Overview / Usage

Urban Driving decision making is challenging due to complex road geometry and multiple agent interactions. Recent Decision making methods are mostly manually designing the driving policy, which yields in suboptimal designs. This application to enable model-free deep reinforcement learning in challenging urban autonomous driving scenarios. Here, State-of-the-art model-free Deep RL algorithms be implemented with several tricks to improve their performance.

Evaluating our method in a most complex scenario roundabout task with various surrounded vehicles in a driving simulator.

Methodology / Approach

Methodology and/or Problem Formulating

Input Representation: it is to directly use the raw sensor data such as front view image. However, the raw sensor data contains extremely high dimensional and complex , with the help of a the Perception module, this complexity can be reduced largely.

Directly applying reinforcement learning algorithms on the raw sensor input can hardly work on the data. To over come this , making it possible to be solved well by the current Model-Free Deep Reinforcement Learning Techniques.

​

Techniques : Deep RL = Deep Learning + Reinforcement Learning (RL)

The Following state-of-the-art Deep RL Algorithms applied in this framework to learn the driving policy.

1. DDQN (Double Deep Q-Network)
2. TD3 (Twin Delayed Deep Deterministic Policy Gradient)
3. SAC (Soft Actor Critic)

Introducing them briefly in the hand-written images in the following sections..

​

Technologies Used

1. CARLA: High-definition open-source simulator for autonomous driving research.
2. Open AI
3. Anaconda,
4. Keras
5. Tensorflow

​

Other links

• Deep RL Course
• Deep Learning Book