Hello, I would like to create a custom environment for training of quadrupedal robots with RL. My requirements are:
parametrized terrains, to apply radomization
creation of custom complex course, like with obstacles elements to test agility
(possibly) dynamic moving parts
I’m wondering which is the more appropriate way to proceed (by code? by custom .usd?). I previously used isaacgym envs but i’ve also seen orbit which claim better performance. Which one fit better the task? which is going to stay? and is orbit already stable?
I’ll answer my own question for the ones interested. I frame the situation as it is today:
Isaacgym preview release for training of policies has a good api for quite simple terrains creation through the use of meshes applied (by code) to the terrain and creation of primitive shapes. You cannot work with .usd but you can load assets as xml or urdf. The GUI is only a viewer.
Isaacsim in particular “Omniverse isaacgym environments” with the RL extension. You can create more complex terrains by loading custom .usd worlds as well of same functionalities provided in Isaacgym preview release but without the interaction with its higher level API. The GUI provided is the Omniverse USD Composer allowing also user interactions with physic on, visual programming with Omnigraph etc. The RL extension is very poor but with some effort for the migration you can deploy agents trained with isaacgym preview release in there.
Isaac Orbit, built on top of Omniverse extensions and IsaacSim (therefore it has the USD composer GUI), provides photorealism, better physic control (?) and should also be more efficient than isaacgym preview release (from graphs in the landing page). Great support for RL training libraries but is still in development at the time with missing sensors and functionalities. So if you plan complex situations as in my case, it’s still early
p.s. thanks for the clear naming convention and overlapping products…