Candy: Self-driving in Carla Environment.
What is candy? A model with the structure: Hierarchical Observation -- Plan&Policy -- Hierarchical Actions
(This project is still working in progress.)
- Download Carla-0.8.2 from here.
- Start CarlaUE4 engine in server mode, using command from here.
- Run python main.py.
- Use C3D for observation abstraction.
- Combining imitation learning and reinforcement learning.
- Using MCTS for planning.
- Plenty of auxiliary tasks for robust hidden representation.
- Persistent training process and flexible architecture.
- Policy embedding? Curiosity, Attention, Memory ?