This repository documents the design, implementation, ablation, and rigorous benchmarking of two distinct AI approaches for autonomous racing in Unity: a deterministic waypoint-based heuristic controller and a deep reinforcement learning (DRL) agent built with Unity ML-Agents (PPO). The work provides an academically robust framework for comparing classical and modern approaches emphasizing reproducibility, open science, and in-depth technical documentation for research.

Try the interactive playtest prototype on Itch.io: Play SPEED Racing Agents Demo

Experience the Enhanced Natural DRL agent in action, compare its behavior with your gameplay, and see the research findings firsthand.

Heuristic Agent	Natural DRL Agent
Characteristics:	Characteristics:
• Precise waypoint following	• Adaptive behavior learning
• Deterministic responses	• Human-like smoothness
• Collision-free navigation	• Dynamic decision making

Autonomous racing is a prime testbed for intelligent agent research. It requires tight integration of perception, sequential policy, real-time optimization, and adaptability to diverse environmental conditions. This project addresses critical open questions:

• When and where do engineered heuristic controllers outperform or underperform RL agents? • How does reward shaping and neural architecture affect RL agent learning, stability, and generalization? • Can DRL agents demonstrate truly human-like behavior, as measured by both efficiency and naturalness?

• Heuristic Controller: Fully deterministic, high performance waypoint follower. Adaptively controls speed and heading for efficient, collision-free lap completion. All parameters and logs are documented for reproducibility. • DRL Racing Agent: PPO-based agent using raycast-derived sensor input, curriculum learning, and custom reward structure. Multiple ablation studies highlight the trade-offs in architectural depth, reward complexity, and agent robustness. • Human Benchmark Suite: Tracks and scenes support human demos for direct comparison like lap times, steering smoothness, and behavioral metrics. • Open Experimental Pipeline: Source code, configuration files, trained models, raw logs, and exhaustive documentation are provided to enable full reproduction of all reported findings.

SPEED-Intelligent-Racing-Agents/
├── Heuristic-Agent-Project/     # Deterministic controller (Unity 2021.3.45f1)
│   ├── Assets/
│   ├── ProjectSettings/
│   └── Packages/
├── DRL-Agent-Project/           # PPO agent, custom reward, evaluation scenes
│   ├── Assets/
│   ├── ProjectSettings/
│   ├── Packages/
│   └── Trained-Models/
├── Results-and-Analysis/        # Performance logs, comparison data, CSVs
├── Documentation/               # Research paper, user/setup guides
│   ├── SPEED-Dissertation.pdf
│   └── Setup-Instructions.md
└── README.md

• Game Engine: Unity 2021.3.45f1 LTS • ML Toolkit: Unity ML-Agents 0.30.0 (PPO algorithm) • Programming: C# (Unity scripting), Python 3.8+ • Hardware: 8GB+ RAM, CUDA GPU recommended for DRL • Visualization: Unity scene GUI, TensorBoard, custom plots • Versioning: Git, atomic commits for all experiments

• Unity 2021.3.45f1 LTS (mandatory version) • Python 3.8+ with pip • ML-Agents 0.30.0 (pip install mlagents==0.30.0)

1. Clone the repository:

git clone https://github.com/Sujyeet/SPEED-Intelligent-Racing-Agents.git
cd SPEED-Intelligent-Racing-Agents

2. Open projects in Unity Hub: select either Heuristic-Agent-Project or DRL-Agent-Project. • Only the folders Assets, Packages, and ProjectSettings are strictly required. • On first open, import will take several minutes due to dependency resolution.

3. (Optional) Set up a Python environment for training:

python -m venv racing_env
# Windows: racing_env\Scripts\activate
# Mac/Linux: source racing_env/bin/activate
pip install -r requirements.txt

• Load Heuristic-Agent-Project in Unity • Select a race scene (e.g., Assets/Karting/Scenes/MainScene.unity) • Press Play to run the agent; logs are saved for every episode (Paste the desired output directory in the editor)

• Open DRL-Agent-Project in Unity, ensure ML-Agents is present • Assign any .onnx model from Trained-Models/ to the kart agent in test scenes • Set Behavior Type to "Inference Only" and press Play to visualize evaluation • Retraining: See Documentation/Setup-Instructions.md for hyperparameters, scripts, and training protocol

• Scenes allowing manual play (keyboard/controller) are included, with instructions in the UI • Run with identical evaluation pipeline to ensure compatibility of all recorded metrics

• Evaluation on fixed-seed, multi-lap racing tracks for statistical robustness • Lap time, completion rate, collisions, steering profiles, and path deviation are recorded per episode

• Reward structure: From complex (multiple terms) to minimal (progress, smoothness, speed) • Network architecture: 3-layer/256 vs. 2-layer/128 PPO; training stability and convergence tracked • Simulation time scale & curriculum: Real-time vs 10x acceleration; effect on sample efficiency • Generalization: Evaluations on alternate tracks for out-of-distribution robustness • Result: Simpler reward and lightweight networks consistently improved DRL reliability. See main paper for summary tables (e.g., Table IV, Table V) and cumulative reward learning curves.

Beyond speed, the work quantifies "natural" behavior: • Path deviation and steering variance—plotted time series show DRL agents closely track human smoothness, unlike the mechanical, deterministic baseline • Error and recovery: Analysis of wall-collisions and recovery strategies • Comprehensive logging: All logs available for reproduction in Results-and-Analysis/

• Over-complexity (deep networks, multi-term rewards) led to instability and slow convergence • Well-tuned heuristics provide reliability but limited adaptability to novel track layouts • Simplified DRL with targeted reward shaping yields both superior lap times and higher human-likeness scores in controlled experiments

• All experiments use documented seeds and config files; every result in the paper can be regenerated from scripts in this repository • Paper draft: Documentation/SPEED-Dissertation.pdf (complete analyses, figures, and extended results) • Installation and reproducibility guide: Documentation/Setup-Instructions.md (step-by-step for every OS)

• Academic collaboration: Designed for MSc and PhD research use; contact author for dataset sharing, integration studies, or advanced experiments • Contributing: Issue tracker and pull requests are welcome for reproducibility, cross-validation, or improved agent design studies

• Sujit G (MSc Autonomous Racing, [sujitvarma70@gmail.com]) • Queen Mary University of London (for academic coordination) • For technical questions, open a GitHub Issue or contact by email

All project source, code, models, data, and documentation are released under the MIT License. See LICENSE for details. Intended for research, education, and open collaboration.