This page gives brief descriptions of some of the projects that I am, or have been, involved with. Some of these projects have their own set of web pages, with a much more comprehensive description of the work. If you're reading this page from a machine with a Washington University IP number, you will also see a list of available projects. These are projects that we haven't started yet (or that are currently in hibernation), and that we're looking for help with. If you see something here that interests you, read our rules for joining our group, and send me email. Current Projects Action Jackson Action Jackson is a robot abstract artist, designed to produce paintings inspired by those of Jackson Pollock. The robot has a moving painting head, supplied with various colors of paint, that moves over a canvas lying horizontally. By controlling the position and velocity of the painting head, and the color of paint, we can produce some interesting paintings. The goal of the project is to connect the painting system to interesting sensing devices, so that we can create pieces that are driven by activity in the real world. People: Topher McFarland, Bill Smart, Pehr Hovey Reinforcement Learning Control of a Small Inspector Satellite We are applying techniques from reinforcement learning to the problem of controlling Bandit, a small inspector spacecraft. The goal is to robustly navigate Bandit around its parent spacecraft, Akoya, based on a variery of reward criteria, such as minimal fuel use. The problem is interesting because Bandit moves using a set of eight cold gas thrusters, which delivery a highly stochastic amount of thrust with each activation. We believe that RL will deal better with this stochasticity than traditional satellite control techniques. While most of the work is done in simulation, the ultimate goal is to deploy the learned controllers on the flight version of the Bandit. People: Rob Glaubius, Bill Smart, Stu Glaser Learning Effective Gaits for Complex Robots Learning a good gait for a complex robot, such as a humanoid is extremely difficult. The state and action spaces of these systems is typically continuous and high-dimensional. The curse of dimensionality prevents us from learning a controller that covers the whole state space, so we must find some way of focusing our attention on the parts of the space where the system will actually operate. This project is looking at ways to identify and represent that manfold over which the state actually moves, and at using techniques from reinforcement learning to learn good controllers over this manifold. People: Tom Erez, Bill Smart, Yuval Tassa, Eitan Marder-Eppstein Aircraft Transparency Inspection This project is a collaboration with The Boeing Company to use computer vision and machine learning to find and quantify defects in aircraft windscreens. We take a high-resolution digital images of a large gridboard through the windscreen, and compare it to a reference image of the gridboard. We can use these images to esimtate the distorion cause by the windscreen. We then take this distortion map, along with a human-supplied rating for the windscreen, and use machine learning to learn a classifier. The goal is to have a fully automated system for evaluating the optical distortions in a windscreen, and for providing a rating that agrees with a human evaluation of the windscreen. People: Bill Smart, Robert Pless, Michael Dixon, Rob Glaubius, Chris Wilson Software for Robots How do we design effective software architectures and middleware for robot systems? We are mostly interested in architectures and middleware for mobile robots that give transparent access to sensors and actuators, support graceful degradation in the face of hardware and software failures, and allow different research groups to effectively share implementations of their algorithms. We believe that there is no One True Architecture, but that the best solution is to provide infrastructure and middleware that allows a research group to assemble their own architectures, based on the tasks that their robots are trying to perform. However, having some commonality at the lower levels (in data structures and primitive calls to the robot hardware) is essential if we are to be able to effectively share software between research groups. People: Fritz Heckel, Bill Smart, Nik Melchior Brain-Computer Interfaces for Robot Control This project looks at the problem of controlling a robot, or other physical device, using signals recorded directly from the human brain. We are collaborating with colleagues in Biomedical Engineering and Neurological Surgery to develop direct brain-robot interfaces that will allow subjects to control a mobile robot and an anatomically-correct robot hand using only their thoughts. People: Doug Few, Bill Smart, Zac Freudenburg, Robert Pless, Eric Leuthardt, Dan Moran, Tim Blakely Human-Robot Interaction How can we design effective interactions between humans and robots? We are particularly interested in the case where the robots are not very expressive (ie don't have human-like limbs and features), and the humans do not know much about robots. How can we get the robots to understand what the humans are trying to do, and to react accordinly? How can we get non-expressive robots to give the humans around them an idea of "what they're thinking", so that the humans feel more comfortable? We are taking ideas from psychology, philosophy, and performing arts to allow us to build better, more natural human-robot interfaces. People: Chris Wilson, Bill Smart, Anna Pileggi, Fritz Heckel Old Projects The Robot Photographer The goal of this project was to create a fully autonomous robot photographer, capable of taking candid, well-composed photographs of humans at social events. The robot wanders around the event, constantly evaluaing photographic opportunities. When it identifies a likely shot, it uses standard photographic rules (such as the rule of thirds) to compose the picture, moves into position, and takes the shot. The photographer was deployed successfully at a number of public events, including SIGGRAPH 2002, Washington University's 150th birthday celebration, a science-writers reception, and an actual wedding. People: Bill Smart, Cindy Grimm, Michael Dixon, Zach Byers, Kevin Goodier

Page written by Bill Smart.