Autonomous Valet Parking with DRL under LiDAR Noise Conditions

This repository contains the MATLAB/Simulink implementation of our research on Autonomous Valet Parking (AVP) using three state-of-the-art Deep Reinforcement Learning (DRL) algorithms: Soft Actor-Critic (SAC), Proximal Policy Optimization (PPO), and Twin Delayed Deep Deterministic Policy Gradient (TD3).

The study evaluates each algorithm's performance under four LiDAR noise scenarios:

1. Without Noise (Baseline)
2. Internal Noise (Gaussian noise from internal LiDAR disturbances)
3. External Noise (Gaussian noise from external light interference)
4. Combined Noise (Internal + External)

A total of 12 experimental scenarios were conducted (3 algorithms × 4 noise conditions).

---

1. Research Overview

• Objective:
  Compare SAC, PPO, and TD3 algorithms for AVP in noisy perception conditions and evaluate robustness, efficiency, and generalization.

• Key Contributions:

  1. Implementation of three DRL algorithms in MATLAB–Simulink.
  2. Four distinct LiDAR noise injection schemes for realistic perception challenges.
  3. Comprehensive evaluation using average reward, success rate, collision rate, and generalization.

• Reference Paper:
  Autonomous Valet Parking Based on Soft Actor-Critic Approach (2025)
  (https://drive.google.com/file/d/1mVdTLA1qDqTndRphbDPnqWv1KBYgANBk/view?usp=sharing)

• Documentation:

  

---

2. Repository Structure

├── Combined-Noise/ # Simulink environment for combined noise scenario

├── External-Noise/ # Simulink environment for external noise scenario

├── Internal-Noise/ # Simulink environment for internal noise scenario

├── Without-Noise/ # Simulink environment for baseline (no noise)

│

├── Camera.m # Camera sensor model

├── LIDARSensor.m # LiDAR sensor model

├── MainProgram-PPO.mlx # PPO main training script

├── MainProgram-SAC.mlx # SAC main training script

├── MainProgram-TD3.mlx # TD3 main training script

├── ParkingLot.m # Parking lot model

├── ParkingLotSimulator.m # Simulator for parking lot environment

├── autoParkingValetParams.m # AVP parameters

├── autoParkingValetResetFcn.m # Reset function for RL environment

├── createMPCForParking.m # MPC creation for search mode

├── getCarSegmentLengths.m # Utility function for geometry

├── getRefTraj.m # Reference trajectory generator

├── lidarSegmentIntersections.m # LiDAR ray intersection calculations

├── parkingVehicleStateFcnRRT.m # Vehicle state function for RRT

├── parkingVehicleStateJacobianFcnRRT.m # State Jacobian for RRT

├── rect2segs.m # Rectangle to segments utility

├── rlAutoParkingValetAgent.mat # Pretrained RL agent (optional)

├── vehicleStateFcn.m # Vehicle kinematic model

├── vehicleStateJacobianFcnDT.m # State Jacobian function

---

3. Prerequisites

• MATLAB 2023a or later
• Required MATLAB Toolboxes:
  • Reinforcement Learning Toolbox
  • Model Predictive Control Toolbox
  • Deep Learning Toolbox
  • Simulink

---

4. Reproducing the Experiments

Step 1 – Select Noise Scenario

• Noise scenario environments are stored in:

  • Without-Noise/ → Baseline, no noise injected

  • Internal-Noise/ → Zero-mean Gaussian internal noise

  • External-Noise/ → Zero-mean Gaussian external noise

  • Combined-Noise/ → Both internal and external noise

• Open the .slx Simulink file inside the chosen scenario folder.

Step 2 – Select Algorithm

• Choose one of the three main training scripts:

  • MainProgram-SAC.mlx → Soft Actor-Critic

  • MainProgram-PPO.mlx → Proximal Policy Optimization

  • MainProgram-TD3.mlx → Twin Delayed DDPG

Step 3 – Select Mode

• In each main program (MainProgram-SAC.mlx, MainProgram-PPO.mlx, MainProgram-TD3.mlx), you will find the following code snippet:

doTraining = true;

if doTraining
trainingResult = train(agent, env, trainOpts);
save("TrainingResult.mat", "agent");
else
load("rlAutoParkingValetAgent.mat", "agent");
end

• Set doTraining = true → The script will start training the agent from scratch using the selected algorithm and environment.

• Set doTraining = false → The script will skip training and instead load a pre-trained agent from the file rlAutoParkingValetAgent.mat or your previous Training Result for testing and evaluation.

Step 4 – Monitor Training

• Episode reward, Average reward, Episode Q0 curves will appear during training.

• Models and logs are stored in MATLAB workspace or saved manually.