New plugin for Design Optimization

[hugtalbot]

We are excited to announce the launch of the toolbox for Design Optimization with SOFA. Ideal for optimizing your robot design, it could also be used to calibrate mechanical parameters of your model!

An accompanying paper was recently accepted for publication in RA-L (see the available preprint here). Multi-objective design optimization is an essential paradigm in soft-robotics that is changing the way in which designs are conceived, shared and adapted.

Don’t hesitate to reach out to @TanguyNav and @sescaida if you need support when trying out the Toolbox 🆘

📢 Discover the plugin 📢

In terms of architecture, the Toolbox features three main parts:

• a Configuration module, where the user specifies a parametric design created in Gmsh through the Python 3 interface that grants access to the constructive geometry engine used by Gmsh. The user specifies the parameters that will be explored during the optimization loop, including the corresponding intervals.
• a Design Optimization Loop, where a user configured scene is run in SOFA for each parameter set (configuration), evaluating the fitness functions associated with the optimization problem. Surface and volumetric meshes are generated for each configuration on the fly, so the process is fully automated. Currently, Optuna is used as the solver for multi-objective optimization. However, other solvers can be easily integrated, which is part of the ongoing work.
• a Results Visualization module, where the user has the ability to visualize the Pareto Front resulting from the problem, with the ability to inspect design parameters, etc.

In the paper, authors showcase the functionality of the Toolbox on the example of a cable-driven sensorized soft finger. The use of the toolbox allows us to automatically find designs that are optimal tradeoffs between sensitivity in the deformation measurements and ability of the finger to bend. They found that points on the extremes of the Pareto Front obtained from simulation are indeed far apart in the established metrics when the fingers are fabricated and tested, giving encouraging results for the sim-to-real behavior of this approach. Furthermore, for the sim-to-real approach, S. Escaida N. et al. discussed the effect (or lack thereof) of mesh density in the simulation, mechanical parameters and fabrication tolerances. In addition, authors show that the fabrication of the fingers can also be automated, as the molds needed for casting are derived from the design parameters as well.

📥 Check out the code 📥

The work on the toolbox is ongoing and we expect to provide new examples, showcasing different features, use-cases and results, in the upcoming months