Week 4 left me with a board I could trust and an arm I still couldn't move cleanly. VR teleop still shook every trajectory I tried to record. I poked at sensitivity and smoothing on the Quest for another stretch, then quit.
The headset was costing evenings. I needed those evenings for the bench.
follow the other arm
Switched to leader–follower. Grab one arm, move it by hand. The other copies the joints.
Should have done this earlier. Leader is the arm I drag around. Follower is the one that has to play chess later. No headset, no floating controllers. Just two arms and a cable between them, more or less.
go to a5
Demo for the week: tell it to go to A5 and watch whether it does.
Two weeks ago that request would have failed for three different reasons. The board scan was still lying, VR was too shaky to trust, and square names were just labels on a picture. This week the week-4 localisation and the follower finally line up. A5 is a pose. The arm gets there.
No pick or place yet. The arm does show up where I asked, though.
grasp, almost
Same week I kept training grasp. Better than last week. Not good enough to call done. I keep recording another episode and promising myself the next one is the last.
So: go-to works. Grasp doesn't, yet. I'll take the half win.
side channel
Also talking to people who might care about this when it isn't a chess robot on a lab table. Not going into it here. Too early, and I don't want to jinx whatever it turns into.
Some of this summer stays off the blog until there's something concrete to say.
what I took from it
Week 4 was seeing. Week 5 was admitting the Quest was the wrong tool for recording moves. Leader–follower is ugly and simple and it actually tracks.
Named squares work as targets now. Grasp is still the open problem. And there's a second thread running next to the robot that I'm not ready to write about.
Next: get grasp reliable, chain go-to into pick-and-place, keep the other conversations offline until they earn a post.