LIBERO Layout: GT and DA3
LIBERO spatial, object, goal and 10 suites: 1,695 successful replay episodes, 274,008 frames, 10 Hz. Derived from LIBERO demonstrations for layout-conditioned robot policy training.
Files
data/andmeta/: LeRobot v2.1 trajectory dataset. RGB main and wrist images are 256x256; state float32[8]; actions float32[7]. Tasks provide language instructions.layouts/: episode-indexed arrays. For each of pre, post_raw, post_retreat: RGB uint8[E,256,256,3], GT depth float16[E,256,256] metres, DA3 depth float16[E,256,256] metres.layouts/retreat_success.npy: choose post_retreat when true, otherwise post_raw.layouts/meta.json: episode provenance and camera information.norm_stats/norm_stats.json: openpi normalization statistics, action horizon 10.
GT and DA3 share the exact same trajectories. Select *_depth_gt.npy for simulator depth or *_depth_da3.npy for DA3METRIC-LARGE estimated metric depth (not aligned to GT). Layout row e matches trajectory episode_index e. Pre is the initial frame; post_raw is the state after the final action; post_retreat is captured after withdrawing the robot.
A training sample combines current main/wrist RGB, state, a 10-step action chunk, language and pre/post layout images. Depth arrays store raw metres; image encoding happens at training time. Per-frame auxiliary depth supervision is not included.
Source benchmark: https://github.com/Lifelong-Robot-Learning/LIBERO
- Downloads last month
- -