JunpengGao233 / ImplicitSoftBody

ImplicitSoftBody

Package

.
├── implicit_soft_body
├── energy : all the energy class
├── network.py : neural network class
├── system.py : base robot class
├── robot_model.py : include all the robot models
├── Sim.py : differentiable simulator

Script Folder

The folder scripts contains different scripts using the package.

• normalize_mesh.py is to normalize the given mesh to make it located within the boundary of visualization.
• pretrain.py uses the given log of input and output of a control policy to imitate this control policy
• simulate.py loads a trained model to simulate and output the sequence of actuations for visualization
• train.py trains a model to predict the actuation given the position and velocity of nodes of soft robot.
• visualize_actuation_seq.py: visualize the motion of soft robot given the sequence of actuation.
• robot_model_euler.py: use implicit euler integrator to optimize the walk but stuck in local minimum, results in assets/example/vis_euler_12.html
• supervised_learning.py: use supervised learning to learn the actuation signals directly, resutls can be found in assets/example/final99.html
• robot_model_bdf.py: use bdf2 integrator to optimize the walk, results in assets/example/bdf31.html
• SITL.py: use Solver-in-the-loop training strategy.

To run the code

1. Install Depencencies (jinja2,numpy, Pytorch)and locally install the package (Tested on Python 3.9) pip install -e . . In addition, Pytorch should be installed.

2. Run the script case_0.py or Run the script in the following order: train.py -> simulate.py -> visualize_actuation_seq.py

3. In the output folder, you can find the visualized index.html file

We provide different training cases. Not all cases are stable. The training on the differentiable simulator is not easy. We did add some noise to the actuation during the training. Training of control policy of simple robots is easier than that of complex one.