Adaptive Control Strategies For Energy Efficient Walking On Biped Robots

Compared with wheeled and caterpillar track robots,biped robots have stronger adaptability to the complex environment and can work with or instead of humans to complete complicated and dangerous work,which are the future development trend of mobile robots.However,low efficiency and poor stability are two major obstacles that hinder the practical application of biped robots.Researchers found that energy efficiency and stability are usually mutually restrictive.Therefore,how to improve the stability on the basis of high energy efficiency has become an urgent problem to be solved.In order to solve this problem,this dissertation designs energy-efficient walking gaits based on the advantage of underactuated walking in energy efficiency,and proposes controllers with high adaptability to improve the stability of underactuated gaits.By theoretical analysis and experimental verification,this dissertation provides support for the practical application of efficient walking control.The contributions of the research are listed in the following aspects.Firstly,in order to improve the anti-disturbance performance of underactuated biped walking,this dissertation develops an active disturbance rejection controller design scheme for planar un-deractuated biped robots.Specifically,based on the idea of transverse coordinate transformation,the ADRC controller design scheme is in the form of a serial system with angular momentum as state variable to overcome the drawback of linear approximation,which could be applied to all pla-nar biped robots with one underactuated degree of freedom.Next,by taking a compass-like biped robot model as an example,this dissertation proposes the specific designing process of ADRC controller.The numerical simulations show that active disturbance rejection controller has faster convergence speed.Secondly,since the method above is still based on the robot model,which limits the appli-cation when the model is unknown or inaccurate,this dissertation supposes that the robot model is not known and proposes a stochastic policy gradient based adaptive control law.The walking controller could update parameters through training,and finally achieve the walking control from standing upright to the desired step length without considering the robot model and target limit cy-cle.The simulation results show that the basin of attraction is enlarged comparing with traditional event-based control.Finally,in order to apply the above two-dimensional walking control methods to an actual robot,this dissertation proposes a method to extend the energy-efficient planar walking gaits to three-dimensional robots and apply it to the humanoid robot NAO.Specifically,this method de-couples the three-dimensional walking problem to the problems of trajectory planning and control in the forward plane and the lateral plane.In the forward plane,the NAO robot is abstracted as a planar five-link model,with which an ADRC controller is designed.In the lateral plane,the reinforcement learning algorithm is used to generate lateral trajectories online to achieve the co-operation with the forward plane.After that,it is applied to the walking experiments of humanoid robot NAO,and the experimental results prove the effectiveness of our proposed control strategy.