[Paper Review] Social Attention for Autonomous Decision-Making in Dense Traffic

Edouard Leurent, Jean Mercat|arXiv (Cornell University)|Nov 27, 2019

Autonomous Vehicle Technology and SafetyEngineering37 references68 citations

TL;DR

The paper proposes an attention-based policy architecture that handles variable numbers of nearby vehicles and permutation invariance to improve autonomous decision-making in dense traffic, with interpretable attention patterns.

ABSTRACT

We study the design of learning architectures for behavioural planning in a dense traffic setting. Such architectures should deal with a varying number of nearby vehicles, be invariant to the ordering chosen to describe them, while staying accurate and compact. We observe that the two most popular representations in the literature do not fit these criteria, and perform badly on an complex negotiation task. We propose an attention-based architecture that satisfies all these properties and explicitly accounts for the existing interactions between the traffic participants. We show that this architecture leads to significant performance gains, and is able to capture interactions patterns that can be visualised and qualitatively interpreted. Videos and code are available at https://eleurent.github.io/social-attention/.

Motivation & Objective

• Motivate learning-based behavioral planning for autonomous driving in dense traffic.
• Critique common state representations (list of features vs. spatial grids) for variable-size inputs and permutation invariance.
• Introduce an attention-based architecture that captures ego-vehicle interactions with multiple nearby vehicles.
• Demonstrate improved performance and interpretability on a complex intersection-crossing task with up to 15 vehicles.
• Show that attention patterns align with intuitive safety and negotiation dynamics.

Proposed method

• Use a DQN framework with a specialized Q-function architecture.
• Encode each surrounding vehicle with a shared linear encoder.
• Apply an ego-attention layer with multiple heads to compute interactions with the ego-vehicle.
• Aggregate attention outputs with a residual connection to the ego encoding.
• Use a linear decoder to produce action values for high-level decisions.
• Demonstrate permutation invariance and variable input size without padding.

Experimental results

Research questions

• RQ1 Can attention-based social interactions improve decision-making in dense traffic over fixed-size representations?
• RQ2 Does ego-focused multi-head attention capture meaningful vehicle-to-ego dependencies and be interpretable?
• RQ3 How does the proposed method compare to list-based and grid-based representations in terms of performance and safety in intersection scenarios?
• RQ4 Can the model exploit interaction patterns and become more or less aggressive depending on road priorities?
• RQ5 Is the approach scalable to scenes with a larger number of vehicles and varying configurations?

Key findings

• An ego-attention architecture yields significant performance improvements over FCN/List and CNN/Grid baselines.
• The model handles variable numbers of vehicles and remains permutation-invariant, improving decision quality in intersection crossing.
• Attention heads specialize to different regions (e.g., left and front/right) and focus more on vehicles that may collide, with attention increasing as targets get closer.
• The policy learns nuanced negotiation of right-of-way and can adapt depending on road priorities, affecting aggressiveness.
• Qualitative visualization shows interpretable interaction patterns and attention dynamics during episodes.