Digital Theses Archive


Tesi etd-09212018-160240

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