Reproduce “Vision Transformer for Learning Driving Policies in Complex Multi-Agent Environments” on OpenResearch

About this paper

Eshagh Kargar, Ville Kyrki

Researchers from Aalto University developed a Deep Q-Network agent employing a Vision Transformer (ViT) as its backbone to learn autonomous driving policies from bird's-eye-view images in complex multi-agent urban simulations. This ViT-based approach achieved superior evaluation rewards and a significantly lower crash percentage (5% for ViT-DINO) compared to ConvNet-based methods and other state-of-the-art reinforcement learning algorithms, demonstrating enhanced contextual understanding and safety.

Key insights

• Vision Transformers are effective at capturing global context and dynamically attending to salient features in complex multi-agent driving scenarios, outperforming traditional ConvNets.
• Self-supervised pre-training (e.g., DINO on ImageNet) for the ViT backbone substantially improves learning efficiency, leading to higher performance with less training data.
• The inherent attention mechanism of ViTs provides interpretability, demonstrating the model's ability to focus on critical vehicles and road areas relevant for decision-making without explicit supervision.

Summary by alphaXiv.