Core Concepts and Processes

DI-drive decomposes Autonomous Driving framework into two parts: Policy and Env. Policy maps observations to agent actions, and Env simulates the world with an output of new observations. Thus, the simulation iterates and information can be used to update models, visualize environments and evaluate policies.

One policy can have different running modes, in order to be suitable for varies kinds of functions in Deep Learning. DI-drive defines 3 modes in policy: collect, eval and learn, together with collector, evaluator and learner to interact. DI-drive supports the policy to be any part of the entire procedure of Autonomous Driving including perception, decision, planning, controlling, etc., each of which can be a neural network or else.

Env follows the standard definition in Reinforcement Learning. DI-drive use the most commonly defined environment form in gym as well as the modified version in DI-engine. Env contains a simulator to interact with, and have some other modules to process reward calculation, success judgement and so on. Planner is used to get a navigation route for vehicle to follow. It also involves lane selection and collision avoidance. Visualizer can help to review and analyze running status of hero vehicle and environments. It can get a rendering image and instant display of running parameters.

All general Deep Learning methods in Autonomous Driving can be operated with these modules. The core engine is the rolling of observation-action flow. Imitation Learning can be represented as sampling data and storing to a dataset, then using data to train NN in policy. Reinforcement Learning similarly sampling data into replay buffer and updating policy immediately or laterly. Other methods such IRL and GAIL can be promoted from these modules.