Third Place Solution in Fall 2024 Simulation Racing Series
Written by Sid Arutla, Jaiveer Gill, Vivan Doshi, and Max Manolov from Bellarmine College Preparatory.
GitHub: https://github.com/sidarutla/ROAR_Fall2024
Regulated Pure Pursuit Documentation: https://docs.nav2.org/configuration/packages/configuring-regulated-pp.html
Table of Contents
Introduction
Our team achieved 3rd place in the ROAR Fall 2024 Simulation Racing Series using a combination of the Regulated Pure Pursuit (RPP) controller and dynamic path optimization. This post details our technical approach, key innovations, and performance metrics.
Regulated Pure Pursuit Implementation
We implemented multiple Regulated Pure Pursuit algorithms to balance aggressive path tracking with stability at high speeds. One innovation was a new adaptive lookahead distance that scales linearly with velocity:
We also clamped the output value between 3-15 meters. This new lookahead was crucial for Monza’s long Curva Parabolica where traditional fixed lookahead values previously caused oscillations:
In our controller, the path-following process uses consecutive waypoint triples to estimate local curvature, followed by velocity regulation based on a simplified geometric relationship. We found that using three-point Menger curvature calculations with adaptive waypoint spacing produced stable behavior at track section transitions. We then implemented velocity regulation based on a square root relationship with turning radius derived from the simplified curvature formula:
Lateral Control System
Our lateral controller combines heading error correction with curvature-based steering limitation. The steering angle calculation:
Where L is lookahead distance, 𝛼 is heading error, and d is waypoint distance. Implemented with velocity-sensitive scaling:
Key innovations in our system include several advanced techniques to enhance stability and performance. Path curvature anticipation is achieved using a 3-point radius calculation, allowing the controller to predict and adjust for turns effectively. Velocity-dependent steering gain is implemented with a base gain of 0.5, paired with slight section-specific adjustments. Additionally, a minimum turning radius enforcement of 0.9 meters prevents oversteer. The regulated_linear_scaling implementation helped maintain high speeds through challenging sections like the Lesmo curves, where traditional controllers often lose time.
Throttle Control System
Our throttle control uses a multi-layered approach to optimize speed and handling. First, physical limit calculations ensure the vehicle is commanded within the possible parameters. Next, our lookahead system enables the vehicle to anticipate and prepare for track conditions ahead. Finally, curvature-regulated speed scaling adjusts speeds dynamically based on track geometry.
The velocity regulation system includes several key components. Curvature-based speed reduction adjusts the speed proportionally to the turn radius, calculated as the inverse of curvature. As usual, section-specific friction coefficient (µ) tuning is applied to optimize traction and handling in various zones.
Future Improvements
One of the most exciting potential improvements for our racing solution lies in the application of Reinforcement Learning (RL) for hyperparameter tuning. Currently, our system relies on manually tuned parameters for steering multipliers, throttle control, and friction coefficients. While these hand-tuned values have proven effective, implementing an RL framework could systematically optimize these parameters across different track sections and racing conditions. We’re particularly interested in using Deep Q-Learning to fine-tune our throttle control parameters and steering multipliers, as these have shown the most sensitivity to track conditions and vehicle dynamics.
Conclusion
Our 3rd place solution in the ROAR Fall 2024 Racing Series demonstrates the potential of the Regulated Pure Pursuit control system. By addressing areas for improvement, such as integrating machine learning, we hope to improve our solution and compete for the top position in the spring. We greatly appreciate the support of Dr. Allen Yang, Mr. Huo Chao Kuan, and the UC Berkeley team throughout this competition, and we’re excited to see how we can continue to progress with our solution in the future.