Open samibouziri opened 7 months ago
We went for Bullet first since historically the environment (handover-sim) and training pipelines (OMG-Planner and GA-DDPG) were both developed and tested in Bullet. That said, we don't see a major blocker for running everything in IG. There may be some minor hiccups:
This is assuming you want to stick with the current ray-based training pipeline. If you are thinking of leveraging GPU simulation, there will be one challenge---you can GPU-parallelize the env with IG, but the expert demonstration with OMG-Planner is not GPU parallelizable.
Hello,
Out of curiosity, I was wondering if there was a reason of training the model in Bullet instead of in IG? Is it due to the OMG planner?