripper conversions to these values.

# Run in openpi's locked environment (LeRobot v2.1, datasets 3.6.0).
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    "myx160/pap_1112",
    revision="v2.1",
    video_backend="pyav",
    delta_timestamps={"action": [i / 30 for i in range(50)]},
)
sample = dataset[0]
# observation.state: (16,), action: (50, 16), RGB images: (3, 480, 640)

State and action

observation.state and action are float32 vectors of length 16. Their order matches the upstream Unitree G1_DEX1_CONFIG:

Indices	Meaning	Original JSON field
0–6	Left arm: shoulder pitch, roll, yaw; elbow; wrist roll, pitch, yaw	states/actions.left_arm.qpos
7–13	Right arm, same joint order	states/actions.right_arm.qpos
14	Left internal Dex1 gripper	states/actions.left_ee.qpos
15	Right internal Dex1 gripper	states/actions.right_ee.qpos

State is the recorded measured joint position; action is the recorded target command. Actions have not been replaced by next-frame states, shifted, normalized, clipped, converted to deltas, or padded. Arm joint positions use the original radian convention. Gripper values remain in the original actuator coordinate convention, not a normalized 0–1 aperture. Feature names are provided in meta/info.json.

Additional recorded values are retained under auxiliary.state.* and auxiliary.action.*: full 35-element body state, left/right six-element arm poses, torso height and velocity, body command, and torso velocity command. Pose values are preserved in their source convention. All recorded body and torso command values in this collection were zero; these are not part of the main 16-dimensional action vector.

Cameras

Feature	Original key	View
observation.images.cam_left_high	color_0	Left half of head stereo camera
observation.images.cam_right_high	color_1	Right half of head stereo camera
observation.images.cam_left_wrist	color_2	Left wrist
observation.images.cam_right_wrist	color_3	Right wrist

All four views are retained. A first openpi policy can select one head view plus both wrist views through its robot-specific transform. Videos use H.264, CRF 18, yuv420p, a 15-frame GOP, and no B frames. This adds lossy video encoding to the source JPEGs; the original raw recordings remain separate. No image inpainting or artifact repair was applied.

Time and episodes

The source JSON has frame indices and a nominal FPS, but no per-frame capture timestamps. Therefore timestamp = frame_index / 30; this is a nominal timeline, not a measured hardware capture clock. There is no resampling, frame deletion, or episode concatenation. Original episode_0000 maps to LeRobot episode index 0, through episode_0200 → 200. Global index is continuous across episodes. All episodes are in the standard train split; no held-out evaluation split has been selected. When evaluating, split by complete episodes rather than random neighboring frames.

Task text and quality

The data owner corrected the instructions by original zero-based episode number. The five ranges below are inclusive and cover all 201 episodes. Raw JSON text.goal, text.desc, and text.steps were corrected in English; LeRobot frame-level task_index, episode task lists, task statistics, and source metadata were synchronized.

Original episode range (inclusive)	Task index	Episodes	English instruction
000–040	0	41	Pick up the pink foam and place it in the yellow box.
041–057	1	17	Pick up the data cable and place it in the yellow box.
058–100	2	43	Pick up the black tape and place it in the yellow box.
101–150	3	50	Pick up the blue pen and place it in the yellow box.
151–200	4	50	Pick up the white charger and place it in the yellow box.

The earlier generic cube instruction has been superseded. meta/source_episodes.jsonl contains the corrected source text and updated raw JSON SHA256, plus the pre-correction hash for provenance. meta/task_relabeling.json records the owner-supplied mapping and validation. No success/failure labels have been inferred.

Four time steps have visually confirmed narrow gray horizontal bands consistent with upstream JPEG decoding artifacts:

Episode	Frame index (zero-based)	Camera
5	188	cam_left_high
19	47	cam_right_high
171	193	cam_left_high
189	61	cam_right_high

These frames remain in training data by the data owner's decision. quality.jpeg_artifact and quality.jpeg_artifact_camera_mask record this known list; zero means not flagged by this inspection, not proof of perfect image quality. Episode 196 was reported to have a recording-time warning, but no corresponding visible gray-band frame was confirmed. Structural validation alone cannot prove task success or absence of all visual defects.

Audio, depth and tactile observations were not recorded. Empty source joint velocity/torque fields are not fabricated into measurements.

Validation and provenance

• meta/conversion.json: schema, source mapping, encoding options and retained-artifact policy.
• meta/validation.json: complete numeric comparison, every-video full decode and timestamp checks, and sampled source-to-video RGB comparisons.
• meta/reader_validation.json: actual readback using openpi's pinned LeRobot reader and datasets==3.6.0, including 50-step action windows, episode-boundary padding, and decoded instruction text for all five tasks.
• meta/checksums.jsonl (produced immediately before upload): SHA256 and Git blob hashes for dataset payload files (excluding the checksum manifest itself).
• tools/convert_v21.py: conversion implementation; tools/validate_v21.py: independent validation against current original recordings.
• meta/task_relabeling.json: inclusive episode ranges, instructions, and validation of the owner's annotation correction.

The upload process additionally checks remote file inventory, sizes and hashes against the local payload, downloads episode 0 into a separate directory and reads it through the pinned LeRobot reader before reporting success. The uploader creates the v2.1 tag only after verification, pointing to the verified upload revision.

No data license has been specified by the owner in this conversion.