RoboReel

Human demonstration videos and simulated robot trajectories for 10 manipulation tasks.

Layout

human_demos/
  train/{no_distraction,with_distraction,pose_defined}/<task>/<episode>/
  eval/{no_distraction,with_distraction}/<task>/<episode>/
  excluded_human_demos_with_distraction/<task>/<episode>/
    depths.h5
    egocentric.mp4  front.mp4  left.mp4  right.mp4
robot_trajectories/
  {with_distraction,no_distraction}/<task>/<episode>/
    <timestamp>.h5
    <timestamp>.json
    0.mp4  base_camera.mp4  hand_camera.mp4  left_camera.mp4  overhead_camera.mp4  right_camera.mp4

Episode counts

Human demos, per task:

Task	train/no_distraction	train/with_distraction	train/pose_defined	eval/no_distraction	eval/with_distraction
bowl_in_plate	80	80	100	20	20
close_drawer	80	80	50	20	20
empty_basket	80	80	100	20	20
open_lid	80	80	100	20	20
pick_cube	80	80	110	20	20
press_toaster	80	80	100	20	20
push_cube	80	80	100	20	20
stack_cups	80	80	100	20	20
toys_in_basket	80	80	100	20	20
toys_in_drawer	80	80	100 (as toy_drawer_close)	20	20

Robot trajectories: 100 episodes per task in each of no_distraction and with_distraction.

Train and evaluation split

In no_distraction and with_distraction, episodes 0000-0079 are in train and episodes 0080-0099 are in eval. All pose_defined demos are used for training only.

human_demos/train/pose_defined/push_cube contains the same recordings as no_distraction/push_cube, including the episodes in eval/no_distraction/push_cube.

human_demos/excluded_human_demos_with_distraction holds 20 further with-distraction episodes per task (200 in total), numbered 0100-0119. They are not part of train or eval.

Human demonstrations

RGB video is stored only in the mp4 files (1280x720, 30 fps, H.264). depths.h5 holds the depth video and calibration. It contains one group per camera (egocentric, front, left, right):

Path	Contents
<cam>/depth/frames	uint16 depth, shape (T, H, W): 720x1280 for egocentric, 768x1024 for the others
<cam>/depth/depth_scale	Metres per depth unit
<cam>/depth/intrinsics, <cam>/rgb/intrinsics	Camera intrinsics
<cam>/depth/distortion, <cam>/rgb/distortion	Distortion coefficients; the model attribute names the model
<cam>/extrinsics/depth_to_rgb_R, depth_to_rgb_t	Depth-to-RGB rotation and translation
<cam>/timestamps	Per-frame capture time, seconds since the epoch

Files also carry task_name and demo_id attributes. These are recording metadata, so they may differ from the folder names and episode numbers. Use the folder paths as the task and episode labels.

Reading depth

Depth is compressed with Bitshuffle and zstd, so hdf5plugin must be imported before reading:

import hdf5plugin  # registers the compression filter
import h5py

with h5py.File("human_demos/train/no_distraction/push_cube/0000/depths.h5", "r") as f:
    cam = f["front"]
    depth_m = cam["depth/frames"][0] * cam["depth/depth_scale"][()]
    timestamps = cam["timestamps"][()]

Synchronisation and alignment

• Cameras are not frame-aligned, and frame counts differ between cameras in the same episode. Use timestamps to synchronise them.
• Depth is not registered to RGB. Their intrinsics differ, and so does the resolution on front, left and right. Use the calibration above to align them.

Robot trajectories

Generated by motion planning in simulation with a Panda arm (panda_wristcam_hd) using pd_joint_pos control. Each .json file records the environment ID and settings for its episode; with_distraction episodes have distract: true.

Each .h5 file has one group, traj_0, including:

Path	Contents
traj_0/actions	Actions, shape (T, 8)
traj_0/obs	State observations, length T + 1
traj_0/env_states/actors/<name>	Per-object state, length T + 1
traj_0/env_states/articulations/panda_wristcam_hd	Robot state, length T + 1
traj_0/rewards	Per-step rewards, length T