Supplementary Videos

HOCBF-Informed Constrained Reinforcement Learning with Adversarial Training for Robust Autonomous Driving

Autonomous navigation Constrained reinforcement learning Control barrier functions Adversarial training

Evaluation under Ideal State Observations

Six navigation tasks. Two cases per task. A comparison of longitudinal driving behavior
with and without the HOCBF-based safety filter.

HC2RL w/o SF

Each video combines the intersection view, a traffic schematic, and 3D views.

01

Navigation Task L1

Left turn to target lane 1

Method
Case 1
Case 2
HC2RL
w/o SF
02

Navigation Task L2

Left turn to target lane 2

Method
Case 1
Case 2
HC2RL
w/o SF
03

Navigation Task S1

Go straight to target lane 1

Method
Case 1
Case 2
HC2RL
w/o SF
04

Navigation Task S2

Go straight to target lane 2

Method
Case 1
Case 2
HC2RL
w/o SF
05

Navigation Task R1

Right turn to target lane 1

Method
Case 1
Case 2
HC2RL
w/o SF
06

Navigation Task R2

Right turn to target lane 2

Method
Case 1
Case 2
HC2RL
w/o SF

Evaluation under Perturbed State Observations

Two left-turn navigation tasks. Two cases per task.
Evaluated with HC2RL-MC under truncated Gaussian noise perturbations.

HC2RL-MC
07

Navigation Task L1

Left turn to target lane 1

Method
Case 1
Case 2
HC2RL-MC
08

Navigation Task L2

Left turn to target lane 2

Method
Case 1
Case 2
HC2RL-MC
09

Notation

Main symbols and their definitions

Table of main symbols and their definitions, grouped into constrained reinforcement learning, reward and safety cost, vehicle dynamics and HOCBF-based safety filtering, reward shaping, observation perturbations, and surrounding traffic.