Research On Obstacle-avoiding Strategy Of Hexapod Robot Based On Deep Learning

The hexapod robot has good motion stability and non-structural environments adaptability.It is the best choice for tasks such as extreme environmental detection,disaster rescue,and national defense security protection.Researchers in the field of robotics from different countries have carried out research on hexapod robots continuously,forming a more complete method of kinematics,dynamics,and foot trajectory planning,which makes the hexapod robot gradually out of manual operation and promotes autonomous movement.There are many problems in achieving full autonomy of the hexapod robot,including environment understanding and how to implement path planning in environment.The foothold of the hexapod robot is widely distributed in the space,so that it can obtain excellent obstacle avoidance ability.Due to the lack of cognition of obstacles by hexapod robots,the obstacleavoiding system used in wheeled or pedrail robots is applied with hexapod robots,which doesn't make foothold create the biggest advantage.Therefore,this topic focus on describing the obstacles and formulating class processing patterns of the hexapod robot dealing with obstacles and achieve the goal of self-implemented path planning.For different tasks,the obstacle avoidance strategy adopted by the hexapod robot should be treated differently even if the same obstacles.Therefore,the hexapod robot is required to identify the obstacle type to determine whether the obstacle belongs to the "obstacle-crossing " obstacle or the "obstacle-avoiding" obstacle,and selects correct strategy.Although the traditional target recognition algorithm can meet certain requirements in speed,it is unable to solve the problem effectively because of the lack of prior knowledge in the feature extraction.It fails to develop accurate recognition.This technology needs a crowd of sample data as a support to perform well in accurate target recognition algorithm based on deep learning performs.The more data size,the more calculating pressure.Under the current technology,the method of pre-training combining with the parallel computing of multiple graphics cards has alleviated this problem to some extent.This topic has used the Mask R-CNN target recognition framework to establish an obstacle recognition model.This algorithm ensures the first classification based on obstacle class,and also completes the second classification based on obstacle size.The generalization ability of the model and the feasibility of the method are proved by analyzing the training error and the testing error through the experimental data.On the basis of hierarchical processing,all the obstacles that adopt the "obstacle-avoiding" strategy are obtained,and environment model is established.The global path planning of the hexapod robot is realized in combination with the reinforcement learning method.