Tuning Legged Locomotion Controllers via Safe Bayesian Optimization

This repository contains the code for our paper Tuning Legged Locomotion Controllers via Safe Bayesian Optimization. A demo video is available at https://www.youtube.com/watch?v=pceHWpFr3ng.

All relevant code for the paper is in gosafeopt/aquisitions/*, gosafeopt/models/__init__.py, gosafeopt/optim/swarm_opt.py and gosafeopt/optim/base_optimizer.py.

Setup

python -m venv .venv
. .venv/bin/activate
pip install -e . 
wandb login

It might be necessary to create a wandb account at wandb.ai if not already existing.

Examples

To train the pendulum toy problem model with a specific aquisition function and show the result plot run

 python examples/pendulum.py train --aquisition GoSafeOpt --seed 42

 python examples/pendulum.py plot --data-path wandb/wandb/{RUN_NAME}/files/res               

The aquisition function can be one of GoSafeOpt, SafeOptMultiStage, SafeUCB, SafeEI, UCB.

Settings

Most settings can be changed in config.txt