[ocayon]

awetrim-rom-2026-11-05 → main. Needs the AWETrim server from develop, which now includes the companion PR (awegroup/AWETrim#31, merged). Older servers fail it, for two reasons:

• The server takes the reel-out angle-of-attack box from the ROM's identified range. Without it, the 10 m/s cases fail with the angle of attack on the old 0° floor.
• The server accepts the steering_limits this branch sends.

The ROM it flies is AWETrim develop's default: the turn-rate steering law, depower box 1.7–2.6 m.

Why

With the new ROM, the 2026-11-07 run at Cabauw 10 m/s reached a power ratio of 0.31, and the 4 m/s cases flew about half the predicted force. The ROM is not the cause.

V3Kite and the ROM do not pull the same at the same power tape. Uwe's validate_lift_drag.jl (V3Kite parked at six tapes, 1.5–2.45 m, v_a 15.3 m/s) against AWETrim's /aero_forces gives the ROM 1.3–1.5 times V3Kite's lift at every tape. Matched by lift + drag resultant instead (AWETrim's examples/server/identify_depower_offset.py), the map is a pure offset:

V3Kite tape	ROM tape at the same resultant	offset	L/D V3Kite / ROM
1.50 m	1.856 m	−0.356	5.95 / 5.71
1.70 m	2.078 m	−0.378	5.08 / 4.77
1.90 m	2.284 m	−0.384	4.26 / 3.96
2.10 m	2.469 m	−0.369	3.61 / 3.36

The offset is −0.372 m with a std of 0.010 m (unchanged with the turn-rate ROM: its lift and drag are the same identification); a linear fit gives a gain of 0.977, so no gain is needed. depower_offset = −0.372 m / 5 m = −0.074. V3Kite's 6 % higher L/D at matched force is why the ratios end up slightly above 1. One tape map cannot close that gap.

The remaining changes get the optimizer to a startup path in strong wind, where the new ROM has to depower far (to about 2.4 m at 10 m/s against a 7.2 kN winch cap).

Changes

• depower_offset −0.074 (traj_opt.yaml), with the derivation in the comment.
• KCU steering limits sent with every /init (steering_limits_from(fcs, set)): the course controller's max_steering and the tape's v_steering, 0.32 and 0.20 1/s, against the optimizer's own 0.357 and 0.29 1/s. V3Kite's rel_steering and the ROM's u_s are the same unit (tape half-difference over 1.4 m). The request key leaves the field out while it is unset, so cached entries stay valid.
• Depower seed bounds 1.7–2.6 m (DEPOWER_SEED_BOUNDS, was 1.1–2.3; the ROM's validity band. AWETrim's ROM starts at 1.7 m because the 2025 flights already stall there) and input_depower_seed_max 2.45 m (was 1.85). At 10 m/s the optimum is near 2.4 m, and the seed march cannot start from 1.85 m.
• Larger high-wind figure: pattern_elevation_amplitude_max_high 17° (was 13°), from 12.5 m/s at 100 m (was 13.2, so Maasvlakte 10 m/s at 12.9 gets it too), and guess_el_center_high 33° (was 30°). With 13° the laps outran the optimizer's steering slew (0.29 1/s).
• Startup wind ramp (guess.startup_wind_ramp_start 7 m/s, startup_wind_ramp_step 1 m/s, 0.0 = off). Above the start wind, the startup solve runs at the start wind and then steps the wind up warm through /step (chain_step with inflow_conditions) to the run's wind. Only the final reply is flown. Cabauw 10 m/s failed cold from every seed; through the ramp it solves on the first seed.
• High-wind start length next to the low-wind schedule (low_wind: section of fc_settings_reelout.yaml, high_wind_speed 16 m/s at 100 m, high_wind_l_tether 180 m, high_wind_l_tether()). At 150 m, Cabauw 10 m/s (19.3 m/s at 100 m) has no feasible startup path: a figure big enough for the steering slew dips under the 50 m height floor. Cabauw 7 and Maasvlakte 10 keep 150 m.

Results (2026-11-10, the AWETrim code now on develop, default turbulence)

Case	Power ratio before → after	Success criteria
Cabauw 4 m/s	0.52 → 1.06	all 10 passed
Cabauw 7 m/s	rejected (turn margin 0.74) → 0.95	all 10 passed
Cabauw 10 m/s	no startup path → 0.88 (from 180 m)	all 10 passed
Maasvlakte 4 m/s	0.58 → 1.02	min elevation, phase 5 not reached, see below
Maasvlakte 7 m/s	→ 1.03	all 10 passed
Maasvlakte 10 m/s	no startup path → 0.93	all 10 passed

These runs include the steering limits. The planned steering rate reaches the 0.20 1/s limit, and every case remained feasible. With the previous (roll-gain) ROM and without the limits, the ratios were 1.11, 0.99, 0.89, 1.21, 1.07 and 0.96.

Open (not in this PR)

• Maasvlakte 4 m/s fails the phase-5 transition. It reels out well: d < 1°, 3 laps, the planned force. At the maximum length the state stays 4, steering saturates at +32 %, and the kite sinks to −2° elevation.
• The turn-rate table should be extended at the depowered end. With the 1.7 m floor the kite flies rel_depower 0.226–0.38. At 10 m/s that is past the table's identified range (0.20–0.375), so phase 5 and the re-optimization curvature checks read "margin Inf" there. Re-identifying the table up to about 0.40 is the fix.

Tests

• New: startup_wind_ramp (fake server: init at 7 m/s, warm steps to 9 m/s), the high_wind_l_tether checks, and steering_limits in the request key and JSON.
• Re-run on the last commit: test_awetrim_client, test_opt_request, test_reopt_chain and test_startup_pipeline pass, apart from the write_run_done failure below. The full count of 1588 is from the commit before.
• Updated: the shipped-value pins of the high-wind box. seeding_solve runs with the ramp off.
• 1588 passed. The three failures (script_inputs, write_run_done, ensure_gui_state_file) also fail on a clean main and look Windows-specific (path separators, file mode). The runs were made on Windows.

After merging

Delete the optimizer caches (output/opt_failure_cache.yaml, output/opt_chain_cache/): their keys cannot see the server's ROM. Then pull AWETrim develop and restart the server from it (python examples/server/run_reelout_server.py). Redo the offset whenever the ROM or V3Kite changes: run validate_lift_drag.jl, then python examples/server/identify_depower_offset.py output/lift_drag_validation.json in AWETrim.

Written by Claude, reviewed by Oriol Cayon

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.