Hi, great work you have presented so far with the CommonRoad-RL framework! I have a few questions.
Does this framework provide low-level controllers to track planned trajectory, e.g. PID?
You also end the CommonRoad-RL paper with the following future plans:
closed-loop simulation: we will add simulation to the
existing interface with the traffic simulator SUMO [4],
an interface to CARLA, and integrate behavior models
directly into CommonRoad-RL;
I’m not quite sure what you mean with the SUMO and Carla interfaces. Could you maybe elaborate? Is it similar to the existing SUMO <-> Carla co-simulation feature provided from Carla’s side?
The behavior models you’re mentioning: Does this mean you will try to get rid of SUMO for non-ego traffic control via native behavior models inside Commonroad, so implementing car following (e.g. IDM)/lane change models? If so, will it face similar issues as the the lane-change model in SUMO (very unrealistic physics, vehicles floating sideways and stuff)?
And lastly: Do you have a general timeframe on when these features are planned for release?
I’m very very excited about the future of this work! Thanks for your time!
Thanks a lot for your interest in our work. For your questions:
We have two kinds of action space. One is to directly plan the control inputs of the selected vehicle models. The other one is to plan high-level driving decisions such as turn left. We will provide a sampling-based low-level trajectory planner instead of a low-level controller for this mode. The trajectory planner is based on this paper. In both kinds we assume that the vehicle follows the planed motions perfectly so a controller is not needed.
By the SUMO interfaces we meant that we will control the other vehicles in CommonRoad-RL through our existing CommonRoad-SUMO interface. We are also developing interface to the CARLA simulator. Currently I have implemented this part for evaluating a trained model (not released yet) but I haven’t tried it for the training so I cannot estimate the runtime performance.
If the training time is not acceptable with the SUMO interface, we would integrate simple driver models such as IDM, which will also face the similar issues such as unrealistic physics etc. We are also doing research on using imitation learning methods to learn naturalistic driving behaviors. If the results are good there, we will also integrate them in the CommonRoad-RL framework to control other vehicles.
The closed-loop evaluation in SUMO will be released really soon. I cannot estimate the time for the CARLA interface as it depends on the progress of CommonRoad-CARLA interface. If you are interested in using CARLA to train/evaluate instead of SUMO, I suggest you to post another question under the interactive scenarios session. The responsible developer will answer you there.
