How to Design Training Scenarios for RL-based Collision Avoidance in CARLA?

[itroe]

Hi everyone,

I’m currently working on developing a collision avoidance system using reinforcement learning in the CARLA simulator.

To build a robust RL-based logic, I understand that designing appropriate training scenarios is crucial. However, I’m a bit unsure about how scenarios are typically handled in CARLA for reinforcement learning.

From what I’ve found, Scenario Runner seems to be mainly used for evaluation and validation rather than for training. If that’s the case, I’m wondering:

• For RL training, do we usually need to implement custom scenarios manually (e.g., spawning actors, defining traffic situations, etc.)?
• Or are there any commonly used frameworks, environments, or best practices for generating training scenarios in CARLA?

If anyone has experience working on similar projects (especially RL-based collision avoidance or autonomous driving in CARLA), I would really appreciate it if you could share your overall workflow, approach, or any practical tips.

Thanks in advance, and hope you all have a great day!