search

fyp21011 / PlasticineLab

Codes for HKU CS Final Year Program 21011: Applying the differentiable physics on deformable objects to enhance the performance of robot learning tasks
1 stars 0 forks source link

What's the usage for torch_nn and torch_lstm? And is that possible to add force information to the state? #12

Open YunchuZhang opened 2 years ago

YunchuZhang commented 2 years ago

As title. I found in torch_nn it is trying to train a small network that maps states to actions. What dose it used for?

And is that possible to add force information in the environment state?

Thanks

Ericcsr commented 2 years ago

The network is similar to policy in RL, instead of optimized by approximated policy gradient or gradient from Critic function as RL, we directly optimize it with gradient from simulator. Please refers to: https://github.com/taichi-dev/difftaichi for more detail The second question: No, our simulator is kinematic based without explicit dynamics, the action is the endeffector's velocity actually. If you refers to internal force that computed inside MPM, it is possible but may need to consult more document regarding MPM.