Reproduce the effects of rl training backflip motion

[Mdwsbb01]

Hi Junjia,

Thanks for your meaningful open-source project, I have tested 'run' motion successfully using custom model at first. However, the 'backflip' motion is pretty hard to reproduce with the original 'stick figure' model or mine after training several million times iterations, and I have read the 'HumanoidAMP.yaml', notice you don't leave 'backflip' motion as the animation file in source code, so I wonder if there is any tricks or could you give me some hints? By the way, I really care the GPL-license 3.0 and the 'honor code', if you concern this part.

Thanks again and wish you have a wonderful New Year!

[Skylark0924]

Which custom model do you use? Do you mean robot model or algorithm model?

[Mdwsbb01]

The robot model actually, to be specific, I change the RSI (reference_state_init) and ET (early termination) parameters, which enabled our self-developed robot model to perform a backflip in Isaac Gym. However, the original stick figure model could not achieve the same result with the nearly same mass distribution and totally same DOFs distribution. I can only reproduce the backflip effect by inferring from your policy, so I wonder, for example, how many steps have you taken for training backflip motion or do we need to modified the Agent parameters in BaseTaskAMPRofuncRL.yaml? Thanks for your time and efforts in replying.

[Skylark0924]

The model training is 1 million steps by default, and since I don't see your robot model, it's really hard for me to help you solve it

[Skylark0924]

I can only reproduce the backflip effect by inferring from your policy

Since you said that you can reproduce the pretrained model I provided, then the problem is most likely on your robot model.