Correcting Autonomous Vehicle Behavior to Ensure Rule Compliance

Accepted at ICRA 2026

[Paper (coming soon)] | Pre-print | Jupyter Notebook | License: MIT

Authors: Felipe Toledo, Trey Woodlief, Sebastian Elbaum, and Matthew B. Dwyer
University of Virginia — William & Mary

---

  

_{Left: Interfuser Baseline running a red light  |  Right: Interfuser + M4PC (GT) stopping at red light}

---

Abstract

M4PC is a method to correct autonomous vehicle behavior at runtime to ensure compliance with traffic rules. We evaluate on CARLA with three state-of-the-art AV stacks: TCP, InterFuser, and Pylot. Results show consistent improvements in driving score and reductions in infractions such as collisions, red-light violations, and stop-sign violations.

Results

All metrics are reported as mean ± std. Higher driving score is better; lower values are better for all infraction metrics.

TCP

Treatment	Driving Score	Coll. Ped.	Coll. Veh.	Red Light	Route Timeout	Stop Sign	Veh. Blocked
Baseline	76.86 ± 23.43	0.00 ± 0.00	0.10 ± 0.36	0.06 ± 0.24	0.00 ± 0.00	1.10 ± 1.23	0.00 ± 0.00
T4PC	78.23 ± 21.76	0.00 ± 0.00	0.12 ± 0.33	0.00 ± 0.00	0.00 ± 0.00	0.94 ± 1.13	0.02 ± 0.14
M4PC (SGG)	79.99 ± 27.97	0.00 ± 0.00	0.22 ± 0.55	0.06 ± 0.31	0.00 ± 0.00	0.20 ± 0.40	0.08 ± 0.27
M4PC (GT)	87.16 ± 24.76	0.00 ± 0.00	0.04 ± 0.20	0.06 ± 0.24	0.00 ± 0.00	0.12 ± 0.33	0.08 ± 0.27

InterFuser

Treatment	Driving Score	Coll. Ped.	Coll. Veh.	Red Light	Route Timeout	Stop Sign	Veh. Blocked
Baseline	59.41 ± 32.28	0.26 ± 0.49	0.64 ± 0.88	0.42 ± 0.67	0.06 ± 0.24	0.00 ± 0.00	0.00 ± 0.00
T4PC	64.05 ± 35.47	0.18 ± 0.39	0.72 ± 1.07	0.28 ± 0.50	0.18 ± 0.39	0.00 ± 0.00	0.00 ± 0.00
M4PC (SGG)	60.14 ± 32.37	0.12 ± 0.39	0.54 ± 0.79	0.42 ± 0.64	0.04 ± 0.20	0.00 ± 0.00	0.12 ± 0.33
M4PC (GT)	69.74 ± 32.42	0.10 ± 0.36	0.34 ± 0.59	0.04 ± 0.20	0.02 ± 0.14	0.00 ± 0.00	0.26 ± 0.44

Pylot

Treatment	Driving Score	Coll. Ped.	Coll. Veh.	Red Light	Route Timeout	Stop Sign	Veh. Blocked
Baseline	68.96 ± 26.70	0.00 ± 0.00	0.32 ± 0.65	0.54 ± 0.86	0.00 ± 0.00	0.44 ± 0.54	0.02 ± 0.14
M4PC (SGG)	75.38 ± 27.05	0.00 ± 0.00	0.24 ± 0.48	0.52 ± 0.99	0.04 ± 0.20	0.24 ± 0.43	0.00 ± 0.00
M4PC (GT)	80.53 ± 24.03	0.00 ± 0.00	0.20 ± 0.45	0.38 ± 0.64	0.12 ± 0.33	0.00 ± 0.00	0.00 ± 0.00

Reproducing Results

The analysis in this repository uses the raw experiment results stored in exp_results/. Each benchmark (TCP, InterFuser, Pylot) contains JSON result files for multiple approaches (Baseline/original, M4PC with GT, M4PC with SGG, and T4PC where applicable).

To reproduce the tables above, run results.ipynb. The notebook loads the JSON results, computes mean and standard deviation per approach, and produces the summary tables.

Citation

@inproceedings{m4pc-icra2026,
  title     = {Correcting Autonomous Vehicle Behavior to Ensure Rule Compliance},
  author    = {Felipe Toledo and Trey Woodlief and Sebastian Elbaum and Matthew B. Dwyer},
  booktitle = {ICRA},
  year      = {2026}
}

License

MIT License. See LICENSE for details.