HumbleKent/carla_rl ? reverse-engineered prompt

Build me a reinforcement learning project for CARLA that trains a car to avoid traffic cones in the simulator.

I want a working setup where I can launch CARLA, start training with PPO, and then evaluate the best saved model later. The agent should use both camera images and a few driving state values like speed, heading, and cone distance, and it should learn steering plus throttle or brake. Please include the training loop, the custom driving environment, reward logic, and a simple way to run multiple simulator instances for faster training.

It should also have a clear way to define cone scenarios, spawn the vehicle, and visualize what the agent is doing while it runs. Save checkpoints and logs during training, and make it easy to inspect progress with TensorBoard. If anything is unclear about the latest CARLA or Stable Baselines setup, look up the current docs online and adapt it.