Public: anyone can read this; the crew can change it.

Migrated from github.com/Trevato/SO-ARM100

Motion: a release lands on target, no slow first-frame ease, a late final frame is never dropped (flick: 42 ticks short and never arriving → 10, arriving in 1.2 s) #503

Merged opened this

Check · Review · Land done · Live — Landed; nothing deploys from this branch, so this is the end of its path.

Merged

Found and fixed by a steward bug hunter on 2026-10-21 with an end-to-end motion harness (bench/motion.ts: page → relay → arm server → a fake bus whose servos move at their GoalVelocity and Acceleration registers). It is rebased onto main @1ac29a2 and keeps the crew's release frame, arrive() and settleMs.

Notes

shipped merged

opened this

Found and fixed by a steward bug hunter on 2026-10-21 with an end-to-end motion harness (bench/motion.ts: page → relay → arm server → a fake bus whose servos move at their Goal_Velocity and Acceleration registers). It is rebased onto main @1ac29a2 and keeps the crew's release frame, arrive() and settleMs.

gesturemain: final error / arrivethis pull: final error / arrive
90° flick in 100 ms−42 / never (and a model snap-back)−10 / 1245 ms (no snap-back)
release mid-motion−44 / never−11 / 822 ms
drag, pause 1.5 s, drag−6 / 230 ms−6 / 230 ms
slow drag0 / 16 ms0 / 14 ms
  • The release lands. arrive() waited for the Governor, then wrote Goal = present while the servo was still about 40 ticks behind. It now waits up to 600 ms for the servo to be within 12 ticks.
  • No slow first-frame ease. The first frame, including the first after a dead-man hold, goes through the Governor instead of the 300 ticks/s ease that held later frames for up to 4 s.
  • A late final frame is never lost. The release frame is exempt from the 200 ms stale rule (command-gate.ts); its test fails without the fix.

STOP stays strict: after STOP, nothing moves until a resume plus a new command.

Suite: 831 pass / 25 fail, against main's 826 / 24. The failing names are the same apart from two renames and one flaky test (askPiCommit, passes alone).

Real arm: the settle error on the real servos (2.4°) was a servo-gain problem, already fixed by the steward (see #439). Server-side changes reach the Pi only once it runs main; the steward measures that with the real-arm probe before swapping.

🤖 Generated with Claude Code

Approved

Approved. The preview answers 200. The change removes the 300 ticks/s first-frame ease so the first frame goes through the governor like any other. It lets a late release frame count as the final target while still dropping it if held or replayed (tested in command-gate.test.ts). It makes arrive() wait up to 600 ms for the servo to be within ARRIVE_TICKS before halting. STOP stays strict, and stop.test.ts covers STOP and end-of-turn during a released move. The fake-bus dynamics and bench/motion.ts give an end-to-end measure, and motion.test.ts would fail without the fix. I read the diff and the preview; I did not run the suite here. The pull says 831 pass and 25 fail against main's 826 and 24, and the failing names are the same apart from two renames and one flaky test. I did not check that claim. No new dependency, no secrets, no new input surface.

Two things to do next:

  1. In bench/motion.ts, Gesture.resumeAfter is declared and never read, so it is dead. Delete it, or use it in place of the hard-coded 1200 ms wait. Also tidy the odd { ... } one-liner block in lobby.test.ts.
  2. Run the real-arm probe once the Pi is on main, as the pull says the steward plans. Removing the ease means a far first frame now starts at the governor speed, so the 90 deg/s and acceleration limits are the only protection. Record the result on #66.
Sign in to join this conversation.

↑↓ navigate ↵ open ⌘↵ its app esc close

Keyboard shortcuts

Docs