CPG를 이용한 ZMP 기반 이족 로봇의 무릎 핀 보행 궤적 생성

Abstract

본 논문에서 제안하는 바는 이족 보행 로봇(HYBRO)을 이용한 베이스 바디의 무릎 핀 보행을 위한 궤적과 보행 시 관절의 반복성을 위한 CPG들의 신경회로망의 조합이다. 또한 로봇의 안정적인 보행을 위하여 지지하는 발의 중심으로부터 roll방향 ZMP을 측정하여 로봇의 길이와 무게를 가지고 안정적인 보행을 하였다. 실제 ZMP의 좌표를 한발 지지발의 pivot인 점을 기준으로 베이스 바디의 기울기 값을 계산하여 SSP일 때 스윙 하는 발에 의한 불안정성과 DSP일 때 무게중심의 이동 시 발생하는 불안정성을 ZMP 제어를 통하여 줄였다. 최근 연구에서는 이족 보행 로봇에 CPG만을 적용한 연구가 진행 되어 왔었다. CPG의 큰 특징은 일정 범위 내에서 불특정 외란을 극복하고 자연스러운 보행 자세를 얻을 수 있다는 점이다. 이것은 CPG 구성상 CPG unit 하나하나가 limit cycle를 그리기 때문에 상태안정화 하는 방향으로 움직이려는 경향이 있고, CPG간에 이루어진 neural network는 보행 자세를 자연스럽게 연결시켜주는 역할을 한다. 하지만 CPG으로만 구성된 로봇은 수많은 미지수들의 조합을 통해 반복 주기와 크기를 결정하기에 마치 태엽 인형처럼 처음 결정한 리듬의 주기와 크기로 보행하여서 방향을 바꾸거나, 속도를 높이거나 혹은 점프와 같은 자유로운 움직임에 적용하기에는 무리가 있다. 또한 보행에 필요한 원하는 자세나 속도 등을 얻기 위해서는 관련된 미지수를 정하는데 무수히 많은 시행착오가 필요하다. 선임 연구자들이 CPG 모델에 필요한 미지수를 정하는데 실험을 통해 지표를 만들거나 비선형인 모델에 적용하기 위하여 여러 가지 방법들로 미지수들을 결정하지만, 아직까지는 정형화 된 것은 없다. 따라서 본 논문에서는 CPG 모델을 이족 보행 로봇인 HYBRO모델에서 보행을 결정하는 양쪽 다리 관절에 전부 적용하지 않고, 각각의 골반과 무릎의 pitch direction에 적용하였다. 무릎 핀 보행을 위해 지지발의 inverted pendulum model과 골반의 pendulum model을 사용하여 발목 joint의 Pitch 방향 토크를 생성하고 동시에 골반의 Yaw 방향 토크를 적용하였다. 무릎과 골반의 CPG와의 조합으로 무릎 핀 보행을 할 수 있다. 이로써 기존 two links의 singularity 문제를 해결할 수 있다. 상체의 모션에서 팔과 허리의 움직임은 적용되지 않았다. IP model과 무릎 쪽 CPG와의 신경회로망을 통한 소통으로 CPG와 베이스 바디의 움직임을 자연스럽게 연결해 줄 수 있다. 또한 IP운동과 동시에 적용되는 골반의 움직임은 로봇의 보행 보폭을 보다 크게 향상 시킨다. 높이 1.2m, 중량 60kg인 HYBRO 모델은 RecurDyn version 6.3동역학 시뮬레이터와 MATLAB SIMULINK에서 3D 모델로 시뮬레이션 되었다|This article suggests a combination of CPG neural networks and kinetic trajectory planning for stretched-knee walking of a biped robot. Furthermore, this paper suggests ZMP compensation for stable walking. Instability when a foot swings during SSP(single support phase) and when the center of mass moves to another point was decreased by calculating the real ZMP coordinate with the slope of the base body with respect to the pivot point. Using the calculated ZMP and applying compensation for the forward and lateral directions at the knees and ankles in real time make it possible to maintain the stability of robot. Especially, how much the robot leans laterally toward the supporting foot can he determined by the ZMP in the lateral direction. From this, it is possible to get toward motions of the ankles in the roll direction. Recently, researches on biped robots using CPG have been intensified. The biggest merit of using CPG is that it can get a stable walking posture to overcome uncertainty and disturbance easily. Since a single unit of CPG generates a limit cycle, the movement can be easily stabilized. The neural network among CPGs also connects the walking posture smoothly. The robot only with CPG but nothing, however, determines the period time with combination of lots of unknown variables. Changing direction and velocity according to the initially determined period time like a mechanical doll or applying to free movement like jumping are very difficult. Furthermore, it needs trial and error to determine unknown variables for desired posture and velocity. However, there is no certain value needed for CPG model even though former researchers have been determining with lots of experiments. So, in the thesis, CPG model is used only to each pitch direction of waist and knees. Inverted pendulum model and waist pendulum model were used for generating joint and pitch torque. Simultaneously, yaw torque of waist is applied. Stretched knee walking has been succeeded with the combination of CPGs of knee and waist. As a result, the singularity problem of the legs has been resolved. The robot body movement was to be natural with the neural communication between trajectory generation and CPG pattern. The movements at the waist with IP movement was to make the locomotion stride wider. However, simulation of biped locomotion based on the trajectory failed due to the instability of locomotion could not be controlled.; This article suggests a combination of CPG neural networks and kinetic trajectory planning for stretched-knee walking of a biped robot. Furthermore, this paper suggests ZMP compensation for stable walking. Instability when a foot swings during SSP(single support phase) and when the center of mass moves to another point was decreased by calculating the real ZMP coordinate with the slope of the base body with respect to the pivot point. Using the calculated ZMP and applying compensation for the forward and lateral directions at the knees and ankles in real time make it possible to maintain the stability of robot. Especially, how much the robot leans laterally toward the supporting foot can he determined by the ZMP in the lateral direction. From this, it is possible to get toward motions of the ankles in the roll direction. Recently, researches on biped robots using CPG have been intensified. The biggest merit of using CPG is that it can get a stable walking posture to overcome uncertainty and disturbance easily. Since a single unit of CPG generates a limit cycle, the movement can be easily stabilized. The neural network among CPGs also connects the walking posture smoothly. The robot only with CPG but nothing, however, determines the period time with combination of lots of unknown variables. Changing direction and velocity according to the initially determined period time like a mechanical doll or applying to free movement like jumping are very difficult. Furthermore, it needs trial and error to determine unknown variables for desired posture and velocity. However, there is no certain value needed for CPG model even though former researchers have been determining with lots of experiments. So, in the thesis, CPG model is used only to each pitch direction of waist and knees. Inverted pendulum model and waist pendulum model were used for generating joint and pitch torque. Simultaneously, yaw torque of waist is applied. Stretched knee walking has been succeeded with the combination of CPGs of knee and waist. As a result, the singularity problem of the legs has been resolved. The robot body movement was to be natural with the neural communication between trajectory generation and CPG pattern. The movements at the waist with IP movement was to make the locomotion stride wider. However, simulation of biped locomotion based on the trajectory failed due to the instability of locomotion could not be controlled.