Co-Optimizing Design and Manufacturing for Low-Cost Robotic Actuation

_{^{Written by Yi Cao, Master’s student in Mechanical Engineering and Applied Mechanics, Researcher in ModLab, advised by Mark Yim}}

_{^{Caption for featured image: Top left: CAD models of the gearbox and valve. Middle left: 3D-printed prototypes. Bottom left: performance data. Right: BO22 Wolfrom bilateral gearbox (Bayesian Optimization trial #22).}}

Robotic systems often rely on high-precision components to achieve reliable actuation. However, rapid prototyping and additive manufacturing now make it possible to fabricate robotic mechanisms quickly and inexpensively. The tradeoff is that low-cost manufacturing methods, such as desktop 3D printing or soft material molding, introduce variability that makes system performance harder to predict and optimize.

In our recent work, “Co-Optimization of Design and Manufacturing Parameters for Low-Cost Robotic Actuation,” published in the proceedings of IEEE/SICE SII 2026, we explore how systematic experimentation and machine learning can improve the performance of low-cost robotic components. This work was conducted in ModLab at the GRASP Laboratory and co-authored by Gregory Campbell, Yi Cao, Hannah Escritor, Zihao Zhou, and Mark Yim.

The central idea is to co-optimize both design parameters and manufacturing choices simultaneously. Instead of relying solely on theoretical models, we combine analytical modeling, structured experimentation, and machine-learning optimization to explore the design space of physical systems more efficiently.

To evaluate this approach, we studied two case studies.

3D-Printed Robotic Gearbox

The first case study focuses on a compound Wolfrom bilateral gearbox, designed to achieve both high reduction ratios and meaningful backdrivability—properties important for applications such as haptics, exoskeletons, and compliant actuation.

The gearbox components were fabricated using desktop 3D printing. To optimize performance, we varied both design parameters (gear geometry and clearances) and manufacturing parameters (material, lubrication, and gear thickness).

To explore the design space, we compared two approaches:

• Taguchi Orthogonal Arrays, where parameter sets were selected using a heuristic search over the analytical gear model
• Bayesian Optimization, used as an active learning approach to sequentially guide new experiments

Using this workflow, we identified a gearbox configuration achieving a 63.6:1 reduction ratio while remaining backdrivable with less than 0.35 Nm of torque, improving the optimization score by 49% compared to the Taguchi baseline.

Soft Pneumatic Valves

The second case study uses the same two approaches to examine a soft silicone check valve used in pneumatic systems for soft robotics. By varying material, geometry, and fabrication parameters, we optimized the valve to match target pressure characteristics for a pneumatic test system. Bayesian Optimization guided search reduced pressure error by 55% compared to baseline designs.

Toward Data-Driven Mechanical Design

Together, these results demonstrate the potential of combining rapid prototyping, structured experimentation, and machine learning to improve hardware performance while keeping fabrication accessible and low-cost. They also emphasize existing questions in the field around when to transition from minimizing model uncertainty with a model-improving acquisition function to maximizing outputs with a greedier optimization function. More broadly, this work points toward data-driven mechanical design workflows, where robotic hardware can be iteratively improved through experiments guided by learning algorithms.