The Isaac Gym team has been hard at work improving our Shadow Hand training environment to be able to more closely reproduce those that OpenAI reported, including smoothing the hand motion, Asymmetric Value and Policy networks, supporting LSTM models, and fixing various bugs. With these changes, you can train the Shadow Hand model on a single GPU 3x faster than the supercomputer used in the original experiment.
Thanks the team for the great effort. It seems like you have some soft bodies with franka result from the video. Will this will also be included in the upcoming realease?
Some soft body support is already there in the current release, but we don’t have any documentation around it in part because we don’t expect to support those APIs the same way in the future once Gym capabilities are integrated with Omniverse. The new release won’t have any further soft body features exposed.
Our research team has some papers coming soon that use the existing functionality - if code is shared, it might serve as a more tangible example of the soft body API usage.