rl · controls
↗
Reinforcement-learning agent that adapts the Look-Ahead Distance of a pure-pursuit controller in real time, improving path-tracking accuracy across variable curvature and speed profiles.
RLPure-PursuitMATLAB
- Problem
- Pure pursuit steers toward a point a fixed distance ahead. A short look-ahead tracks tightly but oscillates; a long one is smooth but cuts corners.
- Approach
- An RL agent picks the look-ahead distance at every step. We trained DDPG, DQN and PPO variants in MATLAB's Reinforcement Learning Toolbox, building on Goel & Chauhan (2021).
- My part
- Proposed the idea, surveyed prior work, and wrote the MATLAB implementation in a team of four. The same idea later went into our lab's autonomous race car.