Motion Generation And Control Of Virtual Human Based On Musculoskeletal Model

Virtual character motion generation and control technology is an important research direction in computer graphics and robotics,and has important application value in many fields such as medical rehabilitation,character animation,and unmanned driving.At the same time,the continuous development of society puts forward new demands on the level of intelligence of virtual characters and the fidelity of generated motion.In this context,aiming at the unrealistic response of joint-driven virtual characters to human motion and the limitation of control ability under traditional methods,the paper conducts the following research on the motion generation and control technology for virtual human based on musculoskeletal models:1.A biomechanical-based virtual human musculoskeletal model is constructed,including19 body parts,47 joint degrees of freedom and 154 muscle-tendon units.The Hill muscle model and the Haeufle muscle model are used to calculate the driving force of different types of muscles.The dynamic system of the musculoskeletal model is established by using the Lagrangian method,elucidating the relationship between human posture and the degree of muscles activation.Meanwhile,the constraining effect of muscles on human joints is introduced.Thus,the law of human motion is revealed,and a theoretical model is provided for the motion generation and control of virtual human.2.A motion generation method based on deep reinforcement learning is proposed,and various muscle-driven virtual human motions are generated through the joint action of trajectory tracking policy and muscle control policy.The trajectory tracking policy is learned through the Proximal Policy Optimization,and the deep neural network structure,state space,action space and reward function of the trajectory tracking layer are designed,thus realizing the trajectory imitation of the reference motion.The muscle control policy is learned by using the supervised learning and experience replay mechanism,and the deep neural network structure and loss function of the muscle control layer are designed,thereby realizing the minimization of muscle energy consumption.3.A motion control method combining goals and curriculum learning is proposed to construct a user interaction-oriented virtual human control system.On the basis of completing the motion imitation,the target reward reflecting the completion of the task is added,the structure of the deep neural network is optimized,and the early termination condition is added to the training,thereby realizing the real-time response of the virtual human to the user’s instructions.Curriculum learning is introduced to assist policy training,and the setting of curriculum learning content in the training process is explained in detail,thereby improving the efficiency and success rate of policy training.4.By combining the constructed virtual human musculoskeletal system,the proposed motion generation and control policy,and the simulation environment established by using DART as a physical engine,a series of simulation experiments for virtual human motion generation and control are completed.The experimental results firstly prove that using the method proposed in the paper can generate virtual human motions that conform to the laws of human muscle control and have a high similarity to the reference motion.Secondly,it is proved that the virtual human can respond to the user’s control instructions during the movement process and successfully reach the target point specified by the user.Finally,the experiment shows that the virtual human has good anti-interference ability to the external environment and stability under the uneven terrain,and moreover,it can simulate various types of myopathy gait through simple operations,which has certain practical application value.