Vision-Based Autonomous Car Racing Using Deep Imitative Reinforcement Learning

Autonomous car racing is a challenging task in the robotic control area. Traditional modular methods require accurate mapping, localization and planning, which makes them computationally inefficient sensitive to environmental changes. Recently, deep-learning-based end-to-end systems have shown promising results for autonomous driving/racing. However, they are commonly implemented by supervised imitation learning (IL), suffers from distribution mismatch problem, or reinforcement (RL), requires huge amount of risky interaction data. In this work, we present general deep imitative approach (DIRL), successfully achieves agile using visual inputs. The driving knowledge acquired both IL model-based RL, where agent can learn human teachers as well perform self-improvement safely interacting with an offline world model. We validate our algorithm high-fidelity simulation on real-world 1/20-scale RC-car limited onboard computation. evaluation demonstrate that method outperforms previous RL terms sample efficiency performance. Demonstration videos available at https://caipeide.github.io/autorace-dirl/.