We present a robust markerless image based visual servoing method that enables precision robot control without hand-eye and camera calibrations in 1, 3, and 5 degrees-of-freedom. The system uses two cameras for observing the workspace and a combination of classical image processing algorithms and deep learning based methods to detect features on camera images. The only restriction on the placement of the two cameras is that relevant image features must be visible in both views. The system enables precise robot-tool to workspace interactions even when the physical setup is disturbed, for example if cameras are moved or the workspace shifts during manipulation. The usefulness of the visual servoing method is demonstrated and evaluated in two applications: in the calibration of a micro-robotic system that dissects mosquitoes for the automated production of a malaria vaccine, and a macro-scale manipulation system for fastening screws using a UR10 robot. Evaluation results indicate that our image based visual servoing method achieves human-like manipulation accuracy in challenging setups even without camera calibration.

翻译：我们提出了一种鲁棒的无标记图像视觉伺服方法，可在1、3和5自由度下实现无需手眼标定和相机标定的精密机器人控制。该系统使用两台相机观察工作空间，并结合经典图像处理算法与基于深度学习的方法来检测相机图像中的特征。对两台相机放置位置的唯一限制是相关图像特征必须在两个视野中均可见。即使物理配置受到干扰（例如相机被移动或操作过程中工作空间发生偏移），该系统仍能实现机器人工具与工作空间的精确交互。该视觉伺服方法的实用性通过两个应用场景得到验证与评估：一是用于解剖蚊子以自动化生产疟疾疫苗的微机器人系统标定，二是基于UR10机器人进行螺丝紧固的宏观尺度操作系统。评估结果表明，即使在无相机标定的挑战性场景下，我们提出的图像视觉伺服方法也能实现接近人类水平的操作精度。