基于交互目标视觉跟踪的共融机器人表达型拟人抓取动作规划研究

With the fast development of artificial intelligence and robot technology, there is more pressing demand to Tri-Co robot’s expressive anthropomorphic motion grasp planning in the areas of health care, education, entertainment and so forth. At present, domestic and foreign research is more concentrated in robotic pragmatic grasp motions, and the research of robotic expressive grasp motions faces with many difficulties and challenges, because of the intelligent perception in complex environment and the optimization of human-robot interaction. This project is oriented toward the interaction between human and robot in the human-robot integration environment, studies the visual tracking method to targets, and proposes the visual recognition and graspable region detection methods based on deep learning. The human-robot interaction platform is finally constructed to verify and evaluate the proposed theoretical results, and it could provide ideas, mechanisms, and scientific support for the research of the Tri-Co robot’s anthropomorphic interaction.

随着人工智能与机器人技术飞速发展，医疗、教育和娱乐等领域对共融机器人取动作抓取规划提出了更加迫切的需求。目前国内外研究主要集中在务实型抓取动作，而表达型抓取动作由于涉及复杂环境的智能感知与人-机器人交互优化，还面临诸多困难与挑战。本项目面向人机共融环境中机器人物品抓取的交互场景，研究复杂背景下目标物体的视觉跟踪方法，以及基于深度学习方法的视觉识别与可抓取区域检测方法，构建人-机器人物品交接的交互平台，并对理论成果进行验证与评估，为共融机器人系统的交互提供思路、机理和科学支撑。

随着人工智能与机器人技术飞速发展，医疗、教育和娱乐等领域对共融机器人手臂的抓取动作规划提出了更加迫切的需求。在非结构化环境机器人抓取任务中，获取稳定可靠目标物体抓取位姿至关重要。本项目面向人机共融非结构化复杂环境中机器人的抓取场景，首先设计了基于Faster R-CNN的多目标动态检测方法，提出了稳定抓取检测滤波器，抑制噪声与实时检测时的识别抖动问题；然后，设计了深度目标适配器，并在此基础上采用GG-CNN模型预测估算机器人二维抓取位姿；进而，融合目标检测结果、二维抓取位姿以及物体深度信息，重建目标物体点云，并计算三维抓取位姿；最后，搭建机器人抓取平台，验证了本文所提方法的可行性及有效性。本项目提出的方法克服了二维抓取位姿抓取方向固定且单一的缺陷，扩展了机器人抓取任务实现的灵活性。