Creating a photorealistic reiforcement learning environment : Best way/ practices to do it

[umangdighe]

Hi,
I want to train a wheeled robot with differential drive to drive in a photo realistic outdoor environment. The outdoor environment will have multiple objects and obstacles that the robot is not allowed to touch/hit. Considering I want to train the robot to do multiple tasks : For example : Drive only on marked path but cover the entire area in optimal manner, I am still trying to figure out if I want to use manager based RL environment (https://isaac-sim.github.io/IsaacLab/source/tutorials/03_envs/create_manager_rl_env.html) or direct workflow RL environment (https://isaac-sim.github.io/IsaacLab/source/tutorials/03_envs/create_direct_rl_env.html#) .

Going through the tutorials, I see examples of importing usd terrain (https://isaac-sim.github.io/IsaacLab/source/api/lab/omni.isaac.lab.terrains.html#omni.isaac.lab.terrains.TerrainImporter.import_usd).

My question is: Is the best practice/option is to create the entire scene including the objects/obstacles and import the photorealistic scene as a usd file OR only import the ground withe markings/etc as USD and import the obstacles separately? Is there an example of creating and importing a photo realistic training environment?

The question arises because I need to define the rewards based on no go zones/ not hitting objects and completing the exploration optimally.

Thanks! I appreciate any help or pointers.

[yash-s20]

did you find a good solution @umangdighe ?