Skip to content

feat(hwci): bubble-search max_ramp and soft snaps for heavy props#47

Open
AlexKlimaj wants to merge 2 commits into
ark-releasefrom
feat/hwci-max-ramp-bubble-search
Open

feat(hwci): bubble-search max_ramp and soft snaps for heavy props#47
AlexKlimaj wants to merge 2 commits into
ark-releasefrom
feat/hwci-max-ramp-bubble-search

Conversation

@AlexKlimaj

Copy link
Copy Markdown
Member

Summary

  • Replace the max_ramp verify path that started at the field ceiling and backed off by ×0.7 (long desync thrash on 10″ / high-inertia props) with a floor-first bubble search: prove lo is safe, geometric climb, binary refine to the largest safe value; optional physics hint is one probe, not a high-first ladder.
  • Scale the verify snap to the probe envelope: classic 0.30→0.95 for 5–7″ benches; milder spinning-state steps when max_rpm is low or the probe ceiling is modest (e.g. 900KV + 10″).
  • Add tunes/ark4in1-900kv-10inch.yaml (full stages including the new ramp path) and profiles/ramp_snap_10inch.yaml for standalone snap A/B on the 10″ bench.
  • Unit tests updated/extended (test_tuner.py, 56 passed offline).

Motivation

On 900KV + HQProp 10x5x3 + 6S, the old ramp stage spent dozens of trials desyncing from max_ramp=255 downward. Steady probe trials were clean; the failure mode was the search policy + snap geometry, not the efficiency stack. Floor-first search and probe-scaled snaps keep certification without thrashing the plant.

Test plan

  • PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest tests/test_tuner.py (56 passed)
  • Optional bench: hwci tune --spec tunes/ark4in1-900kv-10inch.yaml --config rig.yaml --battery-cells 6
  • Optional A/B: hwci run --profile ramp_snap_10inch --config rig.yaml --battery-cells 6 with current vs older firmware

Stop thrashing the plant with ceiling-first 0.7 backoff verifies. Find the
largest safe max_ramp from the floor (geometric climb + binary refine),
scale the verify snap to the probe envelope (10" no longer forces 0.30→0.95),
and ship a 900KV/10" tune spec plus a ramp_snap_10inch bench profile.
Cut throttle mid-profile when firmware bemf_timeout, RPM collapse, current
spiral, or a blind-step burst appears after spin is established — do not
finish remaining segments into a desynced rotor. Raise the 900KV/10" tune
efficiency probe through 70% with matching safety.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant