Machine learning approaches for soft robot control

This thesis presents the application of various machine learning techniques for control of soft robots. Simulation and experimental studies are described that show the feasibility of kinematic and dynamic controllers developed using learning techniques. The approaches are validated for both open loop and closed loop task space control. Subsequently, the role of morphology and its effect on control strategies are analyzed for two different cases; First, on a simulated octopus model and then experimentally on a soft manipulator for self stabilizing dynamic behavior. Finally, a short foray into embedded sensing is presented to eventually strive towards self sufficient embodied systems For the static case, global inverse kinematic solutions are directly learned, enabling us to develop computationally cheap controllers. The redundancy in the actuation system and hysteresis effects are the main factors to be considered while learning the static model. Using a learned network, equivalent in form to the traditional resolved motion rate controller, we develop accurate and easy-to-develop static controllers. Yet, this kind of controllers is energy inefficient and perform slow motions in order to maintain the statics assumption. Natural and fast motions can be derived using dynamic controllers. The problem is on obtaining the mapping from actuator forces to the time evolution of system states. A recurrent neural network was used to learn the forward dynamic model. Although the fundamental model is more intricate, the sampling and training time to obtain the model is still faster than the static case. With the new forward dynamic model, any numerical optimization method can be adopted to generate the control inputs. Consideration of the manipulator dynamics brings about fascinating motion behaviors. For instance, we were able to determine open loop trajectories that are globally stable and able to reach workspace regions that were not reachable statically. Later, we use a recent technique called model-based reinforcement learning for obtaining global closed loop control policies. These controllers were found ideal for controlling the soft manipulator dynamically when an unknown load is added, Finally, we perform behavioral studies on the reaching behavior of the biological Octopus using the same control approach and a simulated soft manipulator, which is morphologically similar to the animal. This provided us interesting insights into the role of morphology in shaping behavior. A short detour into modelling of soft resistive sensors is then presented. For this work, we adopt an approach similar to the human perceptive system for modelling embedded sensors. We demonstrate multi-modal sensing with randomly embedded strain sensors; all of the same kind.