ROAR_Competition

ROAR Simulation Racing Series — Summer 2026

Monza Map v1.1 : Best clean time 320.20s for v14 on 2026-08-21

• Monza Map v1.1 : Best clean time 320.00s for v16 on 2026-09-13

• Monza Map v1.1 : Best clean time 319.55s for v17 on 2026-09-19

Provenance

The controller evolved through discussions and simulation trials with assistance from AI agents, informed with the ROAR past results, and included tuning and contributions adapted from publicly reviewed repositories.

ROAR Monza optimization results

• Era: Summer 2026 v1–v14 rows tested 2026-08-21; v17 tested 2026-09-19
• Map: Monza v1.1
• CARLA client: 0.9.12
• Simulator: 0.9.12-dirty
• Scoring: official evaluate_solution elapsed simulation time for three laps
• Protocol for runtime numbers: fresh CARLA simulator restart per validated run

Version	Three-lap time	Major collisions	Result	Main change
Public starter baseline	1658.90 s	Not instrumented	Finished	Official public starter submission, locally measured; proportional controller with 20 m/s target
v1	579.70 s	0	Finished	Cyclic curvature profile, braking propagation, dynamic lookahead
v2	513.75 s	0	Finished	8.5 m/s^2 lateral limit, 48 m/s profile cap
v3	461.85 s	0	Finished	11.5 m/s^2 lateral limit, 55 m/s profile cap
v4	426.50 s	0	Finished	15.0 m/s^2 lateral limit, 60 m/s profile cap
v5	DNF	0 observed	Off track/stuck	Global 20.0 m/s^2 lateral limit, 20 m/s minimum
v6	DNF	0 observed	Off track/stuck	Interpolated global limit, 18.5 m/s minimum
v7	409.05 s	0	Finished	17.5 m/s^2 lateral limit with 17 m/s minimum
v8	DNF	0 observed	Off track/stuck	25.0 m/s^2 outside critical corners
v9	400.10 s	0	Finished	20.0 m/s^2 outside critical corners; 17 m/s cap in critical corners
v10	398.05 s	0 observed	Finished	Full 83 m/s straight envelope, one-tick gearbox launch, and validated 17.5 m/s^2 braking propagation
v11	327.70 s	0 observed	Finished	Optimized racing line with Menger-curvature, section-specific friction, steering, and speed control
v12	321.65 s	0 observed	Finished	Reactive three-point Menger-radius speed target, section-specific friction and heading-PID gains, a 0.80 hard-brake threshold, and two fixed low-brake stability zones on one optimized path
v13	320.35 s	0 observed	Finished	Dense racing-line tracker with speed-scheduled and distance-based lookahead, multi-radius braking preview, exact 10-section state, a 24-tick Section 3 braking horizon, and localized Section 5 distance_gain=0.24 tuning
v14	320.20 s	0 observed	Finished	Raise the localized Section 5 distance_gain from 0.24 to 0.245 at waypoints 1320-1359, retaining all other v13 controller settings
v16	320.00 s	0	Finished	Bayesian-optimized steering-scale multipliers on Curva Grande (+6.4%), Lesmo 1 (+13.7%), Lesmo 2 (−4.4%), Parabolica (+3.6%), and a 0.029 brake-release boost on single corners
v17	319.55 s	0	Finished	Modularized onto the HKU platform: hand-built racing line with 13 baked late-apex widening windows (only apexes move; the L2 right-LEFT complex ±0.3 m), refined per-section grip ladder (μ 1:3.02, 2:3.45, 3:3.36, 4:3.24, 5:2.98, 7:2.85, 8:2.98; others per measured caps), BRAKE_K 815, and the Section-3 dual throttle model (μ 3.65)

Best observed validated result: 319.55 seconds best print, 319.60 seconds typical (v17), measured across ~30 clean fresh-CARLA runs in the 319.55–319.80 s band, all zero-collision; a four-generation benchmark run of this exact committed configuration printed 319.45 s. That is an 80.7% reduction from the starter baseline and approximately 5.19× faster.

Design overview

The submission is a six-module set (competition_code/, loaded by the unchanged competition runner):

1. submission.py — RoarCompetitionSolution orchestrates the modules once per simulator tick and is the only interface required by the competition runner; it also keeps the section state (ten calibrated ranges used for steering, friction, preview, and brake-recovery gain scheduling), waypoint progress tracking, and launch handling.
2. WaypointLine.py — the hand-built 5,865-point racing line with the 13 late-apex widening windows baked in as raised-cosine left-normal shifts, plus the target-snapping used by most sections.
3. LateralController.py — pure-pursuit bicycle steering (4.7 m effective wheelbase, gain 1.5) toward the section-scheduled lookahead target.
4. ThrottleController.py — multi-preview curvature into grip-limited target speeds via the per-section μ ladder, BRAKE_K-projected backward through the braking-distance model, stateful bang-bang actuation with brake-hold latches and per-section throttle-recovery envelopes, and the Section 3 dual model with its own prediction window.
5. SectionStats.py — per-section live statistics used by the throttle model's prediction logic.
6. SpeedData.py — the speed-recommendation record shared between the throttle model's preview stages.

v17 Technical Analysis

The baked widening windows (the racing line)

WaypointLine.py carries the hand-built 5,865-point racing line (5,593 m) with 13 late-apex widening windows — raised-cosine left-normal shifts, positive = left:

1313:1393:+0.9, 1393:1473:-0.9     sec-2 entry/apex    (R 54.5 -> 63.6 m)
 883: 963:-0.8,  963:1043:+0.8     sec-1 entry/apex    (R 37.4 -> 41 m)
4036:4116:+0.5, 4116:4196:-0.5     sec-6 entry/apex    (R 65 -> 73 m)
1688:1768:+0.3, 1768:1848:-0.3     sec-3 entry/apex    (R 50 -> 55 m)
1790:1815:+0.15                    lesmo2_in compensator (wall-margin restore)
2790:2860:-0.3, 2860:2900:+0.3     sec-4 late-apex pair
2915:2985:+0.3, 2985:3055:-0.3     L2 right-LEFT complex

Only apex-side windows move — exit shifts were measured to fail. Magnitude is per-corner and narrow: each step past these values has a measured crash boundary. The line-grip coupling is exploited: each widened corner's μ was re-raised to its new measured cap.

The per-section grip ladder (ThrottleController)

μ = sqrt-model grip per section: 0: 3.05, 1: 3.02, 2: 3.45, 3: 3.36, 4: 3.24, 5: 2.98, 6: 3.30, 7: 2.85, 8: 2.98, 9: 2.10. Every value is one step from a measured crash boundary (S1 3.03 crash-loops, S2 3.50 crashes, S4 3.29 departs at the L2 turn-in, the S3 new-model 3.7 crashes at the lesmo_exit). BRAKE_K 815 sizes the backward braking projection (down from 825 — the L2 turn-in margin).

Measured performance

Metric	Value
Campaign band (~30 clean fresh-CARLA runs)	319.55–319.65 s, best print 319.55 s
Lap structure	110.25 s standing / 104.35 s flying / 104.85 s flying
Top speed	257 km/h (drag-limited)
Per-section traverses (flying lap)	S0 21.4 · S1 9.1 · S2 8.0 · S3 16.8 · S4 4.5 · S5 7.9 · S6 11.4 · S7 1.9 · S8 15.3 · S9 7.0 s

v16 Telemetry Analysis: Anatomy of a 320-Second Lap (historical — the v16-era predecessor)

Full interactive report with SVG charts: docs/v16_telemetry_report.html (open in a browser — self-contained, light/dark themes, no dependencies) This analysis was measured on the v16 stock-line configuration; v17 (this branch) superseded it by moving the line itself — see v17 Technical Analysis above for what changed and why.

Per-tick telemetry analysis of the v16 best-config 3-lap run (320.00s, 0 collisions) on the Monza simulation circuit. Flying lap 2 data (2,091 ticks, 0.05s timestep).

Summary metrics

Metric	Value
3-lap total	320.00 s
Flying lap (lap 2)	104.50 s
Top speed	71.5 m/s (257 km/h)
Track length	5,588 m
Average speed	53.5 m/s
Collisions	0
Throttle usage	93% WOT, 7% brake, 0% coast

The grip ceiling

The car is scrub-limited, not grip-limited. At every corner apex the car is at full throttle (throttle = 1.0) yet achieves only a_lat ≈ 22 m/s² — well below the tire grip limit of 33 m/s². This 37% unused grip headroom is the single largest source of lost time.

Metric	Value	% of tire limit
Tire grip limit	33.0 m/s²	100%
Max a_lat achieved	26.8 m/s²	81%
P90 sustained a_lat	20.7 m/s²	63%
Unused grip headroom	—	37%

Corner-by-corner analysis

Each corner's geometry, achieved speed, lateral acceleration, and grip utilization. The "v@33" column shows the theoretical speed at the tire limit — the gap between achieved v_min and v@33 is the time lost to scrub.

Corner	Sec	R_min (m)	v_min (m/s)	v_max (m/s)	a_lat p90	a_lat max	v@33 (m/s)	Thr mean	Brk mean	Type
Rettifilo	0	74	27.1	64.3	11.7	13.4	78.3	1.00	0.000	Chicane
Biassono	1	51	34.9	68.4	22.3	25.5	41.0	0.86	0.146	Straight+
Curva Grande	2	64	39.6	52.4	25.2	26.6	45.9	0.97	0.030	Single
Roggia	3	70	41.2	68.6	18.4	25.3	48.1	0.98	0.026	Chicane
Lesmo 1	4	84	43.8	70.9	22.0	23.9	52.7	0.80	0.206	Single
Lesmo 2	5	99	45.3	58.7	20.4	22.3	57.1	1.00	0.000	Single
Ascari	6	75	43.6	71.2	25.7	26.8	49.7	0.94	0.068	Chicane
Vialone	7	77	43.3	47.5	20.3	24.6	50.4	1.00	0.000	Chicane
Straight	8	203	47.6	71.3	8.0	11.2	82.0	1.00	0.000	Straight
Parabolica	9	25	24.7	71.5	24.6	26.2	28.7	0.79	0.211	Single

The scrub signature: at every single-corner apex (Curva Grande, Lesmo, Ascari, Parabolica), throttle mean is 0.79–1.00 and brake mean is 0.00–0.03 — yet a_lat plateaus at 20–27. The car is wide open at the apex and still can't reach the tire limit. The v@33 column shows how much faster each corner could theoretically be: Parabolica could do 28.7 m/s (vs achieved 24.7), Curva Grande 45.9 (vs 39.6). That 4–6 m/s gap at every corner is the time the scrub limit costs.

The throttle/brake trace: bang-bang with no trail-brake

The car is WOT (throttle = 1.0, brake = 0.0) 93% of the time. Braking occurs in sharp 1–2 second pulses before corners, then immediately releases to full throttle — even as the car continues to decelerate through the apex. This "WOT at apex" behavior is the scrub: the car can't sustain the planned speed, so it scrubs off 5–10 m/s at full throttle. There is no trail-brake or lift-off coast phase — the car is either full-throttle or full-brake, never in between.

The steering command trace

Chicanes (Rettifilo, Roggia, Ascari, Vialone) produce sharp left-right direction reversals where any controller perturbation crashes. Single corners (Curva Grande, Lesmo, Parabolica) produce sustained steering deflections where the v16 steering-scale tuning has effect. The per-section multipliers (Curva Grande +6.4%, Lesmo 1 +13.7%, Lesmo 2 −4.4%, Parabolica +3.6%) adjust authority on these sustained deflections without touching the chicane transitions.

Why 320 seconds is the ceiling hard to break

The grip-limit lap (all apexes at v@33) would take ~96 seconds. The achieved flying lap is 104.5 seconds. The 8.5-second gap is almost entirely scrub — the car enters each corner at v_target (above sustainable grip), releases to WOT, and scrubs 4–6 m/s before the apex. No steering-gain or brake-release tuning can close this gap because it doesn't change v_target. Breaking 320 seconds requires a controller that sustains a_lat = 33 at the apex — which means model-predictive brake/throttle timing, not reactive bang-bang.