PyTorch problem with jetson-reinforcement on the Xavier

Hi,

I was trying to run jetson reinforcement demos on the Xavier and ran into problems.

I had to compile gazebo and pytorch from source. I installed pytorch v1.0 to be able to run on CUDA 10. I got the “Cartpole” demo running with a simple chege on the code, but the simulation was not improving over time. Then, I realized that there were several breaking changes introduced in pytorch version 0.4. I tried to port the code by reading the PyTorch 0.4.0 Migration Guide, but failed.

Is there a plan to port these demos to pytorch 1.0 or 0.4?

Has anyone been able to run the demos on Xavier somehow?

Thanks.

Hi Miguel, you can try this similar CartPole DQN example which appears to have been updated for PyTorch 1.0.0: https://pytorch.org/tutorials/intermediate/reinforcement_q_learning.html

Here is another post about installing PyTorch on Xavier with CUDA 10.0:
https://devtalk.nvidia.com/default/topic/1041716/jetson-agx-xavier/pytorch-install-problem-solved-/post/5284747/#5284747

The entire repo would take more time to port over to PyTorch 1.0, the timeline for doing that is still TBD.