mwnuk's picture
Add the data files used by the guide notebooks
c3788b8 verified
|
Raw History Blame Contribute Delete
3.46 kB
metadata
license: other
license_name: mixed-see-card
pretty_name: Robot Learning Guide Data
task_categories:
  - robotics
tags:
  - robotics
  - imitation-learning
  - libero
  - tactile
  - audio
  - tutorial
size_categories:
  - 1K<n<10K

Robot learning guide data

Small, ready-to-load files used by the notebooks at github.com/wnukmat/robot-learning-notebooks, which accompany the guide at matthewwnuk.com/guide.html.

Everything here is derived from openly licensed sources and reduced in size so that a notebook downloads megabytes, not gigabytes. Nothing here is a new recording. If you need the full data, use the original datasets linked below.

from guide_utils import data
d = data.libero()             # actions and states of 379 demonstrations
f = data.libero_frames()      # 1,516 frames with CLIP embeddings

Files

File Contents Derived from Licence of the source
libero10/actions_states.npz action (101469, 7), state (101469, 8), episode, task, frame for all 379 demonstrations, 10 steps per second lerobot/libero_10 Apache-2.0
libero10/tasks.json The instruction for each of the 10 tasks same Apache-2.0
libero10/frames.npz Four frames per episode (1,516) at 96 × 96, their CLIP ViT-B/32 image embeddings, 16 × 16 grey pixels, and CLIP text embeddings of six queries same; embeddings computed with openai/clip-vit-base-patch32 Apache-2.0
libero10/text_queries.json The six text queries written for the guide MIT
libero10/sequences.npz Six-second clips (60 frames at 128 × 128, starting 2 s into the episode) from 12 episodes, scene and wrist cameras lerobot/libero_10 Apache-2.0
tactile/press.npz One press: force-torque timeline, pressure summary, and three gel-camera and pressure-array frames Voxel51/TaF-Tactile-Force, a conversion of jiamig/taf-dataset MIT
audio/pour.npy, audio/knock.npy Two clips resampled to 16 kHz mono PDSounds via Wikimedia Commons: pouring water by earthcalling, knocking by stephan Public domain

Field notes

  • LIBERO state is hand position (3), hand orientation (3) and two gripper finger positions. Action is the change in hand position (3), the change in orientation (3) and a gripper command of −1 (open) or +1 (close).
  • frames.npz frac is how far through its episode each frame was taken: 0.1, 0.35, 0.6 or 0.85.
  • press.npz t is seconds, ft is force (N) and torque (Nm) as fx, fy, fz, tx, ty, tz, pr is total pressure, peak cell and number of cells in contact, idx gives the three time indices of the stored images tac and pm.

Citation

If you use the underlying data, cite its authors:

  • Liu et al., 2023. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning. arXiv:2306.03310
  • The TaF tactile-force dataset: see jiamig/taf-dataset