LQR redesign for delay tolerance + velocity-aware catch condition (opt-in) - #4
Open
baggepinnen wants to merge 4 commits into
Open
LQR redesign for delay tolerance + velocity-aware catch condition (opt-in)#4baggepinnen wants to merge 4 commits into
baggepinnen wants to merge 4 commits into
Conversation
Replaces NearTop with CatchCondition: optionally gates the swingup-to-LQR handover on an ellipsoidal sublevel set V = e'Se <= c_engage of a Lyapunov function of the implemented LQR loop, with hysteresis release at c_release. S is designed in robustness/07_design_catch_condition.jl and calibrated for zero false positives against the empirical catch-region grids of four perturbed plants. With use_ellipsoid = false (the default) the component reproduces the previous angle-only condition exactly. The design script also revealed that the implemented LQR gains do not correspond to a fresh LQR design from this model's linearization with the documented weights, so the Lyapunov function is computed for the implemented loop rather than from the design Riccati solution. Evaluated with the paired Monte Carlo campaign (300 identical draws, tf = 20 s): angle-only switch 93.0%, ellipsoidal gate 79.7% at the zero-false-positive calibration and 76.3% loosened. The gate is computed from the nominal model, which is exactly what mismatch invalidates, and in simulation failed catch attempts are cheap, so refusing marginal attempts costs success. Its benefits are qualitative: single chatter-free engagement and mean arrival speed below half of the angle-only switch, which matters on hardware where failed high-speed catch attempts stress the system. Hence opt-in, with the trade-off documented in robustness/README.md. Verified: behavior-neutrality regression passes exactly; defaults reproduce the previous behavior; the opt-in gate swings up nominally with a single engagement. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NV21TmGAkB1uH6auMT78vj
The original LQRstabilizer gains did not correspond to any LQR design from this model's linearization with the weights documented in test/runtests.jl and the loop had no delay margin: one sample of actuation delay dropped the swingup Monte Carlo success from 93.0% to 57.0%, and two samples failed everywhere. New gains (robustness/08_lqr_redesign.jl): discrete LQR with the documented output weights and control weight 300, selected by sweeping the control weight and validating each candidate in simulation (classical margin numbers are misleading for this unstable plant). Results, paired MC at tf = 20 s: - swingup success 93.7% (was 93.0%) - with 1-sample actuation delay 92.0% (was 57.0%); catches with 3 samples - LQR catch region grows ~30% on all tested plants and 24x under delay (8/625 -> 191/625 grid points) - equal catch capability at the top (2.5 rad/s, retained under +20% mass) The ellipsoidal catch condition is recalibrated for the new loop (retains 63-68% of the catchable set, was 40-44%, zero false positives) and reaches 85.7% when opted in (was 79.7%) - still below the angle-only switch, which remains the default. The campaign harness pins the original gains in its "baseline" configs, so historical comparisons and the behavior-neutrality regression are unaffected (verified exact). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NV21TmGAkB1uH6auMT78vj
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NV21TmGAkB1uH6auMT78vj
baggepinnen
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #3 (base branch
robust-swingup; retarget tomainafter #3 merges).Summary
Implements recommendation 2 of the robustness campaign (velocity-aware handover) and redesigns the LQR from the actual model linearization after the design work revealed the implemented gains were stale.
LQR redesign (
robustness/08_lqr_redesign.jl)The original
LQRstabilizergains (L₂ ≈ 394) did not correspond to any LQR design from this model's linearization with the weights documented intest/runtests.jl(fresh design at those weights: L₂ ≈ 120), and the loop had essentially no delay margin. New gains: discrete LQR with the documented output weightsdiag([1000, 10, 1, 1])and control weight 300, selected by sweeping the control weight and validating candidates in simulation (classical margin numbers are misleading for this unstable plant).Paired MC (300 identical draws, tf = 20 s, robust swingup config):
The campaign harness pins the original gains in its "baseline" configs, so historical comparisons and the behavior-neutrality regression are unaffected (verified exact).
Velocity-aware catch condition (opt-in)
CatchConditionreplaces the angle-onlyNearTop: optional gating of the handover on an ellipsoidal sublevel setV = eᵀSe ≤ c_engageof a Lyapunov function of the implemented LQR loop, with hysteresis release atc_release, calibrated for zero false positives against empirical catch-region grids of four perturbed plants (07_design_catch_condition.jl, recalibrated for the redesigned LQR: retains 63–68% of the catchable set, was 40–44%). Withuse_ellipsoid = false(default) behavior is identical to before.Honest negative result in simulation: paired MC gives 85.7% for the gate (was 79.7% before the LQR redesign) vs 93.7% for the angle-only switch — the gate is computed from the nominal model, which is exactly what mismatch invalidates, and failed catch attempts are cheap in simulation. Loosening the thresholds made it worse (76.3%; subset-based tuning did not generalize). Benefits are qualitative and hardware-relevant: exactly one chatter-free engagement and mean arrival speed under half of the angle-only switch. Hence opt-in, trade-offs documented in
robustness/README.md.Verification
adaptive_ab_newlqrcampaign config exactlyrec1_defaults_smoke.jl,rec2_vswitch_smoke.jl, candidate validation in08b_lqr_candidates.jl🤖 Generated with Claude Code
https://claude.ai/code/session_01NV21TmGAkB1uH6auMT78vj