Stephen Miller, Hertz Fellow 2011, has a passion for making robots interact with the real world. As an undergraduate at UC Berkeley he focused on physical manipulation of deformable objects, training a Da Vinci-style surgical robot to autonomously tie sutures, and the Willow Garage PR2 to pair socks and fold laundry. By grounding his research in practical applications, he hoped to identify crucial bottlenecks in current robotics research.
Now a PhD student at Stanford, his interests have shifted to what long nights of debugging have convinced him is the biggest bottleneck: visual perception. While research in the field has classically been split between 2D images (“Computer Vision”), synthetic meshes (“Computer Graphics”), and expensive laser range finders (“Robotics”), the lines are beginning to blur. The past few years have seen an outbreak in inexpensive sensors like the Microsoft Kinect, which provide registered color and depth images at real time speeds. These RGB-D (Red, Green, Blue, Depth) sensors have made it possible for consumer-grade products to reason about the 3D world explicitly, replacing “pixels” with “centimeters” for roughly the cost of a point-and-shoot camera.
But with this new modality comes new challenges: how to compensate for poor depth quality with an abundance of data, and how to adapt Vision, Graphics, and Robotics techniques to use it. To tackle these problems, Stephen's work has ranged from low-level data processing to high-level semantic understanding. At the lowest level he is interested in calibration, developing techniques to both intrinsically calibrate a moving sensor and extrinsically register multiple static sensors, each requiring no explicit human intervention. On the semantic level, he has looked at unsupervised object discovery and instance recognition. His current focus is on surface reconstruction: how to create clean, metrically accurate surface models given a handheld sensor and a non-technical user. In all of this, he hopes to develop usable tools, which make reasoning about the 3D world simple and intuitive: for consumer-grade product designers, frustrated Computer Vision students, and laundry-folding robots.
Stephen's research has been published in the International Journal of Robotics Research (IJRR), as well as the proceedings of the International Conference on Robotics and Automation (ICRA), International Conference on Intelligent Robots and Systems (IROS), Robotics Science and Systems (RSS), and the Workshop on the Algorithmic Foundations of Robotics (WAFR). His work has been featured on the Discovery Science Channel, CBS Smart Planet, and the New York Times; and parodied on Attack of the Show. When not programming he enjoys traveling, reviewing movies, and playing acoustic guitar.