stray-light/zerograv-manipulation-trajectories-v0
⚠️ Superseded — do not use for distillation These trajectories were collected from policies trained before the sim-real action-space alignment. Measured on this data: mean |arm action| is 0.72, with 46-53% of actions at the clip boundary of [-1, 1]. For comparison, the real robot's own pi0.5-DROID policy outputs a mean of 0.066 and human teleop 0.101 — so these actions sit 7-11x outside the distribution the base model has seen. Distilling them would push pi0.5-DROID's action… See the full description on the dataset page: https://huggingface.co/datasets/stray-light/zerograv-manipulation-trajectories-v0.
Mark superseded: action distribution 7-11x outside pi0.5-DROID's range
Add pusht-droid, document its stricter QA filter and schema difference from the other 3 tasks
Add pusht-droid LeRobot-format trajectories (200M-step ent_coef=0.002 checkpoint, QA-filtered incl. require-success-at-end)
Mirror repo is now 1:1 with this one; link it directly
Document filter thresholds, per-task retention, episode-length protocol, and source checkpoints
Fix broken load example (root= is local, not remote subfolder); add camera resolutions
Update for video inclusion, new episode counts, and fix unclipped-actions claim
Update pickcube-droid-longhorizon: single-pass collection now includes all 3 camera videos
Update liftpeg-droid: single-pass collection now includes all 3 camera videos
Update pushcube-droid: single-pass collection now includes all 3 camera videos
Add dataset card describing per-task structure, QA filtering, and license
Add pickcube-droid-longhorizon LeRobot-format trajectories (100M-step DROID checkpoint, QA-filtered)
Add liftpeg-droid LeRobot-format trajectories (100M-step DROID checkpoint, QA-filtered)
Add pushcube-droid LeRobot-format trajectories (100M-step DROID checkpoint, QA-filtered)
initial commit