About Unitree Robot Models and Training PHC

[budiu-39]

Thank you for your excellent work on Perpetual Humanoid Control (PHC) and for continuously updating it. It’s impressive to see that PHC now supports Unitree H1 and G1 robots, along with code for retargeting and training.

I am interested in working on motion tracking for robots based on your work. Given the complexity of training the model from scratch (as you mentioned, "Training PHC is not super automated yet, so it requires some (a lot of) manual steps"), I would like to ask:

1. Do you plan to provide pre-trained network models for Unitree robots?
2. If I have to train the model from scratch, do you have any recommendations or best practices?
3. Based on your experience, how does training a single primitive compare in performance to training the full PNN model for a robot model (e.g., Unitree robots)?

[WesleyXu224]

我也想问这个问题，对训练模型表示苦恼中（用于language2control的model）

[ZhengyiLuo]

Thank you for your excellent work on Perpetual Humanoid Control (PHC) and for continuously updating it. It’s impressive to see that PHC now supports Unitree H1 and G1 robots, along with code for retargeting and training.

I am interested in working on motion tracking for robots based on your work. Given the complexity of training the model from scratch (as you mentioned, "Training PHC is not super automated yet, so it requires some (a lot of) manual steps"), I would like to ask:

1. Do you plan to provide pre-trained network models for Unitree robots?
2. If I have to train the model from scratch, do you have any recommendations or best practices?
3. Based on your experience, how does training a single primitive compare in performance to training the full PNN model for a robot model (e.g., Unitree robots)?

Hi! Thanks for your interest in our work. I plan on releasing unitree models for PHC at some point; however, it will probably be a policy that would be able to transfer to real rather than just in sim. This will take some time.

For training from scratch, a single primitive is sufficient for most of the use cases, especially now I have incorporated better training scheduling than the original ICCV paper.