aagha314/ur-pick-place-cube

UR5 pick-and-place cube (real robot)

Teleoperated demonstrations on a real UR5 with a Robotiq gripper: the arm picks a black cube from one trivet and places it on the other. Episodes were recorded back to back, so the direction alternates every episode (trivet at x ~ 0.30 m -> x ~ 0.10 m, then back).

Format is the raw layout of REAL_WORLD_DIFFPO_MEMORY (replay_buffer.zarr + videos/<episode>/<camera>.mp4).

Contents

path	what
raw/replay_buffer.zarr	low-dim streams at 10 Hz
raw/videos/<ep>/<cam>.mp4	4 cameras at 30 fps (3 video frames per step)
bird_main/pick_place_cube_crop.bird_main.camera_0+camera_2.96x96.zarr	the converted dataset the BIRD policy was trained on

To use the raw data with the repo's loaders, put both under one directory: <dir>/replay_buffer.zarr and <dir>/videos/.

Episodes

• 70 episodes, 9,616 steps at 10 Hz; length 103-234 steps (median 135).
• The recording had 71 episodes. Original episode 35 was removed (10 steps, no grasp, its camera videos disagree in length); later episodes were renumbered, so the current indices are contiguous 0-69.
• Kept, but not clean:
  • episode 10: the gripper closes, reopens in place, and regrasps before carrying the cube;
  • episode 46: the operator opens the gripper mid-transport (step 93), the cube falls on the table, and it is picked up again (step 154) and placed.

raw/replay_buffer.zarr

key	shape	meaning
data/action	(T, 7)	target pose: position (3), axis-angle (3), gripper width (1)
data/robot0_eef_pos	(T, 3)	end-effector position (m)
data/robot0_eef_quat	(T, 3)	end-effector rotation as axis-angle, despite the name
data/robot0_gripper_qpos	(T, 2)	finger positions (+w, -w)
data/timestamp	(T,)	seconds
meta/episode_ends	(70,)	exclusive end index of each episode

action[t] equals the measured state at t+1 (position, rotation, gripper width); the last step of each episode holds. Rotations sit near |angle| = pi, so the axis-angle wraps (12 times over the dataset); convert to rotation_6d or a matrix before learning on it, as both training pipelines do.

Cameras (raw/videos)

camera	resolution	view
0	1280x720	third person, side of the table
1	1280x720	third person, front; shows the teleoperator and the leader arm
2	1440x1080	wrist (fisheye)
3	1440x1080	wrist (fisheye)

Camera 1 sees the operator moving the teleoperation leader arm, i.e. the action; both policies trained on this data exclude it and use cameras 0 + 2.

bird_main/...96x96.zarr

Produced on branch bird_main_sepmae of REAL_WORLD_DIFFPO_MEMORY by

python -m algorithms.bird_main.real_data --dataset <dir> \
    --views image=camera_0,wrist_image=camera_2 --size 96

key	shape	meaning
data/img	(T, 96, 96, 3) uint8	camera 0 (agent obs key image)
data/img_wrist_image	(T, 96, 96, 3) uint8	camera 2 (agent obs key wrist_image)
data/state	(T, 11)	eef position (3), rotation_6d (6), gripper qpos (2)
data/action	(T, 10)	position (3), rotation_6d (6), gripper width (1)
meta/episode_ends	(70,)

Images are centre-cropped to square and resized from the full-resolution frames. The conversion stores frames as JPEG2K (level 50) on the way, so they differ from a direct resize by about 0.75/255 on average. eval/eval_bird_main.py reads this zarr for the observation spaces and the action normalization.

The diffusion-policy baseline's cache (<dir>.baseline.70.rotation_6d.zarr.zip) is not included; the repo rebuilds it from raw/ on first use.