Continual Adaptation for Autonomous Driving with the Mixture of Progressive Experts Network

TL;DR

The paper tackles the challenge of continual adaptation in autonomous driving by proposing a dynamic progressive optimization framework that fuses reinforcement learning data generation with continual imitation learning. At its core is the Mixture of Progressive Experts (MoPE) network, which expands task-specific expert branches over time, uses a gating mechanism to allocate tasks to relevant experts, and employs lateral connections to share historical knowledge while freezing prior components to prevent forgetting. Through experiments in MetaDrive simulating urban driving with complex scenarios, MoPE demonstrates strong scenario adaptation and case generalization, outperforming baseline continual learning methods and achieving up to 7.8% performance gains in intricate environments. The work offers a scalable pathway for continual learning in autonomous driving, enabling robust knowledge transfer and rapid adaptation to evolving traffic conditions."

Abstract

Learning-based autonomous driving requires continuous integration of diverse knowledge in complex traffic , yet existing methods exhibit significant limitations in adaptive capabilities. Addressing this gap demands autonomous driving systems that enable continual adaptation through dynamic adjustments to evolving environmental interactions. This underscores the necessity for enhanced continual learning capabilities to improve system adaptability. To address these challenges, the paper introduces a dynamic progressive optimization framework that facilitates adaptation to variations in dynamic environments, achieved by integrating reinforcement learning and supervised learning for data aggregation. Building on this framework, we propose the Mixture of Progressive Experts (MoPE) network. The proposed method selectively activates multiple expert models based on the distinct characteristics of each task and progressively refines the network architecture to facilitate adaptation to new tasks. Simulation results show that the MoPE model outperforms behavior cloning methods, achieving up to a 7.8% performance improvement in intricate urban road environments.

Figures (7)

• Figure 1: Overview of dynamic progressive optimization framework for autonomous driving. Reinforcement learning experts generate high-quality, multi-scenario data in simulation, which is then aggregated by the Continual Model using imitation learning. The model progressively expands and dynamically weights expert modules to adapt to new scenarios or cases while retaining past knowledge, ensuring robust performance in complex traffic environments.
• Figure 2: The diagram of the Mixture of Progressive Experts.
• Figure 3: The performance of the proposed method is evaluated against baseline models in scenario adaptation, including unprotected left turns (a), sharp circular curves (b), and multi-lane roundabouts (c).
• Figure 4: Results of the proposed algorithm in case generalization. (a) Dynamic traffic flow featuring typical scenarios described in Section IV.B, each labeled with timestamps. The green rectangle is the ego vehicle. (b) Comparison of the proposed method's performance with baseline models.
• Figure 5: The ablation result shows MoPE surpasses PEN, which plateaus without the mixture of experts integration.