Human-Like Autonomous Driving on Dense Traffic

TL;DR

This paper introduces a mixture density network-based imitation learning model for autonomous highway driving that mimics human behavior, addressing challenges like generalization and outlier prediction using real traffic data and collision penalties.

Contribution

The paper presents a novel mixture density network approach combined with collision penalties to improve imitation learning for autonomous driving in complex traffic environments.

Findings

Effective in complex highway scenarios

Reduces errors compared to traditional imitation learning

Improves decision-making in diverse traffic conditions

Abstract

This paper proposes a imitation learning model for autonomous driving on highway traffic by mimicking human drivers' driving behaviours. The study utilizes the HighD traffic dataset, which is complex, high-dimensional, and diverse in vehicle variations. Imitation learning is an alternative solution to autonomous highway driving that reduces the sample complexity of learning a challenging task compared to reinforcement learning. However, imitation learning has limitations such as vulnerability to compounding errors in unseen states, poor generalization, and inability to predict outlier driver profiles. To address these issues, the paper proposes mixture density network behaviour cloning model to manage complex and non-linear relationships between inputs and outputs and make more informed decisions about the vehicle's actions. Additional improvement is using collision penalty based on the GAIL model. The paper includes a simulated driving test to demonstrate the effectiveness of the proposed method based on real traffic scenarios and provides conclusions on its potential impact on autonomous driving.