🏎️ Self-Driving Car in CARLA Simulator

This project implements a self-driving car pipeline in the CARLA simulator using Imitation Learning (IL) and Reinforcement Learning (RL). The system evolves from expert data collection to Behavior Cloning (BC) with a custom ResNet18 architecture, followed by fine-tuning using Soft Actor-Critic (SAC). A ZeroMQ-based client-server setup connects the agent to the CARLA environment.

🚀 Project Overview

✔️ Stages:

1. PPO Baseline (RL)

   • Initial experiments using Proximal Policy Optimization (PPO) to validate environment setup and agent control loop.

2. Behavior Cloning (Imitation Learning)

   • Collected expert data with RGB + segmentation images and control actions (steer, throttle, brake).
   • Trained a custom ResNet18 model with 5-channel input (RGB + Seg) as a multi-output regressor.

3. SAC Fine-Tuning (RL with IL Warm Start)

   • Used the pretrained BC model to warm-start a Soft Actor-Critic (SAC) agent.
   • RL fine-tuning improved robustness and long-term performance.

• ✅ Multi-stage learning: BC → SAC
• ✅ Custom ResNet18 with 5-channel input (RGB + Segmentation)
• ✅ Modular architecture using ZeroMQ for different environment communication
• ✅ Multi-output regression for steering, throttle, and brake

🖼️ Data Collection

Expert data includes:

• rgb_image/: Front camera RGB images
• seg_image/: Semantic segmentation images
• logs/logs.json: GNSS, IMU, speed, control actions (steer, throttle, brake)

📦 Installation

# Clone repo
git clone https://github.com/1Kaustubh122/Self_Driving_Car
cd Self_Driving_Car
cd Self_Driving_Car/Imitation_with_RL_SAC_Self_Drving

## for single env
python train.py

## for multi env
python train_multi.py

## Carla core dumped safe training run: (Only for windows OS)
start_monitor.bat