Research for 3D Space Map Acquisition for Robot by Combining Active Motion and Visual Perception

该研究项目旨在开发一种主动视觉策略，使机器人能够从其立体眼睛获得3D场景理解。也就是说，机器人使用自己的眼睛(相机)和自己的有意动作来构建未知环境的3D地图。使用视觉来了解外部世界的任务对于机器人这样的人工系统来说是绝对困难的，而我们人类在没有老师或事先知识的情况下很容易做到这一点。在这项研究中，我们给出了一个基于动作不变性的关键概念的答案。我们报告了一种在机器人中实现对三维空间的直接感知。我们提出了一种标定机器人立体眼睛的方法，以便在他的大脑中建立3D地图。它不需要校准对象的外部参考，而是主动运动和视觉感知的结合，这种思想在某种程度上是所谓的自校准。但是，在自标定中，3D地图的标定和后续的构建仍然是分开考虑的。在我们的方法中，我们将动作和视觉结合起来，实现了完全高效的标定和三维地图的构建。我们只依赖于校准结果的一致性。我们介绍了这样一个事实：“无论是在真实环境中，还是在机器人的3D地图上，静止的物体都不会移动，即使当机器人四处移动时也是如此。”这意味着我们接受构建的3D地图的校准误差和由此产生的几何失真。我们认为，对于3D地图来说，最重要的是机器人的预期运动与机器人感知的图像之间的一致性，并通过大量的仿真实验和在实际机器人上的实现证明了这一想法的可行性。

This research project aimed to develop an active vision strategy to let a robot obtain 3D scene understanding from his stereo eyes. That is, the robot constructs a 3D map of an unknown environment by himself using his own eyes (cameras) and his own intentional actions. The task to get to know the external world using vision is absolutely difficult for artificial systems, such as robots, while our human achieves it easily without teacher or advance knowledge. In this research, we placed an answer based on a key concept of the invariance against our actions.We reported a realization of the direct perception of three-dimensional space in a robot. We proposed a method to calibrate the stereo eyes of a robot to build 3D map in his brain. It does not need an external reference of calibration objects but a combination of active motion and visual perception.The idea is common to so-called the self-calibration to some extent. But, the calibration and the following construction of the 3D map still considered separately in the self-calibration. In our method, we combine action and vision to achieve totally efficient calibration and the construction of the 3D map. We rely only on the consistency in the calibration results. We introduced the fact that "stationary object both in the real environment and on the robot's 3D map never moves even when the robot moves around". This means that we accept a calibration error and resulting geometric distortion of the constructed 3D map. We consider that the most important for the 3D map is the consistency between the intended motion of the robot and the perceived image by the robot.We showed the feasibility of this idea with plenty of simulation experiments and implementation on actual robots.