NVlabs/ProtoMotions ? reverse-engineered prompt

Build me a GPU accelerated framework for training simulated humanoid characters and humanoid robots.

I want it to support motion learning from human motion data, retargeting motions onto different robot bodies, and training control policies that can run in a few different physics simulators. It should be able to switch between at least a couple simulator backends, make it easy to try different robot models, and include a simple way to test the same policy across simulators.

Please also include a clear training workflow, dataset preparation steps, and a deployment path that exports a single model I can use for real robot inference with raw sensor inputs. If there are example motions, demo scripts, or quick start commands that make the project easier to try end to end, add those too.

Keep the code modular and practical, with good docs so I can understand how to run it locally. Look up current docs online if you need to, especially for the simulator integrations.