F1Tenth Implementation & Vehicle Dynamics Modification

[loitonyngo]

Hi, first of all, I would like to thank the people who are contributing to this project. I am doing my master thesis on reinforcement learning application on the F1Tenth car, and I appreciate what you are doing.
I was wondering what is the state of F1Tenth implementation. What differs from the other two cars? Is it mainly just the URDF model and some config file parameters? What is missing to merge it into the main branch?
Secondly, I am interested in knowing if the car dynamics and physics are constrained to the Isaac Sim physics engine PhysX, or if they can be adjusted or modified (using explicit functions or NN) in the environment/during the simulation.
Thanks again for the effort!

[geiko246]

Hi! Thanks for the question!

Quite a few people have been asking about the F1Tenth integration, so I'll double check on this with @byx4 and see how soon we can have it merged. To answer your question: we are adding mainly the URDF and actuator parameters. Once these are in, it should be able to be trained. We'll verify these things in the actual PR (#6).

Great question! We actually also have been thinking about doing some Real2Sim which would override the sim step. I'm not sure if Isaac Lab natively provides support for this but I would love to be proved wrong. Some things you can do meanwhile:

1. Learn an actuator network - this is standard practice for humanoids and quadrupeds and may be relevant for aggressive mobile robots.

2. Override the sim.step or env.step - this is something we were thinking about doing but would take some work as you have to call the PhysX API to set the agent state (assuming you also want to do sensor simulation). In theory, this is totally possible but I'm not aware of anything so far that has done it.

[loitonyngo]

Thanks a lot for the quick answer! I will start playing around with the code and see what I can bring out from it