Learn by Doing - NeurIPS 2021 Competition (ROBO Track)

This repo contains the code for our submission of the learn by doing ROBO track.

Folder structure

• training contains the scripts that we used to train the models.
• controllers contains the controller.py script that, depending on the system, loads a trained model and predicts the next control action.

Models used

Model 1 (Polynomial features with linear regression)

These models consist of an imitation learning set up, in which we try to learn the mapping f(X, Y) -> U from the given training data. We used polynomial feature augmentation (degree 2) on the current state variables.

Tricks used for the models involved clipping the predicted U vector and normalizing inputs.

Model 2 (Two step linear regression) (Bumblebee only)

We trained a linear system model that predicted the next state given the current values of the state and the input. We then implemented a one-step dead-beat controller with clipping.

How to train the models

The models should be trained from the root directory using one of the scripts on the training directory. The bumblebee systems are trained using the linear.py script and other systems using the polynomial_features.py.

The training data should be on the training_trajectories and create a models directory. The models will be used for the training scripts to save the models.