morphlng / scenario_runner_rl Goto Github PK

This repository is a fork from the official carla-simulator/scenario_runner. It is meant to be used combined with model/checkpoint trained by our fork of MARLlib.

Currently based on branch tag/0.9.13.

Currently, you can try with:

cd scenario_runner

python scenario_runner.py --host 127.0.0.1 --port 2000 --openscenario srunner/examples/RLTest_Town01.xosc --sync --reloadWorld 

Note: In order to keep the repository small, we will distribute the trained model/checkpoint in Releases. You can download them and put them in the srunner/examples/rl_models folder.

The RL inference support is added as an "extension" to the original scenario_runner, with minimal changes to the original code. This project should be fully compatible to run non-RL scenarios.

1. srunner/extension. Folder for extensions, currently only rl_integrate folder.
2. srunner/examples/rl_models. Folder for RL models/checkpoints.

1. srunner/autoagents/agent_wrapper.py

   • Code formatting and type annotations.
   • Import rl_integrate module. This is done in a try-except block, so that the original scenario_runner can still run without the module.
   • Add support for RLAgent class, which is a wrapper for RL inference.
   • Complement the destruction logic for AgentWrapper class.

2. srunner/autoagents/sensor_interface.py

   • Add support for carla.ColorConverter. This will not change the default behavior of the original scenario_runner.

3. srunner/scenariomanager/carla_data_provider.py

   • Code formatting and type annotations.
   • Add new data pool _actor_velocity_vector_map and _actor_acceleration_map.
   • For all get_xxx functions, add new function called get_xxx_by_id, which takes an actor id as input.
   • Other changes are made for our custom UE4 build, which should not affect the original scenario_runner.

4. srunner/scenariomanager/scenario_manager.py

   • Import rl_integrate module. This is done in a try-except block, so that the original scenario_runner can still run without the module.
   • Register all agents to RL blackboard
   • on each tick, update information for RL (RlAgent.on_carla_tick())
   • add some function to check and enable heterogeneous agents action space padding.

5. srunner/examples/catalogs/ControllerCatalog.xosc

   • Add a new controller catagory "RLControl"