stray-light/zerograv-manipulation-trajectories-state-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⚠️ Superseded — do not use for distillation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 head far outside its training distribution.

They also predate several sim-real alignment fixes: real gripper timing (0.9 s to close, vs 0.1 s here), torque rate limiting, gravity-residual bias torque, a 40 N contact reflex, and home-pose noise (the arm starts from an identical pose in every episode here — the first 13 state dims have zero variance).

A replacement collected with arm_action_bound=0.2 is in progress. Kept public for provenance and reproducibility only.

Zero-Gravity Manipulation Trajectories (State-Only) Dataset Card

Same source checkpoints as stray-light/zerograv-manipulation-trajectories-v0, collected fresh at larger scale with a stricter success criterion and no camera/video observations.

Tasks and Checkpoints

pushcube-droid

• Checkpoint: ppo-pushcube-droid-v3-w-3d-s1-100M-ne4096
• Peak eval success: 1.00
• Episodes: 923

liftpeg-droid

• Checkpoint: ppo-liftpeg-droid-v3-w-3d-s1-100M-ne4096
• Peak eval success: 0.75
• Episodes: 1001

pickcube-droid-longhorizon

• Checkpoint: ppo-pickcube-droid-longhorizon-v3-w-3d-s1-100M-ne4096
• Peak eval success: 0.94
• Episodes: 1050

pusht-droid

• Checkpoint: ppo-pusht-droid-ent002-w-3d-s1-200M-ne4096
• Peak eval success: 0.828 (success_at_end: 0.727)
• Episodes: 940

Differences from zerograv-manipulation-trajectories-v0

• State-only: no observation.images.* fields (no agent_view/base_camera/wrist_camera video). Collecting with cameras rendered for 256 parallel envs proved impractically slow (a single environment step could take minutes); state-only collection at the same env count is roughly an order of magnitude faster with identical episode outcomes (same policy, same physics — only rendering is skipped), which is what made a ~1000-episode-per-task collection practical.
• Stricter success criterion: collected with --require-success-at-end, so every episode's next.success is True on its final frame, not merely at some point during the episode. v0 counted episodes toward its target on success_once (ever succeeded) without this stricter filter.
• Two additional per-frame fields: episode.success_once and episode.success_at_end (booleans, constant within an episode) — added to the collection pipeline after v0 was published, so v0 does not have them. Included here for convenience even though this dataset's own collection already guarantees episode.success_at_end == True throughout.
• Larger scale: ~1000 episodes per task here vs. 29-220 in v0.
• Same 4 source checkpoints, same QA filtering (ejection-velocity and settle-velocity checks) as v0 otherwise.

Key Notes

The checkpoints themselves are not published here — they live in the training run's Weights & Biases project, not on the Hub.

Actions are the policy's raw output, not clipped to [-1, 1] — clip before using them as training targets (ManiSkill's controller clips internally before physically applying the command, so the recorded physical trajectory itself is correct either way; only the saved action array reflects the unclipped raw policy output).