Christian, Prathyush, and I came up with a rough plan on how to proceed with our new research and some goals/questions to address for this quarter. It can be found below:
- Parameterized robot specification
- What parameters are required to fully specify a robot?
- Create a composable that can export robot in a way that can be used in our chosen simulation environment (see below)
- Simulation
- Research current offering of physics engines (needs to support soft-bodies)
- Chrono
- Bullet
- Voxelyze (do we want to use voxels?)
- Research machine learning libraries
- Tensorflow
- Torch
- Research current offering of physics engines (needs to support soft-bodies)
- Machine Learning Research
- Background research on reinforcement learning methods and principles
- Controller (multiple options)
- Come up with a PID controller for the Jack’s car
- Might be good to start with as it is a simple controller
- Choose form of feedback
- A common form of finding gains for PID is the Ziegler-Nichols Method
- Use machine learning to calculate gains for the controller that give similar or better results as the Ziegler Nichols method
- Neural Network based controller
- More complicated but in theory could be trained to be more robust than simple PID
- Come up with a PID controller for the Jack’s car
This is a work in progress and will be updated in the future.