Bayesian Optimization for MPPI Planar Pushing

In this project, we implement the Bayesian Optimization Algorithm (BOA) for autotuning the hyperparameters of Model Predictive Path Integral (MPPI) control to solve a planar box pushing task with non-trivial obstacles.

Quick Start

1. Install all necessary dependencies using bash install.sh.
2. Run our demo using python demo.py. Instructions and results will be printed in the terminal and a live pushing demo should show up in pybullet GUI.

API for Easy BOA Application

This implementation provides an easy-to-use Bayesian optimizer with full GPU pipeline support. The optimization results are at least comparable to the most popular BOA implementation in python.

Check here for more details.

Dependencies

• Python >= 3.8
• PyTorch >= 1.11
• GPytorch >= 1.9.1

Reference

BOA

f: objective function
X: support of variables
S: acquisition function
M: GP model
D: dataset
x: variable
y: observation

MPPI