Research On Modeling And Autonomous Motion Control For Modular Snake-like Robots

Being with multiple joints and hyper-redundant freedom,modular snake-like robots have high terrain adaptability because they can adaptively change their locomotion gaits with respect to various environments.Such robots have demonstrated their high potential in areas of disaster search and rescue,military reconnaissance and pipeline inspection.At present,the research on snake-like robots mainly follows the bionic principle to design their structure and gaits.Although the bionic research alleviates the difficulty on robot design,it also has many limitations.For example,the robot's structure and gaits have to be like a real snake,which implicitly exclude many other better structure and gait options.Thus,in this thesis,we aim to model snake-like robots and design their gaits from a perspective of learning by themselves,instead of using bionic prototypes.The thesis content is highlighted as below.1.Screw theory-based kinematics and dynamics modeling.We analyze the movement mechanism of a modular bionic parallel joint on a snake-like robot.In particular,based on the screw theory,the mechanism's constraints and degrees of freedom are determined.We find an analytical approach for solving the parallel joint's velocity and acceleration in the support of reciprocal screws and the Lie algebra.Then,we build a kinematic model describing the relationship between the snake-like robot's joints motion states and overall pose when it moves with the serpentine gait.We also construct a balance equation of force and moment,on top of which we derive the relationship between the robot's joints driving torque and motion states.Consequently,we obtain its dynamics model that can provide a theoretical support for the subsequent path-following control method;2.Optimal path-following control based on approximate dynamic programming.We derive a reduced dynamic model of a snake-like robot by using the differential geometry theory and virtual holonomic constraints.Subsequently,an optimal path-following objective function of the robot with uncertainties is constructed using this reduced dynamic model and a line-of-sight guidance strategy.We build a critic network to solve the corresponding Hamilton-Jacobi-Bellman equation of the above objective function.Through the approximate dynamic programming and the experience replay technology,the critic network's weights updating strategy is developed,and in the end,an online learning framework for optimal path-following control of the snake-like robot is constructed;3.Locomotion strategies' emergence based on deep reinforcement learning.We derive a fusion-based policy gradient method that combines both on-and off-policy gradients by using the Taylor expansion of on-policy gradients.It improves on-policy reinforcement learning methods'sample-efficiency,reduces the variance of policy-gradient estimation,and stabilizes locomotion learning processes.In order to improve the locomotion exploration efficiency and guarantee output actions' consistency,we present an efficient exploration mechanism for the robot's locomotion control strategy.Then,combining with an asynchronous parallel technology,we develop a distributed decoupling actor-critic framework to maximize data throughput of the proposed locomotion learning algorithm.As a result,a model-free control method for the snake-like robot's locomotion emergence is obtained;4.Learning locomotion control strategy rapidly via meta-learning.We present an anti-overfitting training method for a snake-like robot's dynamic ensemble by using a heteroscedastic data enhancement technology.This method breaks the correlation between historical data sets of the robot,and thus improves the robustness of trained models.We then construct a multitasking approximate meta-reinforcement learning framework based on the meta-learning theory,which reduces the high-order policy-gradient estimation'variance.Subsequently we use this framework to sample"imaginary" trajectories from the robot's dynamic ensemble.It relaxes the data requirement from a real environment and improves its sample-efficiency.Via the parallel computing technology,we present a distributed training method for locomotion search framework of the snake-like robot that greatly improves the efficiency of locomotion search processes.