Learning methods and new sensors for autonomous robots

研究使用宾夕法尼亚大学开发的一些新的生物启发传感器的可能性，例如跟踪和方向传感器。现在，我们的大型机器人使用连接到数码相机的笔记本电脑对视频图像进行全面分析。大部分的处理都花在寻找图像中的场特征（线、角、颜色过渡）上，以便找到机器人的位置。使用Penn的跟踪传感器，可以从现场的已知位置开始跟踪机器人的运动，以这种方式获得独立于机器人里程计的更新。到目前为止，Penn的传感器一直用于跟踪在固定系统前面移动的物体-我们的系统将移动，而环境将是固定的。Penn的方向传感器还检测我们感兴趣的一些低级特征（线条，边缘，线条停止），探索如何将其集成到我们当前的系统中会很有趣。B）研究自主机器人的自动颜色和光线校准。我们的机器人现在是手动校准的，即由操作员根据他在计算机屏幕上看到的内容调整相机和颜色表。这是所有其他RoboCup团队所遵循的方法。我希望能够把一个“天真”的机器人放在球场上，它环顾四周，并根据它对球场几何形状和颜色的了解自动校准自己。机器人将学习相机所做的几何投影的特征，并根据其在场地上的位置生成颜色表。主要目标是能够将机器人从盒子中取出，几秒钟后，机器人应该能够探索周围环境并玩耍。这种学习方法可以扩展到其他机器人，比如索尼的腿式机器人。我还想测试一些新的传感器来完成这项任务，比如斯图加特微电子研究所（IMS）的视网膜相机。c）使用强化学习研究机器人的控制。到目前为止，我们都是手工编写完整的控制例程。每次我们改变电机或硬件的某些部分时，我们都必须调整机器人的参数或编写新的控制程序。我想让机器人学习如何自动移动，通过在球场上移动和翻滚，在开始的时候。如果（B）点成功，机器人应该能够学习如何移动，如何为电机提供动力，以及如何制动。

Investigate the possibility of using some of the new biologically inspired sensors developed at Penn, such as the tracking and orientation sensors. Right now, our larger robots do the full analysis of the video images using a laptop connected to a digital camera. Most of the processing is spent finding field features in the image (lines, corners, color transitions) in order to find the position of the robot. Using Penn's tracking sensor it would be possible to start from a known position in the field and track the movement of the robot, obtaining in this way an update independent of the robot's odometry. Penn's sensor has been used until now to track objects moving in front of a fixed system - our system would be moving and the environment would be fixed. Penn's orientation sensor detects also some of the low-level features we are interested in (lines, edges, line stops), it would be interesting to explore how to integrate it in our current System. b) Investigate automatic color and light calibration for autonomous robots. Our robots are calibrated now by hand, i.e. by an operator that adjusts the camera and a color table according to what he sees on a computer screen. This is the approach followed by all other RoboCup teams. I would like to be able to put a "naive" robot on the field, that looks around, and from its knowledge of the geometry and colors of the field calibrates itself automatically. The robot would learn the characteristics of the geometric projection made by the camera and would generate a color table according to its position on the field. The main objective would be to be able to take the robot out of a box, and after a few seconds, the robot should be able to explore its surroundings and play. This learning approach could be extended to other robots, such as the Sony legged robots. There are also some new sensors I would like to test for this task, such as the retina inspired camera of the Institute of Microelectronics in Stuttgart (IMS). c) Investigate control of the robot using Reinforcement Learning. Until now we have written the full control routines by hand. Every time we change motors or some part of the hardware, we have to adjust parameters of the robot or weite new control routines. I would like to let the robot learn how to move automatically, by moving on the field and tumbling around, at the beginning. If point (b) succeeds, the robot should be able to learn how to move, how to provide power to the motors, and how to brake.