Can Transformers Perform Low-Level Control In-Context?

Ebonye Smith, Aidan Andrews, Phoenix Wilson, Alex Frias, Ayush Pandey, Gireeja Ranade
University of California Berkeley
Univiersity of Illinois Urbana-Champaign
University of Southern California
University of California Merced

This repository contains the implementation of low-level transformer controller framework that does stabilizing control for cartpole system and an acrobot system via incontext learning.

Setup

This repository requires two conda environments. Create a conda environment using environment_test.yml. This conda environment will be used for generating data, training, and doing inference. Create a second conda environment using environment_gym.yml. This conda environment is used only for visualization purposes when running gym_acrobot_inference.py or gym_cartpole_inference.py.

1. Creating a Dataset

To generate a dataset, use the following command:

python generate_dataset_cartpole_ebonye.py --config conf/_XandYtest_cartpole_ebonye.yaml

---

2. Training the Model

To train the model, run:

python trainSequential_ebonye_cartpole_zerodyn.py --config conf/_XandYtest_cartpole_ebonye.yaml