Exact twin of physical-intelligence/libero (the pi05 LIBERO finetune set):
1693 episodes / 273,465 frames, with byte-identical state, actions, proprio,
episode-selection and no-op filtering. The ONLY differences by design:
images recolored (agentview + wrist): every object tinted its per-episode color
task prompts rewritten to a minimal color grammar (e.g. put magenta on red and put green on blue)
LeRobot v2.0. Train with… See the full description on the dataset page:
https://huggingface.co/datasets/adamzinebi/libero-pi05-colormin.
