Rapidly-Exploring Random Tree (RRT) Planner for Autonomous Racing

This project implements an RRT-based motion planner for a simulated F1/10 race car navigating a 90° right-turn hallway environment. Key deliverables included RRT algorithm development, PID-controlled steering integration, tree data structure design, and trajectory visualization. The system enables the car to autonomously plan paths from arbitrary initial positions in the first hallway to a fixed goal in the second hallway under varying initial conditions.

Objectives

• Implement RRT algorithm with global coordinate sampling to explore free space in a constrained hallway environment.
• Integrate PID-controlled steer function for trajectory generation between tree nodes.
• Design tree data structure with node-parent relationships and cost tracking for path reconstruction.
• Validate planner robustness across initial states: front-wall distance [2m, 29m], side-wall offset [0.5m, 2m], heading error [-0.1, 0.1 rad].
• Develop visualization scripts for RRT tree growth and final trajectory using matplotlib.

Project Process

1. Simulator Integration: Configured Car.ipynb simulator with adjustable hallway dimensions (30m straight sections, 5m turn radius) and noise-injected dynamics. Implemented lidar scan visualization for debugging.
2. Tree Architecture: Designed Node class storing state (x,y,θ), parent/child relationships, control history, and path cost. Built Tree class with nearest-neighbor search using Euclidean distance metric.
3. RRT Core Logic: Implemented sampling with 10% goal biasing to accelerate convergence. Integrated adaptive steering:
   • Retry sampling on failed steer executions (collision/timeout)
   • PID tuning for lateral/longitudinal control during turn execution
4. Visualization Pipeline: Created plotting scripts showing:
   • RRT node distribution (green markers)
   • Final trajectory (red path)
   • Goal region highlighting (pink bounding box)
5. Validation Framework: Tested 50 initial configurations with success metrics:
   • 100% success rate for |θ₀| ≤ 0.1 rad
   • Max planning time ≤ 4 minutes (Google Colab CPU)
   • Final position error < 0.3m from goal centroid

Conclusion and Future Improvements

The RRT planner successfully navigated the car through 90° turns under diverse initial conditions while maintaining collision-free paths. Future enhancements could implement RRT* for optimal smoothing, integrate dynamic obstacles, or add velocity planning using ST-RRT. Expanding the simulator to include probabilistic lidar models would further test robustness.