Integration with TorchRL
fedebotu opened this issue ยท 3 comments
fedebotu commented
๐ Feature
I think it would be great to integrate the current RL bolts models with the new TorchRL library
Motivation
While gym
is a great library for environments, it may not be very flexible for certain applications and has limitations (such as the need to run environments on CPU). This is where the new PyTorch library torchrl
can help with features such as:
- Batched environments
- TensorDicts are a great way to handle data
- Can directly run the environments on single and multiple GPUs
- Can wrap multiple environments, including
gym
- Convenient implementations of models and methods such as rollouts
Additional context
I uploaded in this repo an example notebook of the official TorchRL's pendulum tutorial with a custom environment that can run on batches and multiple GPUs