| In recent years,along with the wave of development of artificial intelligence,robots have been increasingly used in human production and life.In indoor construction scenes,humanoid robots are more flexible and effective than wheeled robots,and gradually become a hot spot in the field of robotics research.Indoor building environment is a complex dynamic environment,and humanoid robots need to walk stably and steer constantly to avoid obstacles.However,this is a challenging problem for humanoid robots with high degrees of freedom.Although humanoid robots have made tremendous improvements in motor skills over the past few years,many methods of generating motion still require tedious and timeconsuming manual calibration due to hardware discrepancies and inaccurate sensors.In order to improve the locomotion of humanoid robots and achieve stable humanoid robots in indoor architectural environments.In this thesis,two aspects of gait planning and path planning of humanoid robots are investigated as follows.In order to improve the walking ability of humanoid robots in indoor buildings,a gait optimization method based on the Parallel Comprehensive Learning Particle Swarm Optimizer(PCLPSO)is proposed.Firstly,the key parameters affecting the walking gait of the humanoid robot are selected based on the natural zero moment point trajectory planning method.Secondly,we decompose the gait training task by changing the slave group structure of PCLPSO algorithm and build a parallel distributed humanoid robot gait training framework based on Robo Cup3 D.Finally,a hierarchical learning approach is used to optimize the turning ability of the humanoid robot.The experimental results show that the PCLPSO algorithm achieves the optimal solution faster,and the optimized humanoid robot has fast and stable gait and excellent turning ability.A parallel Deep Deterministic Policy Gradient(DDPG)algorithm for model-free gait optimization of humanoid robots is proposed to address the time consuming modeling of humanoid robots.The conventional approach cannot fully utilize the autonomous exploration capability of the humanoid robot.To expand the exploration range and improve the training efficiency,a multi Actor-Critic(AC)network is established.An experience filtering unit was introduced to optimize the experience playback mechanism,and the cosine similarity method was used to classify the experience.Then,a Markov Decision Process(MDP)model based on knowledge and experience is designed to solve the payoff sparsity problem.Finally,experimental results show that the parallel DDPG algorithm can make the humanoid robot walk faster and more stably with a speed of 0.62m/s.In order to improve the path planning capability of the humanoid robot in a real-time dynamic environment like indoor buildings,a dynamic path planning method based on the DDPG algorithm is proposed.First,the threat distribution model of the opposing player is designed using the two-dimensional normal distribution method.A global threat distribution model is constructed by combining the relationship of distance,relative orientation and relative speed between our player and the opponent player.Secondly,the MDP model for dynamic path planning of the robot is built according to the working environment and motion characteristics of the humanoid robot.Finally,the experience generated from training is stored and extracted in a classified manner according to the importance of the experience.The experimental results show that the humanoid robot reaches the target point in 20.1%shorter time,also avoids dynamic threats,and reduces the gait transition time when walking by 51.9%. |
