liveProject: Reinforcement Learning for self driving vehicles

: Build & Research solutions for autonomous driving agents

This github repository contains implementations of liveProject: Reinforcement Learning for self-driving vehicles for reference purposes. This repository includes two different types of implementations. First, containing agent implementations from scratch and second approach uses stable-baselines or rl-agents like packages to directly import algorithms for fast prototyping to evaluate the performance of the models.

In this liveProject we explore the tools, techniques and methodologies used by AI researchers to quantify the performance of an agent in a given environment to achieve a given task. The use-case that we explore is navigating and performing different tasks across multiple driving environments. We explore different basic and abstract driving environments for prototyping some RL algorithm implementations in this beginner centered up-for-grabs liveProject.

Note: For all the Colab Implementation Links below please create your own copy and proceed with the code editing, execution and testing process.

Agent implementations for milestones

This liveProject is analyzed based on three simple working principles in AI: optimize, explore & simplify. The practical applications of these concepts are elaborated over from Milestone Two, Three & Four. Also, we have an introduction Milestone that focuses primarily on visualization techniques.

Milestone One: Visualization

• Milestone One: Visualization in Reinforcement Learning for problem & result analysis.
  • Getting started with Google Colab environments & notebooks with OpenAI.
  • Introduction to RL with Value Iteration algorithm in simple environments.
    • Limitations for Value Iteration algorithm in Taxi-v3 scenario. Colab Implementation Link
  • Developing Value Iteration for MDP based driving scenarios in the highway-env package. Colab Implementation Link
  • Deliverable One: Comparative result Visualizations for discussed algorithms in MDP based driving scenarios.

 

Figure 1: Lane change and highway merge (additional) tasks plots for reference.

Milestone Two: Optimize

• Milestone Two: Optimizing a given agent's performance to its fullest potential.
  • Setting up Google Colab notebooks for rendering OpenAI environments outputs. Colab Implementation Link
    • Value function approximation implementation with Q-table and exploration for OpenAI's MountainCar-v0 environment. Colab Implementation Link
  • Handling sparse rewards scenarios in goal based environments like MountainCar-v0 environment. Colab Implementation Link
  • Deliverable Two: Evaluating & documenting variations in agent behavior for our hand-built reward functions to develop the most optimized agent in MountainCar-v0 for this approach.

   

Figure 2: Reward Shaping approach output (Left) & its Optimized Iteration output (Right) respectively.

Milestone Three: Explore

• Milestone Three: Exploring new algorithms for achieving better performance in goal based driving tasks.
  • State–action–reward–state–action (SARSA lambda) agent implementation for MountainCar-v0 environment to learn agent designing concepts. Colab Implementation Link
  • Exploring new RL algorithms for goal based parking task on highway-env package with HER and SAC from stable-baselines3 RL algorithms package. Colab Implementation Link
  • Deliverable Three: Implementation, Analysis and Documentation of HER supported algorithms like DDPG, TD3 and SAC for the parking task on third party highway-env package in parking-v0 environment.

 

Figure 3: Mean average reward plot for performance evaluation of SAC, TD3 & DDPG equipped with HER.

Additionally, if you are interested in reviewing the code base of the Tensorflow based stable-baselines package, you can check their earlier archived repository. Also, you can also refer to theTensorflow backend based implementation of HER based SAC agent as well, just to see turnarounds required for running the earlier stable-baselines. Colab Implementation Link

Milestone Four: Simplify

• Milestone Four: Designing plan oriented agents that use searching capabilities to navigate across optimal paths in an environment.

  • Monte Carlo Tree Search (MCTS) Implementation for Toy Text Taxi-v3 environments.

  • Introduction to navigation environments in highway-env package and MCTS agent prototyping for all these navigation based tasks with rl-agents package.

  • Deliverable 3: Creating a MCTS learning and evaluation functional structure to evaluate the agent performance for different environments present in the highway-env package.

    Note: The current generic implementation doesn't work optimally for these different environments in the highway-env package. You can skip the optimization and performance analysis part.

Certification Test

• Certification Test: Final certification test checks the understanding developed by you during the liveProject.

     

Figure 4: Outputs of different trained MCTS agents with the same parameters that are under evaluation in the certification test.

Building Your Own Environment

This is a supplementary section that covers an additional topic of making a custom environment as per one's need.

• Building Custom Environments In this notebook we create a sample environment for highway-env package and use the rl-agent package to create a baseline planner agent corresponding to that environment. Colab Implementation Link

   

Figure 5: U-Turn environment baseline agents namely MCTS Agent (Left) and Deterministic Agent (Right).

Acknowledgements

• Thank you to for providing and maintaining the and packages.
• Thanks to team and   for providing efficient and well documented implementations of RL algorithms in a single .
• Also, thanks to for an interesting implementation use-case of reward shaping.
• Great improvement suggestions were provided by , & . It's because of them this project has reached its current stage. Thank you for all the support & help!