grabette_pick3_graspproj_480

SteveNguyen/grabette_pick3_cartesian_480 with the gripper channels re-expressed. Everything else — images, arm deltas, episode structure — is byte-identical to the source.

What changed

The last two channels of action and of observation.state are no longer joint angles in radians. They are now:

channel	meaning
strategy	the distal joint, normalised to its reachable travel (102°). The shape of the grasp.
closure	the proximal joint, normalised to its reachable travel (93.5°). How far the grasp is closed.

In action, closure = 1.0 means "drive fully closed" — the object stops the fingers, at the servo's torque cap. Every other frame passes the demonstrated closure through unchanged, so the only thing this conversion rewrites is the grasp itself.

In observation.state, closure stays continuous (the measured closure), so "am I actually holding something" remains observable: the jaws reach the mechanical stop when empty and are blocked short when not.

Why

Replayed joint angles under-close. The demonstrated angle is where the human's fingers sat while pressing the object, and a position servo reproducing it stops just short and grips nothing. Measured across these episodes, demonstrations use only 38–60% of the proximal range. Commanding closure = 1 removes the need for the policy to predict an object-dependent angle at all.

Measured on hardware (Diffusion policy, 5/5 grasps): the resulting grasp reaches 60–68° of proximal travel against 48–51° in the demonstrations, and the achieved angle tracks the object's local width — narrower features close further.

Contents

554 episodes over three objects, concatenated from three recordings:

task_index	object	episodes	strategy at the grasp
0	can	166	0.182 ± 0.071
1	mustard	199	0.134 ± 0.139
2	cup	189	0.334 ± 0.090

The cup separates clearly; can and mustard overlap. Between-object versus within-object variance of strategy is 0.67, so the object drives the grasp shape less than episode-to-episode variation does — worth knowing before relying on the strategy channel being learnable from vision alone.

Event labelling found exactly one closing event in 94.9% of episodes (14 with none, 14 with several).

Decoding back to angles

gripette.grasp_projection.GraspProjection in the grabette monorepo. Encode and decode are exact inverses, so an eval loop must decode before sending joint goals — and must encode the live gripper position for observation.state, or the state channel is out of distribution.