Challenges in Human-Robot Interaction: Rethinking Robot Eyes

Every year the Pohang Institute of Intelligent Robotics hosts the Korean Robot Grand Challenge Competition, where autonomous robots from several teams compete in sort of scavenger-hunt game. Robots try to reach a specified floor of a building by taking the elevator, find a specific object (pack of milk, ball, pen...) in one of the rooms and bring it back to the starting point. Needless to say that no team ever succeeded yet in completing this challenge (though similarly complex challenges as the DARPA one, has been completed in the past). What is remarkable about this competition is not much the results but the challenge it proposes (a great challenge that is still unsolvable), highlighting and emphasizing few avenues of research necessary for the Human Robotic Interaction. In fact, although people give often for granted that robots can easily navigate in a space meant for humans, this is often not the case, as I have also discussed in my previous post. Indeed, autonomous navigation remains still a complex task and an open research area.

Dr. Jean-Charles Bazin, graduated from KAIST in the Robotic and Computer Vision laboratory and now working on his postdoc in the Computer Vision laboratory of the University of Tokyo, is one of those researchers, whose work will surely help for completing the Korean Robot Grand Challenge. Dr. Bazin has been working now for several years on a broad spectrum of computer vision techniques for robotic navigation, suggesting that we should rethink how robot see and how the robot eyes work. His main idea is that a technological breakthrough that changes the way robot can see would lead to new opportunities and ways to deal with common robotic vision problems. Consider the camera, the technology behind almost every robotic vision system. Despite the technical improvements of camera technology in hundred years, camera still work according to the same basic principle, known as the pinhole principle: basically the image capture by a camera is a small window to the world in a particular direction, pretty much as it happens with human eyes. Therefore, what would it be if robots could have a 360 degree vision like flies for examples? The amount of data that they could be parsing in real time would be much higher (the ability to visually explore a larger space), probably leading to a better navigation system.

Dr. Bazin has hence worked for several years on the way to enable a 360 degrees vision (omni-vision) to robots both using multiple cameras (camera cluster is the way that Google Map works) or catadioptric cameras (basically a camera pointing to a concave mirror). By building a system that maps any camera frame to a spherical representation, and an algorithm to detect the camera orientation from the curved lines in the image, he was able to provide a framework for robots navigation based on computer vision. In fact, this system has been mounted on some of the robots at KAIST and Tokyo University and results have been obtained for various applications like visual odometry (estimating the camera trajectory from a sequence of images), 3D reconstruction (build the 3D model of an object or the scene) and e-heritage (digital preservation of archeological sites). He is currently investigating additional applications of omnidirectional vision such as video stabilization, robot guide and photometry.

So, where the future will lead us? Undoubtedly in the future, due to better hardware and algorithms, we will have smarter autonomous robots. However, we can also anticipate some problematic issues: who will be responsible for the mis-behavior of autonomous robots? The owner? The robot itself? The person interacting with it? The engineer who programmed it? Furthermore, better vision system necessarily raise concern on the privacy for image acquisition: who will be the owner of that visual material? Dr. Bazin and I were talking about these issues when we realized that there was a camera-mounted kiosk next to us: the future of computer vision, technically but also philosophically, is near.

Further readings:

Catadioptric Vision for Robotic Applications: From Low-Level Feature Extraction to Visual Odometry by Jean-Charles Bazin.