This presentation is about the development of a general optimization framework for soft robotic actuators using auto-differentiation. This development is motivated by inherent difficulties in modelling and development of both soft robots and their controllers due to the non-linear and time-variant properties of the forces acting on them. Furthermore, solutions from comparable literature tend to lack generality and restrict themselves to certain use-cases or offer methods to work around the issues but are more time-consuming in return.
The goals of the project were both the development of the framework as well as testing its capabilities in a number of exemplary optimizations. A literature study is first conducted related to soft body simulation and optimization procedures to find valid approaches in modelling soft robotic actuators and performing auto-differentiation. After an approach is chosen and implemented, a verification of the simulation is performed by comparing multiple simulations to an established FEM solver. Before optimizations can be performed the issue of exploding gradients from auto-differentiation needs to be addressed by proposing a scaling scheme. To demonstrate the use of the developed 3D optimization framework for controller synthesis, a small neural network feed-forward controller is set up for a pneumatic endoscope actuator model with three pressure chambers and trained using the derived gradients from the simulation. Furthermore, a meta-optimization scheme is presented, where the damping factor of the simulation is split into 50ms time-windows and optimized with the intention of shortening the time until the deformation of the actuator reaches its final state.
The developed system is shown to be able to derive meaningful gradients that can be used to optimize different components of the simulation. The proposed scaling scheme to avoid exploding gradients requires the user to fine-tune a few parameters to get optimal results. The scheme has been shown to produce useful gradients for an exemplary pressure optimization and controller synthesis. In comparison to reinforcement learning, the controller synthesis requires about 1 order of magnitude fewer iterations steps to converge in addition to more smooth and reduced loss fluctuation over the course of training.
The meta-optimization managed to reduce the required number of time-steps by 25%. The loss function showed strong signs of being ill-defined but the optimization still succeeded based on the gradients and changes in the damping factors, which implies that more complex and well-behaved formulations have the potential to give even better results.
To join the presentation via Microsoft Teams click here