e2 to e4. The pawn moved.
First grab that counted. Week 5 could reach a square. This week the arm picks and places. One catch: there was only one piece on the board when I ran it. A full board still needs work.
first grab
Pick from e2. Drop on e4. It works. Still not fully accurate on the square, but the piece leaves and lands. Close enough that I stopped treating grasp as the blocker.
That run was on an almost empty board. One pawn, nothing crowding the path, nothing next door to clip. Real chess is crowded. Neighboring pieces, tighter clearances, more ways to knock something over. I have not shown the same grab there yet.
I started with head-cam trained movement. Got the arm in the neighborhood. Then switched toward the wrist cam when I needed accuracy. Head cam finds the square. Wrist cam sees the piece. That swap is why the empty-board grab finally stuck.
grab and drop
Grab and drop training is done and works well on leader–follower. Same setup I switched to last week. Glad I ditched the headset before this part. No Quest tremor baked into the episodes.
The demos I trust so far are still the clean ones: sparse board, simple pick, simple place. Next batch needs occupied boards, or I'll just be teaching it a version of the game that only exists in demos.
still off the record
Ordered parts for a use case that isn't chess. Talked to more people who might care about that version. Still not naming anyone or anything here. Too early.
what I took from it
Two weeks ago I could see the board. Last week the arm could go to a named square. This week it picks a pawn and puts it down, at least when the board is empty enough to be nice about it.
Square centering still bothers me. So does the full-board case. Wrist cam is the fix I'm chasing for both.
Next: train and test with more pieces on the board, cleaner landings, more wrist-cam episodes. Keep the other thread quiet until the parts arrive.