Building Safer Autonomous Agents by Leveraging Risky Driving Behavior Knowledge

The highway-env reinforcement learning tasks provides a good abstract testbed for designing driving agents for specific driving scenarios like lane changing, parking or intersections etc. But, generally these driving simulation environments often restrict themselves to safer and precise trajectories. However, we clearly know that real driving tasks often involve very high risk collision prone unexpected situations. Hence, the autonomous model-free driving agents prepared in these environments are blind to certain low probability traffic collision corner cases. In our study we systematically focus on generating adversarial driving collision prone scenarios with dangerous driving behavior and heavy traffic in order to create robust autonomous agents. In our experimentation we train model free learning agents with additional collision prone scenario simulations and compare their efficacy with regular simulation based agents. Ultimately, we create a causal experimentation setup which successfully accounts for the performance improvements across different driving scenarios by utilizing learning from risky driving situations.