MO-BBO: Multi-Objective Bilevel Bayesian Optimization for Robot and Behavior Co-Design

Research supervised by Professor Kris Hauser

1. Overview

Robot design is complicated because of the time-consuming performance evaluation, multiple performance metrics to be optimized, and behavior optimization requirements in as many environments as possible. We propose multi-objective Bilevel Bayesian optimization (MO-BBO) technique to automate the process of form-behavior co-design.

Figure 1. Metric values of 200 gripper designs explored by MO-BBO, of which 13 designs are on the Pareto front


2. Results

We evaluate MO-BBO by applying to (i) gripper design and (ii) bimanual arm placement. (i) In the gripper design problem, we use mass metric and the time elapsed until the gripper drops an object during lifting and shaking as performance metrics. And our object set consists of 13 objects with varying shapes and difficulty of grasping. (ii) In the bimanual arm placement problem, we use reachability metrics and measure that in 5 scenarios {Ground, Low-shelf, High-shelf, Table-Horizontal, Table-Vertical}.


Figure 2. Four gripper designs from the Pareto Front of Fig. 1. Images captured during lifting and shaking. Some grasping poses have force closure but objects can still slip off the gripper. A full fledged dynamic simulation can take these issues into account.

Figure 3. Video of grasping objects by design that has max elapsed time metric.

Figure 4. Four robot arm designs generated by MO-BBO. Best performers highlighted in bold.