多视角视觉目标生成的空间机器人操作学习  

Learning Space Robotic Manipulation via Multi-View Visual Goal Generation

作　　者：李林峰 王勇[1,2] 解永春 胡勇[1,2] 陈奥[1] LI Linfeng;WANG Yong;XIE Yongchun;HU Yong;CHEN Ao(Beijing Institute of Control Engineering,Beijing 100094,China;Science and Technology on Space Intelligent Control Laboratory,Beijing 100094,China)

机构地区：[1]北京控制工程研究所,北京100094 [2]空间智能控制技术重点实验室,北京100094

出　　处：《空间控制技术与应用》2022年第2期18-28,共11页Aerospace Control and Application

基　　金：国家自然科学基金资助项目(U20B2054)。

摘　　要：空间机器人操作一般采用多视角相机为机械臂提供视觉信息引导,全局相机引导远距离的前段操作,局部相机引导近距离的末端精细操作.针对多视角视觉信息融合与利用的自主性不足问题,提出一种多视角视觉目标生成的新方法,采用分层策略架构,将图像输入映射到SE(3)空间上,形成多视角视觉目标,并根据相对误差建立多视角目标决策机制.地面试验结果表明,该方法能够快速响应(>30 fps)多类不同尺度的操作对象,并对静态、动态遮挡具有较强鲁棒性,具备执行视触融合、长序列复杂操作任务的能力.For space robotic manipulation,multi-view cameras are generally used to provide visual information guidance for the robotic arm.In such a configuration,the global camera guides the long-distance operation while the local camera guides the proximity operation.To increase the autonomy level of multi-view visual information utilization,a novel method called multi-view visual goal generation is proposed in this paper,which adopts a hierarchical architecture.The goal generator maps the multi-view input images to visual goals expressed in SE(3)space.Furthermore,a multi-view goal decision mechanism is established according to the relative error of the generated visual goal.Experimental results demonstrate that the proposed method can respond quickly(>30 fps)to numerous manipulation objects at different scales,and meanwhile show the robustness to static and dynamic image occlusion.The capability on long-sequence complex manipulation tasks with visuo-tactile integration is also verified.

关 键 词：空间机器人操作 基于视觉的机器人操作 视觉目标生成 视触融合 多任务学习 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]