[transfix]

A port-Hamiltonian vehicle model fitted to the physics (a learned plant on SE(2), CVC-Lab/LieSPHGP's method) needs two things this PR adds:

• the physics' trajectories in a plant-learning format;
• a baseline to beat. P8a's dynamic model has never been measured against Jolt; this measures it.

What it adds

• cvc::physics::plant_log (cvc/physics/plant_log.h) samples a sim_world's agents on a nav_plant every tick, as planar rigid bodies in SI. Each sample holds:

  • the centre of mass's pose on SE(2) and its body twist;
  • the steering angle;
  • the command nav held over the tick (a zero-order hold);
  • the measured acceleration, static contacts, and wheels on the ground.

• one_tick_residuals: the per-tick drift (D14's method) with diff::dyn_plant beside the kinematic bicycle.

  • Each tick, both models step from Jolt's state with the held command and are compared with where Jolt was at the tick's end.
  • Ticks are binned by the lateral g Jolt measured.
  • Ticks with a static contact, a wheel off the ground, an impact (above 1.5 g) or the vehicle at rest are left out and counted.

• cut_flights / write_npy: equal-length windows of [x, y, psi, vx, vy, r, drive, steer], float64.

• cvc plant-log (a cvc-cli subcommand over cvc/physics/plant_log_scene.h) drives the Humvee or the RoboRacer on Jolt through nav, in one of three scenes:

  • a slalom on open ground, with swerves past 0.3 g;
  • the coef trainer's city;
  • a cvc::world city bundle: vehicle routes on its rasterized map, vehicles spawned on its terrain.

  It writes flights.npy, meta.json and residual.json. The scenes and the writer are library calls (plant_log_options, drive_plant_log_scene, write_plant_log), so a trainer can drive them in process.

• src/cvc/tools/liesphgp_pack.py (standard library and numpy only) packs runs into the dataset pickle LieSPHGP's loaders read (t, train_trajectories, test_trajectories, test_trajectories_noisy, settings).

  • Layouts: the 20-wide SE(3)-lifted rows its ROV loader takes with control_dim: 2, or the 8-wide SE(2) rows.
  • Inputs: zero-order hold, with row 0's inputs zero.
  • Noise: observation noise at a stated sigma on the planar state.
  • The split is by seed. --check re-reads the file against that contract.

• nav_plant_options::ground_height: each start's ground height, for a terrain that is not flat. Empty means ground_z, as before.

• Roadmap: the drift section now publishes the dynamic model's residual (below).

Measured (local, Linux, Jolt 5.6.0+cvc.6 from cvcpkg; 4 seeds x 60 s, mu 0.9; spec-sheet dyn_bike, before #753; see "After #753" below)

The models' one-tick position residual at nav's 0.06 s tick (6 substeps), in mm, dynamic / kinematic:

lateral g	Humvee slalom	Humvee city	RoboRacer slalom	RoboRacer city
0.0-0.1	2.0 / 3.9	2.0 / 3.4	0.9 / 0.7	2.5 / 1.9
0.1-0.2	5.1 / 11.6	4.8 / 11.1	3.3 / 5.6	4.3 / 4.6
0.2-0.3	7.1 / 17.9	7.0 / 18.1	5.0 / 10.7	5.8 / 9.4
0.3-0.5	8.1 / 27.5	8.1 / 26.2	6.9 / 15.8	7.1 / 13.8
>= 0.5	5.2 / 94.0	4.5 / 52.0	6.5 / 25.1	4.8 / 19.0

• Position and heading: the dynamic model is closer wherever the tires carry load, 4-18x above 0.5 g. At the 0.02 s tick the pattern is the same: Humvee slalom above 0.5 g, 0.9 mm against 31.0 mm.
• RoboRacer speed: the dynamic model is worse, 0.02-0.12 m/s rms against 0.007-0.06. Its 0.1 s drive lag (dyn_bike_si_of's drive_lag) is slower than the vehicle's driver model.
• RoboRacer yaw rate: in the city it is off by 0.1-0.3 rad/s rms.

The 50 Hz datasets are 5 s flights from the slalom, the city, and the 3 km Austin south bundle, with seed 101 held out:

vehicle	train flights	test flights	windows (101 points, stride 75)
Humvee	472	166	1,416
RoboRacer	245	111	735

• Kinematic consistency: in the clean test flights, dx/dt matches R v_b to 0.024 m/s at p99.
• Austin yield. About half the windows are dropped, and the route planner fails on about 60% of the goals it picks. The Humvee spends 5.6% of its samples with no wheel down; those ticks and windows are left out.
• RoboRacer in the city. Its chassis touches the ground: 732 static-contact ticks, and 296 of 384 windows dropped.

Tests: physics_plant_log_test 4/4:

• a run a model made leaves that model no residual (kinematic and dynamic, nsub 1, 2 and 6);
• the flights, their layout and the .npy;
• a Humvee slalom on Jolt: row k's command is the one held over the tick, the centre of mass moves with its logged twist to 0.023 m/s, and the peak is 0.87 g;
• ground_height spawns as ground_z does.

All nav/physics ctest cases pass (402/402). The tool and the packer were run as above; the packer's check passes on all four pickles.

Follow-ups (later commits on this branch)

• Fitting dyn_bike to Jolt (fit_dyn_bike, dyn_bike_cost, cvc plant-log --fit-dyn): a coordinate descent on the one-tick twist residual over the knobs a spec sheet does not give: drive and steering lags, factors on the tires' B, the yaw inertia, and drag + rolling resistance. It fits on seeds 100-102 and is judged on seed 103.

  run (50 Hz)	drive lag	tire B front / rear	yaw inertia	drag + rolling	held-out twist residual, (m/s)²
  Humvee slalom	0.15 s	x0.40 / x0.30	x1.00	x1	0.00287 -> 0.00045
  Humvee city	0.02 s	x0.40 / x0.30	x1.20	x1	0.00280 -> 0.00060
  RoboRacer slalom	0	x0.30 / x0.25	x0.85	x0	0.00035 -> 0.00001
  RoboRacer city	0	x0.30 / x0.30	x1.75	x0	0.00229 -> 0.00018

  • Tires: the spec sheet's Pacejka B is 2.5-4x stiffer than Jolt's tires behave.

  • RoboRacer drive: its driver model tracks the commanded acceleration with an integral trim. So it has no drive lag, and it makes up for drag and rolling resistance.

  • Held out, at nav's 0.06 s tick: the fitted model's position rms against the spec sheet's and the kinematic bicycle's, in mm:

    • Humvee slalom, above 0.5 g: 3.0 / 6.4 / 55.2;
    • RoboRacer city, 0.3-0.5 g: 1.6 / 6.1 / 11.2.

    The RoboRacer's speed residual comes down to the kinematic bicycle's.

  • The Austin fits are not trustworthy: they land on the grids' edges, absorbing what the planar model leaves out on rough terrain.

  • Adopted in physics: fit dyn_bike_si_of to Jolt (tires, direct drive); P8b training now helps on Jolt #753 (merged): dyn_bike_si_of now gives Jolt-fitted values. See "After physics: fit dyn_bike_si_of to Jolt (tires, direct drive); P8b training now helps on Jolt #753" below.

  • Test: FitsTheDynamicModelToRunsOfAKnownVehicle checks that a run made by a vehicle off the spec sheet gives back every knob on its grid.

• Austin routes. The "planner fails on ~60% of goals" was the goal draw, not the planner: drawing from the whole map lands 150-450 m away only ~6% of the time, and all 8 tries missed.

  • Goals are now drawn from that ring: 0 unplanned requests.
  • The planner keeps 2 m off building meshes.
  • An agent stopped for 5 s gets a new route from where it stands. The Humvees still wedge on bundle geometry the nav grid calls free (a curb or a building base) about every 16 s, so this recovers them rather than preventing it.
  • Moving time rose from ~45% to 54% (Humvee) and 97% (RoboRacer, seed 100).

• RoboRacer chassis contacts. Its chassis box is centred on the axle line, 2-3 cm above the ground at rest, so at nav's full ±6 m/s² the pitch puts it on the ground. Contact ticks have a median longitudinal acceleration of -6.2 m/s².

  • --accel-margin / --brake-margin (nav_plant's own knobs, recorded in meta.json) at 0.4 take the city's contact ticks from 413 to 12, and the windows kept from 88 of 384 to 352 of 384.
  • The vehicle model itself is unchanged.

• --verbose prints each agent every 5 s. --samples writes samples.csv. Samples now carry the roll.

• Datasets regenerated:

  vehicle	train flights	test flights	before the follow-ups
  Humvee	475	182	472 / 166
  RoboRacer	443	167	245 / 111

  Austin gives 83 Humvee and 67 RoboRacer flights.

• Folded into cvc-cli (2026-10-20). The command line is now cvc plant-log. The standalone cvc-plant-log tool and its CVC_BUILD_PLANT_LOG option are removed, and its scene drivers and directory writer moved into the library as cvc/physics/plant_log_scene.h.

  • The three scenes run through cvc plant-log write byte-identical files to the old tool's (swerve with --fit-dyn --samples, city at 0.06 s with --steer command and the margins, Austin with --verbose --fit-dyn): flights.npy, meta.json, residual.json and samples.csv. Stdout is the same too, apart from the per-tick timing.
  • physics_plant_log_test 7/7, adding the options' defaults and a swerve driven twice to identical runs. cvc_cli_test adds four plant-log cases (56/56 of its help, dispatch and plant-log tests locally).

• CI fixes.

  • Windows: CutsFlightsAndWritesThemAsNpy removed the .npy while its stream was still open; it is closed first now.
  • Master is merged in for nav tests: build nav_material_ext_test without FMA contraction, as the rollouts are #750's macOS test fix. A conflict in src/cvc/CMakeLists.txt with the relay's move to src/arisrv was resolved by keeping both.

• After physics: fit dyn_bike_si_of to Jolt (tires, direct drive); P8b training now helps on Jolt #753 (master merged in; 7d95ba9). Every scene was rerun with dyn_bike_si_of's Jolt-fitted values.

  • Datasets: flights, meta.json and the four LieSPHGP pickles are byte-identical. Jolt drives the vehicles, and dyn_bike_si_of enters only the residuals.

  • Residuals at 0.06 s, position rms in mm, dynamic before physics: fit dyn_bike_si_of to Jolt (tires, direct drive); P8b training now helps on Jolt #753 / dynamic now / kinematic:

    lateral g	Humvee slalom	Humvee city	RoboRacer slalom	RoboRacer city
    0.0-0.1	2.0 / 1.2 / 3.9	2.0 / 1.6 / 3.4	0.9 / 0.3 / 0.7	1.9 / 1.3 / 1.9
    0.1-0.2	5.1 / 1.4 / 11.6	4.8 / 1.6 / 11.1	3.3 / 0.4 / 5.6	3.4 / 0.9 / 5.2
    0.2-0.3	7.1 / 1.3 / 17.9	7.0 / 2.0 / 18.1	5.0 / 0.6 / 10.7	4.8 / 1.0 / 8.6
    0.3-0.5	8.1 / 1.2 / 27.5	8.1 / 2.8 / 26.2	6.9 / 0.8 / 15.8	6.5 / 1.8 / 11.7
    >= 0.5	5.2 / 4.2 / 94.0	4.5 / 3.8 / 52.0	6.5 / 0.7 / 25.1	12.6 / 11.5 / 30.4

    Below 0.5 g the dynamic model is 1.3-9x closer than before. The RoboRacer's speed residual now matches the kinematic bicycle's there (0.001-0.035 m/s rms, was 0.018-0.090). What is left is the RoboRacer in the city above 0.5 g, and its yaw rate there.

  • Refits against physics: fit dyn_bike_si_of to Jolt (tires, direct drive); P8b training now helps on Jolt #753's values (50 Hz, fit on seeds 100-102, judged on 103):

    run	drive lag	tire B front / rear	yaw inertia	held-out (m/s)²: default -> fit
    Humvee slalom	0.20 s	x1.45 / x1.20	x1.00	0.00046 -> 0.00049
    Humvee city	0.04 s	x1.00 / x1.00	x1.00	0.00065 -> 0.00061
    RoboRacer slalom	0	x1.00 / x1.00	x0.85	0.00001 -> 0.00001
    RoboRacer city	0	x1.00 / x1.20	x1.75	0.00015 -> 0.00018

    Nothing is left to fit on flat ground: the defaults do as well as a refit on the held-out seed. The Austin fits still land on the grids' edges.

  • Roadmap: both tables are replaced with these. The datasets' rerun is /home/joe/src/cvc/plant-log-runs-2026-10-20/post753/ (local).

  • Tests: physics_plant_log_test 7/7 after the merge.