딥러닝 기반 상태 추정기를 사용한 모델 예측 제어와 전신 제어를 통합한 보행 제어 전략 (Integrated Walking Control Strategy With Model Predictive Control and Whole–body Control Using Deep Learning–based State Estimator)

서지정보

· 발행기관 : 제어·로봇·시스템학회
· 수록지 정보 : 제어.로봇.시스템학회 논문지 / 30권 / 10호 / 1139 ~ 1146페이지
· 저자명 : 전상현, 권나현, 하성민, 김정엽

초록

This paper proposes a walking strategy that integrates Model Predictive Control (MPC) and Whole-Body Control (WBC) based on a single rigid body model. Additionally, it develops and utilizes a novel state estimator based on LSTM deep learning instead of the commonly used Kalman filter–based state estimator. MPC and WBC are methods that enable walking robots to perform various tasks without falling. Therefore, this study uses MPC based on a single rigid body model to derive the optimal ground reaction force. Subsequently, in WBC, constraints are added to satisfy robot dynamics equations and track the optimal ground reaction force. Finally, through WBC, the optimal joint torque is computed considering whole-body floating dynamics and optimal ground reaction forces. All of these computations are performed based on the newly developed deep learning–based state estimator. The proposed control strategy is validated through Gazebo simulation using a quadruped walking robot, Go2.

영어초록

This paper proposes a walking strategy that integrates Model Predictive Control (MPC) and Whole-Body Control (WBC) based on a single rigid body model. Additionally, it develops and utilizes a novel state estimator based on LSTM deep learning instead of the commonly used Kalman filter–based state estimator. MPC and WBC are methods that enable walking robots to perform various tasks without falling. Therefore, this study uses MPC based on a single rigid body model to derive the optimal ground reaction force. Subsequently, in WBC, constraints are added to satisfy robot dynamics equations and track the optimal ground reaction force. Finally, through WBC, the optimal joint torque is computed considering whole-body floating dynamics and optimal ground reaction forces. All of these computations are performed based on the newly developed deep learning–based state estimator. The proposed control strategy is validated through Gazebo simulation using a quadruped walking robot, Go2.

참고자료

· 없음