DI-drive

Project Description

Implemented macro-level functionalities within the Metadrive simulator' highway environment, employing the DQN algorithm to formulate lane-changing strategies.
Implemented Model Predictive Control (MPC) into the CARLA simulator, seamlessly combining trajectory tracking and path following modes.
Replicated the Disturbance-based Reward Extrapolation (D-REX) algorithm in the Drive environment, enabling the vehicle to acquire effective strateties and surpass provided demonstrations, even in the presence of suboptimal inputs.

Enable Metadrive to behave macro decision making strategies, such as Lane Change Left, Lane Change Right, Accelerating, Deccelerating and Maintain the Current State, like in highway-env.
Adapting Metadrive Simulator to the multi-process training in DI-engine in OpenDILab.
Adapting reinforcement learning algorithms for discrete actions, such as PPO and DQN, to this simulation environment.

In the official Carla platform, there is only a simple PID control moudle available, which limit our ability to perform spatio-temporal planning tasks. For instance, we are unable to spedcify an entire trajectory and demand reaching a designated destination within a predefined time frame.
We designed two modes:
- For Trajectory Planning: Tracking an entire trajectory with specified arrival times at specific locations.
- For Lane-following tasks: Focus on tracking a path without time allocation.
Implementation can be found in this link.

Reproducting a mehotd (D-REX) that utilizes expert data into DI-drive
Enabling the vehicle to acquire effcetive strategies and surpass provided demonstrations, even in the presence of suboptimal inputs.
Implementation can be found in this link.