Shiki42/piperx-general-pickup-preview

Piper X General Pickup — Preview

▶ Play episode 0 with all three synchronized cameras

这是 Piper X 仿真抓取采集 pipeline 的 preview 数据集：16 条成功轨迹、2,894 帧、25 FPS。 顶部、左腕、右腕三路 RGB 均为 640 × 480，同步采集。 在播放器中可切换 episode，并查看活动臂状态与控制目标。

Contents

Field	Value
Format	LeRobot v3.0 (Parquet + H.264 MP4)
Robot	Two Piper X arms; one active arm per episode
Episodes / frames	16 / 2,894
Cameras	top, left_wrist, right_wrist
Frequency / resolution	25 FPS / 640 × 480
Assets	correction_tape/00000, sachet/00000, chocolate_bar/00000
Collection	Four spatially isolated workspaces in one Isaac Sim process

The four fixed asset/seed requests were repeated four times for throughput verification. These are preview examples, not sixteen distinct tasks or a full catalogue dataset.

Behavior and annotations

The expert prioritizes the live mesh center, approaches and descends with open fingers, and closes only after measured alignment and aperture checks pass. It physically lifts the object approximately 10 cm and holds it for one second. The inactive arm stays at joint-zero Home with its gripper open. Objects are not attached or teleported during recording.

• observation.state: six measured active-arm joint positions (radians), then measured gripper-open fraction.
• action: six active-arm joint position targets, then target gripper-open fraction (0 closed, 1 open).
• observation.active_arm: left=0, right=1.
• observation.environment_id: simulation workspace index (0–3).
• annotation.phase: initial=0, approach=1, descend=2, close=3, lift=4, hold=5.
• Images: observation.images.top, observation.images.left_wrist, observation.images.right_wrist.

Observations are captured before each action is held for 40 ms. The top camera is the installed task's oblique overhead view; the inactive wrist camera is not repositioned to track the object. Global HDR/key lighting is shared per collection wave; tabletop, asset scale and layout are sampled per episode.

Validation and provenance

All sixteen trajectories passed center/alignment, open-before-close, inactive-arm Home and physical lift/hold checks. Maximum measured center offset: 9.28 mm. Final physical object lift: 9.783–9.980 cm. All camera streams decoded fully at 25 FPS. The official LeRobot v3 writer and reader passed sixteen-episode export and 48 first/middle/last-frame reads. Non-image trajectories match the synchronous collection baseline element for element.

Provenance and diagnostic metadata are in meta/a2r_manifest.json. Collector code: Shiki42/a2r, PR #1. The preview was collected on Coder A using the already installed RoboDojo, Isaac Lab and cuRobo versions.