Learning to drive in CARLA v0.8.2 using reinforcement learning. The code has been tested with Python 3.5 and Tensorflow 1.14.0.
- Install CARLA v0.8.2.
- Install dependencies (cf requirements.txt).
- (Optional) Download pre-trained VAE here.
- Train a control policy for 10000 steps using Soft Actor-Critic (SAC).
python train.py --algo sac -vae <path-to-vae> -n 10000
- Test the trained agent for 2000 steps.
python enjoy.py --algo sac -vae path-to-vae.pkl --exp-id 0 -n 2000
Check enjoy.py for more options.
- Collect images by manually driving the car around the track. Don't forget to store the images captured by the camera on the car into a folder.
python manual_control.py
- Train a VAE.
python -m vae.train --n-epochs 50 --verbose 0 --z-size 64 -f <path-to-recorded-images>
- Explore Latent Space
python -m vae.enjoy_latent -vae <path-to-vae>
With the current reward function used, the car doesn't achieve a smooth control. A way around the is to restrict the maximum change in the steering angle at each step. This is generally acceptable since humans drive in a similar manner. For more information check this awesome post.
Set the following values in config.py.
MAX_STEERING_DIFF = 0.15 # 0.1 for very smooth control, but it requires more steps
MAX_THROTTLE = 0.6 # MAX_THROTTLE = 0.5 is fine, but we can go faster- Most of the code has been adapted from "Learning to Drive Smoothly in minutes".
- Related Paper: "Learning to Drive in a Day".
- Wayve.ai for idea and inspiration.
- Stable-Baselines for DDPG/SAC and PPO implementations.