Learning Latent Representations to Co-Adapt to Humans

This is a minimal repository for our paper "Learning Latent Representations to Co-Adapt to Humans" [http://arxiv.org/abs/2212.09586]. We include two main things:

• Code for our proposed alogrithm RILI: Robustly Influencing Latent Intent
• Four custom gym environments we used during our simulated experiments

Requirements

• python (v3.9)
• pytorch (v1.14)
• gym (v0.23)
• pybullet (v3.2)
• setuptools (v63.4)

You can install the above packages by running the following command

pip install -r requirements.txt

Then, install the gym environment:

cd gym_rili
pip install -e .
cd ..

Instructions

To train the RILI model with the Circle environment, run the following:

python3 main.py --env-name rili-circle-v0 --save_name []

You can train the model in different environments using the --env-name argument. It has the following values:

• rili-circle-v0
• rili-circle-N-v0
• rili-driving-v0
• rili-robot-v0

To test the trained models, run the evaluate.py script. Pass the correct environment name, and use the --resume argument to pass the name of the saved model that you want to load:

python3 evaluate.py --resume [model_to_load] --env-name [env_name]

Other parameters that you can change are:

• --change_partner: the stochasticity with which the partner changes (a value between 0 and 1)
• --num_eps: the number of episodes for training
• --start_eps: the number of episodes for exploration

We have pre-trained checkpoints saved in the repository for each environment. You can load the desired model by passing the name in the --resume argument:

• For Circle environment:

  • --resume circle_env_30000

• For Circle-N environment:

  • --resume circle_N_env_30000

• For Driving environment:

  • --resume driving_env_30000

• For Robot environment:

  • --resume robot_env_30000