PyTorchRL: Modular and scalable reinforcement learning in pytorch reaching new high in obstacle tower challenge

[giadefa]

Hi, don't know if it is the right place, but it is probably relevant for anyone interested in obstacle-tower-env.

https://arxiv.org/abs/2007.02622

We have scaled up our second position approach in the official challenge. We now reach a max of 23.6 floors on the test seeds and consistently above 20. Good news is that there is more work to do, as it does seem to plateau with the current method. The entire code and trained policies will be available soon.

[If there is a better place to put this let me and delete it]

[MarcoMeter]

Hi!

I could imagine a section on the readme.md to list such publications.

[awjuliani]

Hi @giadefa

This is great work to see, and I have no problem keeping this issue open so it can be discovered. The work is relevant to helping others who might want to get started doing research with Obstacle Tower.

[giadefa]

HI, We have just release the final version of PyTorchRL: https://github.com/PyTorchRL/pytorchrl Main points are:

• It is scalable on multi GPUs and multi nodes and
• It provides a modular approach which helps in developping multiple algorithms, both on-policy and off-policy, We provide several.

and the final paper: https://arxiv.org/abs/2007.02622

The docs: https://pytorchrl.readthedocs.io/en/latest/

I think that this is relevant for people who would like to try multiple algorithms or scale it up. We would be happy to work with people who are keen on this.

Best, Gianni