用预训练编码器提升自动驾驶强化学习的感知能力

Our research investigates the challenges Deep Reinforcement Learning (DRL) faces in complex, Partially Observable Markov Decision Processes (POMDP) such as autonomous driving (AD), and proposes a solution for vision-based navigation in these environments. Partial observability reduces RL performance significantly, and this can be mitigated by augmenting sensor information and data fusion to reflect a more Markovian environment. However, this necessitates an increasingly complex perception module, whose training via RL is complicated due to inherent limitations. As the neural network architecture becomes more complex, the reward function's effectiveness as an error signal diminishes since the only source of supervision is the reward, which is often noisy, sparse, and delayed. Task-irrelevant elements in images, such as the sky or certain objects, pose additional complexities. Our research adopts an offline-trained encoder to leverage large video datasets through self-supervised learning to learn generalizable representations. Then, we train a head network on top of these representations through DRL to learn to control an ego vehicle in the CARLA AD simulator. This study presents a broad investigation of the impact of different learning schemes for offline-training of encoders on the performance of DRL agents in challenging AD tasks. Furthermore, we show that the features learned by watching BDD100K driving videos can be directly transferred to achieve lane following and collision avoidance in CARLA simulator, in a zero-shot learning fashion. Finally, we explore the impact of various architectural decisions for the RL networks to utilize the transferred representations efficiently. Therefore, in this work, we introduce and validate an optimal way for obtaining suitable representations of the environment, and transferring them to RL networks.