Design And Control Of Robotics Based On Dielectric Elastomer

With the development of science and technology,soft active material(SAM)are increasingly being applied to engineering and everyday life.Soft robots designed from SAM have gradually draw scientists' attention.Their bio-affinity,environmental adaptability,and especially the flexibility exhibiting in complex environments are not available in traditional robots.However,its complex nonlinear dynamics bring great difficulties to control.As a typical kind of SAM,dielectric elastomer(DE)has characteristics such as fast response,large deformation and high energy density.This paper focuses on the design and control of soft robotics based on DE.First,we carried out the control of a single soft robot.Based on DE,we designed a kind of robotic cuttlefish,which relies on water jet propulsion.The structural design was analyzed and optimized.Through model-free reinforcement learning,the voltage sequence of the robotic cuttlefish gradually changed from chaos to a periodic pattern.And the swimming speed of the robotic cuttlefish increased from about 0.20 body length per second to about 0.38 body length per second,with about 91%improvement.This work provided a method to control DE driven robotics with model-free reinforcement learning,which can also be applied to other SAM based robotics.Then,we carried out the formation control of multiple soft robotics.Inspired by manta ray,a kind of robotic fish driven by median paired fin was designed.Then we designed control algorithms to mimic the torus behavior and migration behavior of swarms.And we verified the feasibility of the algorithm with simulation.Based on the global visual system,we succeeded in the circular and triangular formation control of three robotic fishes through experiments,which lay the basis for a swarm of soft robotics to explore deep sea in the future.Finally,we studied the application of deep learning(DL)based computer vision in controlling soft robotics.Using the Mask R-CNN deep learning framework,we carried out instance segmentation of the robotic fish in order to track its position,posture,and deformation.The trained neural network could identify the robotic fish with confidence above 0.99 and give masks for covering the robotic fish.Then,a pneumatic soft gripper and a traditional robot arm were combined to form a gripping system.By using instance segmentation to track desired objects,we could adaptively grip desired objects with the gripping system.This visual method,which provides more effective recognition and tracking,will contribute to the control of soft robotics in the future.