Comparing Physics Effects through Reinforcement Learning in the ARORA Simulator

By testing various physics levels for training autonomous-vehicle navigation using a deep deterministic policy gradient algorithm, the present study fills a lack of research on the impact of physics levels for vehicle behaviour, specifically for reinforcement-learning algorithms. Measures from a PointGoal Navigation task were investigated: simulator run-time, training steps, and agent effectiveness through the Success weighted by (normalised inverse) Path Length (SPL) measure. Training and testing occurred in the novel simulator ARORA, or A Realistic Open environment for Rapid Agent training. The goal of ARORA is to provide a high-fidelity, open-source platform for simulation, using physics-based movement, vehicle modelling, and a continuous action space within a large-scale geospecific city environment. Using four physics levels, or models, to create four different curriculum conditions for training, the SPL was highest for the condition using all physics levels defined for the experiment, with two conditions returning zero values. Future researchers should consider providing adequate support when training complex-physics vehicle models. The run-time results revealed a benefit for experimental machines with a better CPU, at least for the vector-only observations we employed.