pick3_graspproj_chunkrel

The pick3 episodes prepared for chunk-relative action training, with the grasp projection applied. Built to be a controlled A/B against SteveNguyen/grabette_pick3_graspproj_480: same 554 episodes, same (strategy, closure) gripper channels, so the ONLY difference is how the end-effector motion is represented.

episodes / frames	554 / 91123
tasks	'pick up the red can', 'pick up the mustard bottle', 'pick up the cup'
action	8D ['x', 'y', 'z', 'ax', 'ay', 'az', 'strategy', 'closure'] — ABSOLUTE pose (axis-angle) + projected gripper
observation.state	2D ['strategy', 'closure']
camera	1 x 480x360

Two things you must know to train on this

The action stats are CHUNK-RELATIVE (chunk_size 50), not absolute: the pipeline converts to chunk-relative offsets before normalisation. action_absolute keeps the originals. Training without the chunk-relative processor step will mis-normalise — measured, the rotation channels land OUTSIDE [-1, 1].

The gripper is projected: closure = 1.0 means "drive fully closed" and the object stops the fingers. meta/grasp_projection.json records the calibration. Eval must decode it (--grasp_projection on).

See docs/relative_actions_lerobot_native.md and docs/grasp_projection.md in the grabette monorepo.