Scalable appearance-based robot navigation

当我走过走廊时，一个个人头从小隔间和办公室里冒出来，所有的目光都转向我的方向。我的同事们有的笑，有的皱眉。一个看起来很害怕，然后逃跑了。我想，当你像机器人一样出现在办公室时，这就是会发生的事情。”这是Erico Guizzo最近在IEEE Spectrum（2010年9月）上发表的一篇文章的开头，他描述了自己穿着机器人化身上班的经历。故事的背景定义了一个直观的场景，说明在哪里以及为什么这项研究是重要和相关的。毕竟，如果有人能创造出一个为你“出现”的机器人，就会有大量的机会，包括在家庭或医院里陪伴和帮助老年人或身体残疾的人。

"As I come down the hallway, heads start popping out of the cubicles and offices, all eyes turning in my direction. Some of my colleagues laugh, some frown. One looks terrified and flees. That's what happens, I suppose, when you show up at the office as a robot". So began a recent article in IEEE Spectrum (September 2010) by Erico Guizzo, to describe his experience of coming to work in his robot avatar. The context of the story defines an intuitive scenario for where and why this research is important and relevant. After all, if one can create a robot that "shows up" for you, a great number of opportunities will open up, including providing companionship and assistance to the elderly or physically handicapped in a household or hospital. But wait, you don't want to be bogged down by your avatar with every tedious and repetitive move, do you? At the heart of this type of autonomous mobile robot is the research of robot SLAM (simultaneous localization and mapping). The research described in this proposal is concerned with a promising approach to SLAM, called appearance-based SLAM (aSLAM). In aSLAM, the robot uses a camera to acquire sensor data and describes places of interest not in terms of 3D landmarks as is done traditionally, but in terms of pictures taken at these places, resulting in simple control algorithms with improved performance. The proposed aSLAM framework departs from the current method of image description based on the so-called bag-of-words (BoW) or vector-quantized visual features, but instead uses carefully selected raw visual features, to address the perceptual aliasing problem of BoW. In addition, our aSLAM framework will be online, through the use of hashing techniques for matching images, rather than tree structures which require offline construction. The research will be conducted in both indoor and outdoor environments, on wheeled robots and a humanoid robot called NAO. With any luck, we may just produce a robot avatar that can work as your electromechanical proxy in the near future!