AEGIS: Human Attention-based Explainable Guidance for Intelligent Vehicle Systems

Research Background and Issues

• Identified Problems or Challenges:

  1. The decision-making processes of current autonomous driving systems (AIVs) lack explainability and transparency, leading to public concerns about their safety.
  2. Mainstream Deep Reinforcement Learning (DRL) methods, while capable of adapting to new environments through trial-and-error learning, exhibit slow convergence and lack explicit scene understanding and reasoning capabilities.
  3. Although existing studies attempt to improve scene understanding by mimicking driver visual attention, such attention models have not been effectively integrated into autonomous driving systems.
  4. Current learning methods (e.g., imitation learning, deep reinforcement learning) are sensitive to distribution shifts, showing subpar generalization performance in testing scenarios.

• Significance of the Research:

  • Enhancing the explainability and safety of autonomous driving systems is crucial for societal acceptance and practical deployment, especially under the frequent public and regulatory scrutiny of autonomous driving.
  • Learning how to embed human attention patterns into autonomous driving systems can not only improve model interpretability but also accelerate learning, thereby reducing development costs.

• Research Motivation and Related Work:

  • Human drivers' visual attention selectively focuses on task-relevant objects (e.g., vehicles or pedestrians ahead), effectively supporting driving decisions, whereas existing models lack this active attention capability.
  • While imitation learning can directly learn from driver behavior, it has limited adaptability to abnormal situations. DRL, on the other hand, is more robust to distribution shifts but requires vast amounts of data and time to converge.
  • Introducing human attention-guided explainable mechanisms can help resolve the trade-off between performance and explainability.

Proposed Solution

• Proposed Method or Solution: The authors propose a novel framework—AEGIS (Human Attention-based Explainable Guidance for Intelligent Vehicle Systems)—which leverages human visual attention to guide reinforcement learning models to focus on task-relevant key areas.

• Innovative Contributions:

  1. Human Attention Guidance Mechanism: Embedding an attention model trained on human eye-tracking data into the reinforcement learning network to guide focus on task-relevant objects.
  2. Enhanced Explainability: Improving model transparency and explainability by comparing machine attention with human attention.
  3. Large-Scale New Dataset Construction: Collecting a large-scale VR eye-tracking dataset with over 1.2 million frames across six scenarios, marking the first high-fidelity eye-tracking driving dataset collected using a VR simulator.

• Implementation Steps:

  1. Data Collection and Modeling: Using a VR driving simulator to collect eye-tracking data from 20 participants and generating a human attention model.
  2. Reinforcement Learning Architecture Design:
     • Developing a policy network based on a self-attention mechanism.
     • Adding a Kullback-Leibler divergence (KL Divergence) loss to align machine attention with learned human attention.
  3. Reinforcement Learning Training:
     • During initial training, using attention distributions derived from eye-tracking data to guide the network to focus on task-relevant areas.
     • Relaxing the strict constraints of human attention in later training stages, allowing the model to further optimize its attention patterns.
  4. Performance Validation: Testing generalization ability, training efficiency, and explainability across various driving scenarios (e.g., car-following, left turns, occlusions).

• Key Technologies Used:

  1. Self-Attention Mechanism: Enhancing the model's focus on critical features of the input.
  2. Reinforcement Learning Algorithms (e.g., TD3 and PPO) combined with auxiliary losses (e.g., KL divergence) to improve learning efficiency.
  3. Pre-trained lightweight U-Net architecture for predicting human attention.

Research Outcomes

• Specific Achievements:

  1. AEGIS significantly outperformed baseline reinforcement learning methods (Vanilla) and imitation learning-based methods (BC) across various driving scenarios.
  2. AEGIS demonstrated faster convergence rates: a 270% improvement in car-following scenarios and a 150% improvement in left-turn scenarios compared to Vanilla.
  3. In explainability surveys, 80 respondents found AEGIS's attention distribution and decision-making process easier to understand and expressed greater confidence in its safety.

• Advantages over Existing Solutions:

  1. Stronger Generalization Ability: AEGIS exhibited lower performance degradation in unseen scenarios, demonstrating better robustness.
  2. Improved Training Efficiency: With human attention guidance, AEGIS achieved better performance in less training time.
  3. Higher Explainability: Geometric similarity metrics (e.g., CC and SIM) between machine attention and human attention showed significant improvement.

• Experimental or Evaluation Results:

  1. In car-following and left-turn scenarios:
     • Success rates were 62% and 65%, respectively, outperforming the baseline BC method (46% and 23%).
     • Machine attention showed a significantly higher focus ratio on critical objects (e.g., vehicles and pedestrians).
  2. In occlusion scenarios, AEGIS achieved results comparable to the state-of-the-art imitation learning method ReasonNet but provided more interpretable attention distributions.
  3. Quantitative metrics indicated that AEGIS's attention distribution closely matched learned human attention and was more concentrated (lower spatial entropy).

• Limitations and Future Directions:

  1. Limitations:
     • The current study focuses on longitudinal control (braking and acceleration) and lacks research on steering control.
     • Reinforcement learning still relies on manually designed reward functions, requiring improvements in the automation of policy learning.
     • Experiments were primarily conducted in simulated environments, and the generalization ability to real-world scenarios remains to be validated.
  2. Future Directions:
     • Expanding the driving dataset to include more complex scenarios and long-term driving tasks.
     • Applying AEGIS to real-world driving systems and validating its applicability in real-world environments.
     • Developing more general human behavior modeling methods to enhance the model's multi-task adaptability.

Through this research, AEGIS demonstrates that human attention can effectively improve the performance and explainability of reinforcement learning models, offering valuable insights for the explainable design and optimization of autonomous driving systems while exploring new directions for integrating human cognition with machine learning.