Abstract

This work studies the design of safe control policies for large-scale non-linear systems operating in uncertain environments. In such a case, the robust control framework is a principled approach to safety that aims to maximize the worst-case performance of a system. However, the resulting optimization problem is generally intractable for non-linear systems with continuous states. To overcome this issue, we introduce two tractable methods that are based either on sampling or on a conservative approximation of the robust objective. The proposed approaches are applied to the problem of autonomous driving.

Video

Reproduce the experiments

Install requirements

• The environment: highway-env
• The agents: rl-agents

Run the benchmark

cd <path-to-rl-agents>/scripts/
python experiments.py benchmark configs/RoundaboutEnv/benchmark_robust_control.json \
                      --test --episodes=100 --processes=4

The following agents will be evaluated:

{
    "environments": [
        "configs/RoundaboutEnv/env.json"
    ],
    "agents": [
        "configs/RoundaboutEnv/agents/DeterministicPlannerAgent/baseline.json",
        "configs/RoundaboutEnv/agents/DeterministicPlannerAgent/assume_random_route.json",
        "configs/RoundaboutEnv/agents/DiscreteRobustPlannerAgent/routes.json",
        "configs/RoundaboutEnv/agents/DeterministicPlannerAgent/assume_random_behaviour.json",
        "configs/RoundaboutEnv/agents/IntervalRobustPlannerAgent/behaviours.json"
    ]
}