You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fix: iterate A/V every cycle and push the feed-forward harder
The chart after 5 cycles settled to ~zero (P/D/I good) but still showed a
steady-speed lag (V too low) and accel/decel spikes (A too low), because A/V
were only tuned on the final cycle and gave up early. Now A/V run every cycle
(true coordinate-descent iteration with P/D/I), with a tighter plateau (5%),
lower cruise target (3), higher caps (A 2e6 / V 10000) and a couple more
attempts, so they push further toward zero error during the move.
48 tests green; typecheck + verify-build OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
exportconstV_CRUISE_OK=5;// |mean P term| in cruise considered ~zero
164
+
exportconstA_MAX=2_000_000;
165
+
exportconstV_MAX=10000;
166
+
exportconstAV_PLATEAU=0.05;// <5% improvement → stop raising (push further before settling)
167
+
exportconstV_CRUISE_OK=3;// |mean P term| in cruise considered ~zero
168
168
169
169
constnoMove: AutoDecision={kind: "fail",reason: "No steady-speed move detected — increase the A/V test move length or speed so the axis reaches cruise."};
0 commit comments