Co-Optimization of Design and Manufacturing Parameters for Low-Cost Robotic Actuation

Additive and low-cost manufacturing techniques promise increased access to robotic actuation at the cost of mechanical precision. In this work, we employ principled Design of Experiments (DoE), including Taguchi orthogonal arrays, in parallel to sequential experimentation enabled by Bayesian Optimization (BO) for co-optimization of design and manufacturing parameters across two design case studies. We optimize for a combination of gear ratio and backdrivability in a 3D-printed compound Wolfrom bilateral gearbox. We also optimize for crack pressure and steady-state pressure differential of an injection-molded silicone check valve. Using BO, we find a 3D-printing compatible gear design with a gear ratio of 63.6 that backdrives without ever needing more than 0.35 Nm of input torque. This represents a 49% increase in ‘score’ over the Taguchi method. Similarly, we find a BO valve with lower combined crack and steady-state pressure errors than the Taguchi trials, decreasing cumulative error by 55%. Tracking model uncertainty throughout training, we conclude that further model training is necessary to reach optimal results in both cases. We further conclude that BO via the Ax platform is not yet as “plug-and-play" as Taguchi arrays.