结合深度学习的机械臂视觉抓取控制
|
摘要
针对基于视觉的机械臂抓取精确抓取的需求,考虑传统的视觉识别算法受环境、对象变化的制约且在识别正确率及快速性上存在的问题,在现有研究的基础上,提出了一种基于深度学习的目标精确检测与识别方法。首先基于深度学习改进了YOLO算法,通过对数据集的训练,基于英伟达Jetson TX1高性能处理单元实现了复杂环境下多目标的识别与定位,给出了目标的类别与位置等信息;以此为基础,结合利用Move It!功能包完成的机械臂运动轨迹的求解与规划,以及基于李群李代数建立的递推正逆动力学模型,设计了机械臂抓取控制的滑模控制律。仿真及实物验证表明,基于深度神经网络的方法学习到的特征对复杂背景具有较强的鲁棒性和稳定性;所设计的滑模控制算法在0.21 s时跟踪误差在2%,取得了较高的控制精度。可为后续视觉抓取任务提供参考。
The precise capture of the robotic arm plays a very important role in actual military or industrial production. Vision is widely used because of its high efficiency,low price,etc. Therefore,robotic arm grabbing based on vision has gradually become a research hotspot in this field. However,the traditional visual recognition algorithms are restricted by the changes of the environment and the objects. At the same time,there exist recognition accuracy and rapidity problems. Based on the existing researches,an algorithm based on deep learning was proposed in this paper to accurately detect and identify the target so as to improve the recognition ability in a complex environment. First,the YOLO algorithm was improved by deep learning. Based on the high-performance processing unit of Nvidia Jetson TX1,the identification and location of multi-targets in a complex environment were achieved through data sets training,and the information such as the category and location of the target were given. Then,based on the solution and planning of the trajectory of the robot arm solved by the Move It! Function Package,and the efficient recursive and inverse dynamic model established according to Lie groups and Lie algebra theory,the sliding mode control algorithm for robotic arm grab control was designed. The simulation and physical verification showed that compared with the traditional image recognition algorithm,the features learned from the deep neural network method have stronger robustness and stability for the complex background. The tracking error of the designed sliding mode control algorithm was 2% at 0. 21 s and with high control accuracy which may provide relevant references for subsequent visual grabbing tasks.
The precise capture of the robotic arm plays a very important role in actual military or industrial production. Vision is widely used because of its high efficiency,low price,etc. Therefore,robotic arm grabbing based on vision has gradually become a research hotspot in this field. However,the traditional visual recognition algorithms are restricted by the changes of the environment and the objects. At the same time,there exist recognition accuracy and rapidity problems. Based on the existing researches,an algorithm based on deep learning was proposed in this paper to accurately detect and identify the target so as to improve the recognition ability in a complex environment. First,the YOLO algorithm was improved by deep learning. Based on the high-performance processing unit of Nvidia Jetson TX1,the identification and location of multi-targets in a complex environment were achieved through data sets training,and the information such as the category and location of the target were given. Then,based on the solution and planning of the trajectory of the robot arm solved by the Move It! Function Package,and the efficient recursive and inverse dynamic model established according to Lie groups and Lie algebra theory,the sliding mode control algorithm for robotic arm grab control was designed. The simulation and physical verification showed that compared with the traditional image recognition algorithm,the features learned from the deep neural network method have stronger robustness and stability for the complex background. The tracking error of the designed sliding mode control algorithm was 2% at 0. 21 s and with high control accuracy which may provide relevant references for subsequent visual grabbing tasks.
引文
[1]邓桦.机械臂空间目标视觉抓取的研究[D].哈尔滨:哈尔滨工业大学,2013.Deng Hua.Research on Visual Capture of Robot Arm with Spatial Target[D].Harbin:Harbin Institute of Technology,2013.(in Chinese)
[2]于灏洋.基于双目视觉的伺服机械臂运动控制研究[D].大连:大连理工大学,2016.Yu Haoyang.Research on Motion Control of Servo Arm based on Binocular Vision[D].Dalian:Dalian University of Technology,2016.(in Chinese)
[3]Redmon J,Divvala S,Girshick R,et al.You only look once:Unified,real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2016:779-788.
[4]Redmon J,Farhadi A.YOLO9000:better,faster,stronger[J].ar Xiv preprint ar Xiv:1612.08242,2016.
[5]The WAMTMArm[EB/OL].https://www.barrett.com/wam-arm/.
[6]孟韶南,梁雁冰,师恒.基于ROS平台的六自由度机械臂运动规划[J].上海交通大学学报,2016(S1):94-97.Meng Shaonan,Liang Yanbing,Shi Heng.Motion planning of six degree of freedom robot arm based on ROS platform[J].Journal of Shanghai Jiao Tong University,2016(S1):94-97.(in Chinese)
[7]刘鹏,王强,张伟.机械臂动力学建模及控制仿真[J].自动化与仪表,2017(03):9-12.Liu Peng,Wang Qiang,Zhang Wei.Dynamic modeling and control simulation of mechanical arm[J].Automation and Instrumentation,2017(03):9-12.(in Chinese)
[8]邵兵,吴洪涛,程世利,等.基于李群李代数的主被动关节机器人动力学及控制[J].中国机械工程,2010(3):253-257.Shao Bing,Wu Hongtao,Cheng Shili,et al.The main passive joint robot dynamics and control based on the theory of Lie groups and Lie algebras[J].China Mechanical Engineering,2010(3):253-257.(in Chinese)
[9]邵兵.基于李群李代数的统一开闭环机械多体系统递推动力学研究[D].南京:南京航空航天大学,2010.Shao Bing.Research on Recursive Mechanical Mechanics of Unified Open Closed Loop Mechanical Multibody Systems based on the Theory of Lie Groups and Lie Algebras[D].Nanjing:Nanjing University of Aeronautics&Astronautics,2010.(in Chinese)
[10]Young K D,Utkin V I,Ozguner U.A control engineer’s guide to sliding mode control[J].IEEE Transactions on Control Systems Technology,1999,7(3):328-342.
[11]Yong F,Yu X,Man Z.Non-singular terminal sliding mode control of rigid manipulators[J].Automatica,2002,38(12):2159-2167.
[12]Jetson TX1 Module[EB/OL].https://developer.nvidia.com/embedded/buy/jetson-tx1.
[2]于灏洋.基于双目视觉的伺服机械臂运动控制研究[D].大连:大连理工大学,2016.Yu Haoyang.Research on Motion Control of Servo Arm based on Binocular Vision[D].Dalian:Dalian University of Technology,2016.(in Chinese)
[3]Redmon J,Divvala S,Girshick R,et al.You only look once:Unified,real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2016:779-788.
[4]Redmon J,Farhadi A.YOLO9000:better,faster,stronger[J].ar Xiv preprint ar Xiv:1612.08242,2016.
[5]The WAMTMArm[EB/OL].https://www.barrett.com/wam-arm/.
[6]孟韶南,梁雁冰,师恒.基于ROS平台的六自由度机械臂运动规划[J].上海交通大学学报,2016(S1):94-97.Meng Shaonan,Liang Yanbing,Shi Heng.Motion planning of six degree of freedom robot arm based on ROS platform[J].Journal of Shanghai Jiao Tong University,2016(S1):94-97.(in Chinese)
[7]刘鹏,王强,张伟.机械臂动力学建模及控制仿真[J].自动化与仪表,2017(03):9-12.Liu Peng,Wang Qiang,Zhang Wei.Dynamic modeling and control simulation of mechanical arm[J].Automation and Instrumentation,2017(03):9-12.(in Chinese)
[8]邵兵,吴洪涛,程世利,等.基于李群李代数的主被动关节机器人动力学及控制[J].中国机械工程,2010(3):253-257.Shao Bing,Wu Hongtao,Cheng Shili,et al.The main passive joint robot dynamics and control based on the theory of Lie groups and Lie algebras[J].China Mechanical Engineering,2010(3):253-257.(in Chinese)
[9]邵兵.基于李群李代数的统一开闭环机械多体系统递推动力学研究[D].南京:南京航空航天大学,2010.Shao Bing.Research on Recursive Mechanical Mechanics of Unified Open Closed Loop Mechanical Multibody Systems based on the Theory of Lie Groups and Lie Algebras[D].Nanjing:Nanjing University of Aeronautics&Astronautics,2010.(in Chinese)
[10]Young K D,Utkin V I,Ozguner U.A control engineer’s guide to sliding mode control[J].IEEE Transactions on Control Systems Technology,1999,7(3):328-342.
[11]Yong F,Yu X,Man Z.Non-singular terminal sliding mode control of rigid manipulators[J].Automatica,2002,38(12):2159-2167.
[12]Jetson TX1 Module[EB/OL].https://developer.nvidia.com/embedded/buy/jetson-tx1.