[transfix]

nav_plant::dyn_bike_si_of now gives P8a's dynamic model the values that Jolt's one-tick residual asks for. Re-run on Jolt, the P8b gate shows that training through the dynamic model now helps rather than only hurting less.

The change

• Tires. Jolt's lateral friction ramps linearly to its peak at tire_peak_slip_angle, a cornering stiffness of 1/α_peak per unit load (jolt_backend.cpp). The old Pacejka tire, peaking at the same angle, started at B·C = C·tan(π/2C)/α_peak, about 4.5× steeper.
  • B is now set so that B·C = kJoltTireStiffness/α_peak, with kJoltTireStiffness = 1.12.
  • Front and rear are alike: front_stiffness defaults to 1 (it was 0.85), since Jolt's tires have no load sensitivity.
  • For both presets this is ×0.30 front and ×0.25 rear against the old values.
• A direct drive (one gear and no torque converter, like the RoboRacer) gets no drive lag and no drag or rolling resistance. The driver model's integral trim tracks the commanded acceleration, so the command is the achieved acceleration.
• A geared drive through a converter (the Humvee) keeps its 0.1 s lag and its resistances.
• drive_lag now defaults to −1 (the vehicle's own value); an explicit value still wins.

Where the values come from. fit_dyn_bike (#749) fitted Jolt runs of both presets (slalom and city, 50 Hz) on seeds 100–102 and judged them on seed 103. The best single rule for both vehicles was the one above: the held-out twist residual came out about 11× lower than with the spec-sheet values (geometric mean over four runs).

The P8b Jolt gate, before and after

City RoboRacers, 20 held-out seeds × 48 agents × 400 ticks, paired bootstrap, Linux, master 65e7230 against this branch:

training	dynamic-trained, before / after	dynamic − untrained seed, before	after
lr 5e-5, 150 steps (the test's default)	50.8% / 51.6%	+0.2 [−1.8, +2.2]	+0.9 [−1.0, +3.0]
lr 1e-4, 300 steps	48.5% / 52.1%	−2.1 [−4.2, +0.1]	+1.5 [−1.4, +4.1]
lr 2e-4, 300 steps	43.1% / 53.8%	−7.5 [−10.0, −5.3]	+3.1 [+0.7, +5.5]
checkpoints selected on 32 validation seeds	52.2% / 53.2%	+1.6 [−0.4, +3.6]	+2.6 [+0.7, +4.6]

• The point-mass arm and the seed are unchanged: the point mass is 51.4%, 49.9%, 22.2% and 52.2% across the four rows, and the seed 50.6%.
• Fitted minus before, on the dynamic arm:
  • lr 2e-4: +10.6 [+8.0, +13.1];
  • lr 1e-4: +3.5 [+0.5, +6.4];
  • default lr: +0.7 [−1.4, +2.9], within the noise.
• Validation reach of the dynamic arm, under selection with 32 seeds, now rises steadily over training, 0.497 → 0.526 at step 210. Before, it fell to 0.41.
• The selected checkpoint is the same with 8 or 32 validation seeds.
• The gate logs and run_gate.sh are in /home/joe/src/cvc/p8b-gate-2026-10-20 on the build box.

Tests

• DynBike.HumveeParametersFromItsSpecSheet asserts the new rule, for both presets.
• DynBike.SteadyLateralNeverExceedsMuG now measures the body's lateral acceleration, vy' + r·vx, over the last second.
  • With the softer tires the Humvee drifts past its limit instead of settling into steady understeer. There, r·vx alone overstates the lateral acceleration (4.80 against μg = 3.92 at μ 0.4).
  • The body's lateral acceleration peaks at 3.67 m/s² (0.94 μg), and more Newton iterations or a smaller step change nothing.
• dyn_bike_golden.json is regenerated. dyn_bike: the golden for libcvc's Jolt-fitted dyn_bike_si_of CVC-Lab/GRL-SNAM#134 carries the same file for the torch twin, which passes against it to 1e-9.
• Local: nav/physics ctest 402/402 (Linux, Jolt 5.6.0 from cvcpkg).

Interactions with open PRs

• nav: the dynamic bicycle in explicit port-Hamiltonian form, with its energy balance #748 (dyn_bike in port-Hamiltonian form): its DynBikePH gates use humvee() and need re-checking after this merges. With softer tires the linear-tire regime reaches further, and the lateral-g bins where the tangent D turns indefinite will move.
• physics: log nav_plant runs for fitting a plant model, and dyn_plant's residual against Jolt #749 (the exporter and fit_dyn_bike): its fitted factors and its roadmap table are relative to the old spec sheet. After this merges, the spec-sheet column there means these values.
• nav: dyn4_plant, the differentiable two-track vehicle with suspension (P8d) #731 (P8d, four wheels): if its golden uses dyn_bike_si_of, it moves too.