V2_packing_1012_87ep

Bimanual AgileX PiPER + 2-DOF neck, packing task: bring a basket into the work area, put 3 items into it, then lift the basket with both hands, align it and move it to the left. Teleoperated demonstrations.

This repository is the preprocessed training copy (zarr, 10 fps, images 240×320) used to train a Diffusion Policy with the flare codebase. It is not in LeRobot format.

Contents

• 87 episodes, 108,204 frames at 10 fps.
• Sources, in order: V2_packing_1012_1 (episode 9 removed), _2, _3, _4, _5, _6, _7 (episode 6 removed), _8, _9.
• The original 30 fps recordings (LeRobot v3.0) are not included.

key	shape	notes
data/observation.state	(N, 16) float32	left_joint1-6, left_gripper, right_joint1-6, right_gripper, neck_pitch, neck_yaw
data/action	(N, 16) float32	same order; gripper commands in direct mode (≈90 open)
data/observation.images.left_main	(N, 240, 320, 3) uint8	main camera (1280×720 originally)
data/observation.images.left_wrist	(N, 240, 320, 3) uint8	left wrist (640×480 originally)
data/observation.images.right_wrist	(N, 240, 320, 3) uint8	right wrist (640×480 originally)
data/episode_index, frame_index, index, task_index, timestamp	(N,)
meta/episode_ends	(87,)	exclusive end index of each episode
config.json		features, normalization stats, source repos, fps

Preprocessing

• 30 fps → 10 fps. Before decimation, state and action are low-pass filtered (4th-order zero-phase Butterworth anti-alias).
• Images are taken at the kept frames and resized with antialiased bilinear interpolation.

Stage labels (not included)

The 3-stage version used for training appends a 2-dim stage code to the state (16 → 18).

stage	code	from
stage 1	[1,0]	episode start
stage 2	[0,1]	first frame where any left-arm joint moved > 2°
stage 3	[1,1]	first frame the right gripper command re-opens (≥ 80) after the 3rd confirmed grasp

A grasp is confirmed when the command stays < 62 and the measured width stays in 22–65 for ≥ 1 s. The boundaries are computed on the 30 fps originals.

Loading

The repository root is the zarr store itself.

import zarr
from huggingface_hub import snapshot_download
path = snapshot_download("mimiminsoo/V2_packing_1012_87ep", repo_type="dataset")
z = zarr.open(path, "r")
ends = z["meta/episode_ends"][:]
state = z["data/observation.state"]