You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
First, thank you for open sourcing this project. I am currently setting up a native Ubuntu 22.04 environment to deploy these tasks on a Unitree G1.
I have encountered a few blockers while trying to run the Sim2Real inference:
Missing Pre-trained Weights: The README states that "The trained policy weights for most tasks are already in the Sim2Real/logs/ folder." However, I noticed that the Sim2Real/.gitignore explicitly ignores the logs/directory. Consequently, the folder is empty after cloning. Are these weights hosted on an external drive (eg, HuggingFace/Google Drive), or is the intention for users to train from scratch?
Also it would be much appreciated if the weights are provided as my isaac is running on minimum vram possible, thus taking much time.
I am trying the pick and place task
Training Iterations: For the pick and place task, what is the recommended number of iterations to achieve convergence for stable sim-to-real transfer? Is the 4000-iteration mark mentioned in similar contexts the target for this specific task?
Activity
Implement reward/termination/obs/residual rework behind HS_REWORK flag (