RL environments and trained agents in CARLA using RLlib
- CARLA: This project uses version 0.9.11
- rllib-integration: The integration API is taken from here (with minor changes)
- Task: Follow the lane with given target speed. Episode ends if ego-vehicle changes the lane or stays idle for given number of steps or travells 200m.
- Observations: Camera image (grayscale), steering, throttle, speed, number of steps the vehicle has been idle
- Actions: Steer right/left, speed up/down
- Reward: If travelling at or below target speed, reward is same as distance travelled in the last step. If travelling above target speed, reward is zero. Reward of -1 if the ego-vehicle changes the lane or stays idle for given number of steps.
- After training for 350k timesteps, the agent behaves as follows:
- To train the agent,
python dqn_train.py dqn_example/dqn_config.yaml --name dqn --stop-timesteps 350000- To run the included trained agent,
python dqn_inference_ray.py dqn_example/dqn_config.yaml _data/checkpoint_000170/checkpoint-170 --num-episodes 10