Motion Generation Of Multi-legged Robot By Using Estimation Of Distribution Algorithm And Domain Adaptation

The problem of motion generation for multi-legged robots is an important area in robotics research.Due to its complex structure and the strong uncertainty of the external environment,the gait optimization of multi-legged robots is faced with the problems of high computational complexity and difficulty in obtaining effective movements in an unknown environment.In this paper,the relevant background and research status of multi-legged robot motion generation are introduced in detail.The six-legged robot model designed in this paper is introduced.The objective function to evaluate the motion effect and energy loss of multi-legged robot is defined,and multi-objective optimization is given.The concept of the problem,and introduced the multi-footed robot's motion generation problem as a multi-objective optimization problem,and used the evolutionary computation method to solve the algorithm framework.In this paper,we propose RM-MEDA-MG,which is based on the rule model distri-bution estimation algorithm for multi-footed robot motion generation method.The macro-evolutionary approach based on search space and probability distribution adopted by the Estimation of Distribution Algorithm(EDA)can solve the multi-objective optimization problem with faster convergence speed and stronger global search capability.In the pro-cess of evolution,the Local PCA algorithm is used to cluster the populations,and then these subpopulations are modeled using a regularity model.The non-linear POS manifold-s are segmented.Finally,each regularity model is randomly sampled to obtain the next generation population.The regularity model is very effective for highly continuous com-plex multi-objective optimization problems with highly correlated decision variables,and it is well suited for high-dimensional multi-foot robot motion generation problems with high-dimensional space-related motion sequences.This dissertation uses the distribution estimation algorithm based on the regularity model as the control iteration model of the multi-legged robot.The experimental results also prove that the RM-MEDA-MG solves the multi-legged robot's gait optimization problem well.According to the idea of transfer learning,DA-MG is proposed,which is a multi-legged robot motion generation method based on domain adaptation.The motion generation problem of multi-legged robots in different environments is regarded as a dynamic multi-objective optimization problem.The locomotion in a known environment is regarded as a source domain,and the gait in another unknown complex environment is regarded as a target domain.The TCA algorithm is applied to map the decision variables of the source domain and the target domain into the same hidden space,classifying the gaits in a known envi-ronment by non-dominated sorting,using the prior knowledge represented by the resulting labels,which helps the target task to select more potential individuals as part of the initial population in the early iterations of the evolutionary algorithm,and optimize on this basis.In this process,the walking knowledge of multi-legged robots transfer from the source task in a known environment to the target task in another unknown complex environment,which reduces the computational complexity of multi-legged robots' motion generation in an un-known complex environment.The robot's gait is more stable,effective and saves energy.