ml racing line

a reinforcement learning model that predicts a racing line.

a learning sandbox for how vehicle dynamics, simulation, reward design and reinforcement learning fit together in autonomous racing. a ppo agent drives a simplified formula sae-style car around a waypoint track in mujoco.

how it works

a custom gymnasium environment sits between mujoco and stable-baselines3. the agent sees vehicle state and track-relative information, outputs steering and throttle, and is rewarded for forward progress and staying on track, with penalties for failure. a local dashboard tracks training live and every launch is archived with its config, logs, telemetry and checkpoints.

where it stands

earlier long runs, past 600k steps, made little progress, and the archived episodes show the same: in the last 100 logged, the car reached at most 4 of 17 waypoints and never finished the course. the environment and model configuration are being reworked before the next round; results and a write-up will go here once there’s something worth showing.