Abstract
Personalization is a key factor in the acceptance of Advanced Driver Assistance Systems (ADAS). Traditional imitation learning approaches for personalizing Automated Lane Change (ALC) systems are computationally expensive and require vast offline datasets. This paper proposes a "lesson learning" framework that enables online, onboard personalization by learning specifically from human takeover interventions. By identifying "perceived safe zones" and refining reward functions through apprenticeship learning, the system achieves successful customization 14 times faster than traditional methods, requiring an average of only 13.8 iterations to align with a driver's unique style.
Introduction
While ALC systems are increasingly common, their "one-size-fits-all" nature often leads to overly conservative behavior that frustrates users, resulting in frequent manual takeovers. The standard solution—imitation learning—attempts to mimic human trajectories by processing massive amounts of naturalistic driving data. However, this approach is computationally intensive and typically occurs offline. There is a critical need for ADAS to evolve "online" and "onboard," adapting to a specific user's preferences quickly without requiring significant computational resources.
The Lesson Learning Framework
The proposed Lesson Learning strategy shifts the focus from matching a human's exact trajectory to avoiding the "perceived unsafe zones" that trigger interventions. The framework consists of three core components:
- Driving Zone Filter: Utilizes Gaussian Discriminant Analysis (GDA) to classify the environment into "perceived safe" and "perceived unsafe" zones based on the spatial-temporal context of user takeovers.
- Driving Experience Aggregator: Employs apprenticeship learning to update the system's reward function. It treats human reactions to unexpected vehicle motions as expert signals to refine the definition of a "good" lane change for that specific driver.
- Lane-Change Trajectory Planner: A Model Predictive Control (MPC) planner optimized via Pontryagin’s Minimum Principle (PMP). It generates personalized paths that remain within the perceived safe zone while adhering to the updated reward function.
The primary innovation is that training is activated only when a takeover occurs, which drastically reduces the data and computational power required for convergence.
Experimental Evaluation
The framework was evaluated using a MATLAB-based simulation platform. The environment featured a 200-meter road section with two lanes, where background vehicles were controlled by the Intelligent Driver Model (IDM) to simulate realistic traffic. Scenarios included various speeds (40–65 mph), congestion levels (headways of 30–50 meters), and distinct driving styles (Aggressive, Neutral, Cautious). Unlike methods relying on external datasets, this study focuses on the real-time generation of personalized data through direct interaction.
Results and Discussion
The lesson learning approach demonstrates significant improvements in both speed and efficiency over traditional imitation learning.
- Evolution Efficiency: The system required an average of 13.8 learning iterations to achieve full customization, making it approximately 14 times faster than conventional baselines.
- Computational Performance: With an average computation time of 0.08 seconds, the framework is highly suitable for real-time onboard implementation.
- Perceived Safety Boundaries: A key insight is that a personalized ADAS does not need to drive exactly like a human; it simply needs to stay within the human's "acceptable domain." By treating takeover points as hard constraints, the system ensures perceived safety more effectively than systems that average human behavior.
Furthermore, the system can carry over "experience" (state-driving zone pairs) to new environments. This "warm start" capability improves evolution efficiency by an additional 24%, allowing the system to adapt even faster in novel scenarios.
Conclusion
The lesson learning framework provides a robust path toward rapidly personalizing ADAS. While currently validated in simulation for highway lane changes, future work will address real-world challenges such as sensor noise, unpredictable actor behavior in dense urban environments, and interactions with pedestrians. Expanding this concept to other features like Adaptive Cruise Control (ACC) and integrating more complex models of human cognition will further enhance the synergy between drivers and automated systems.
Summary updated