Navigation and Wayfinding in Robot teams
A team of robots is deployed to cooperatively generate a map of the area, an area under reconnaissance or an urban disaster site, for example. The objective is to generate an accurate map showing hazards, obstacles, traversable routes, etc., very quickly and to communicate it back to a command center. The map will then be used by a combination of human and robot teams for effective operations in the mapped area.
A team of 16 Pioneer AT3 robots is used as the base for these experiments. Some robots are equipped with laser rangefinders, some with sterevision, some with cameras an sonar. Although each robot carried an onboard computer, the team communicates with an 88 core (11 HP DL160 G5 quad-core dual-processors servers) cluster for performing landmark detection, recognition, mapping , coordinate and reporting.

Terrain Spatiograms are a unique landmark representation in which the image spatial information is replaced by terrain spatial information. We havepresented experimental results for mutual landmark recognition on two different model robots equipped with different stereocameras, and with terrain spatiograms collected on one robot being used on the other and more recently showed that because this representation preserves depth structure, it is possible to identify and filter potential occlusions.

Agile, Legged Locomotion
We have developed a novel tripedal design for energy-efficient legged locomotion. The platform is called a rotopod, because its principle mechanism of motion is rotation around leg endpoints.

What is it?

• The Rotopod is a novel robot mechanism which combines the features of wheeled and legged locomotion in an unusual way.
• This robot has the advantage of legged locomotion in stepping its 1-DOF legs over objects, but its drive mechanism is a rotating reaction mass that rotates the robot, in a controllable fashion, around each of its legs, similar to a rotating wheel.
• The mechanism has the potential to transfer the energy from the rotating reaction mass in an efficient manner to the legs, effecting a spinning forward motion

How does a Rotopod Move?

• The Rotopod moves by rotating around each of its legs in turn.
• A rotation around a leg is call a step, and it results in the center of the robot moving a distance
• A regular sequence of steps is called a gait.
• The robot can produce a broad set of gaits: stepping for various values of q on one leg, or any sequence of legs
• One of the most interesting gaits is spiral walking. This is a very natural behavior for the rotopod mechanism. The final result looks a little like a prolate cycloid. It leads to paths that have a width element as well as a length (i.e. a fractal dimension) and hence may be very efficient for covering space (e.g., searching, surveillance, exploring, etc.)
• Sharp turns are not necessarily a problem for a rotopod, since its continual rotation allows it to change direction dramatically under certain conditions.

Cognitive Robotics (with P. Benjamin, PACE).
One of the objectives of Cognitive Robotics is to construct robot systems that can be directed to achieve real-world goals by high-level directions rather than a complex, low-level robot programming. Such a system must have the ability to represent, problem-solve and learn about its environment as well as communicate with other agents. In previous work, we have proposed ADAPT, a Cognitive Architecture that views perception as top-down and goal-oriented, so that perception becomes part of the problem solving process. This approach is linked to a SOAR-based problem-solving and learning framework.
Consider an object, such as a ball, moving in cluttered environment and consider a robot given the objective of intercepting the object. The robot must track the object as it moves and determine how to intercept it. However, in a cluttered environment, it is highly likely the object will collide and rebound from the environment. Tracking can at most handle this on a collision-by-collision basis. In our approach a world modelling system - a 3D game engine - is intimately linked into the visual perception process. Predictions of complex object behavior such as repeated rebounds from the environment are generated from the world modelling system at very high rates and are available as 'imagined' perceptions, directly comparable with actual perceptions.

We have developed an approach to tracking targets with complex behavior, leveraging a 3D simulation engine to generate predicted imagery and comparing that against real imagery. In this approach, the salient points in real and imaged images are identified and an affine image transformation that maps the real scene to the synthetic scene is generated. An image difference operation is developed that ensures that the matched points in both images produce a zero difference. In this way, synchronization differences are reduced and content differences enhanced.

Sensory Fusion for Multiple Target Tracking
Most existing visual tracking systems do not handle crowded scenes well. Our goal is to develop algorithms that take multiple sensory cues from the video (e.g., target locations, colors, shapes, etc) and fuse this information to robustly track in crowded scenes. We focus on the issue of occluding targets - since this is where a lot of the difficulty in vsiual tracking arises. We use sensory fusion to disambiguate occluding targets. This is a difficult problem, since the process of occlusion gives rise to dramatic and non-linear changes in the feature values. We exploit an approach that determine which cues to use and how to best combine them by looking at the distribution of feature measurement values to candidate targets - the so-called rank-score behavior. Experimentally we have shown that this approach, which we call the Rank and Fuse approach improves on a weighted sum or mahalanobis-sum for fusion.

Automated Management of Multiple Camera Resources.
Our goal is to automate the process of switching between multiple cameras when (manually or automatically) tracking a target. A major question in this is to understand the connectivity between camera views. We have developed algorithms and set of software libraries to automatically learn (using a NN) the candidate handoff cameras for each camera in a building. The cameras do not need to have overlapping views, exist and entrances can be anywhere in the field of view, and no map is needed. Future work will include software to periodically update the handoff information to account for camera or building changes.

Combining Recognition with tracking: Discrete-Event Modeling of PTZ targets
Most PTZ tracking systems decide when to pan, tilt or zoom based only on providing the best operator view of the target. While the operator view is clearly an important end-goal for tracking, it is not the only constraint that needs to be acknowledged. A second constraint is that the tracker be able to robustly recognize the target. There is no reason that these two constraints should always agree, and ignoring the second constraint means the operator may get an excellent view of the wrong target! We have developed a discrete-event control approach to modelling the target shape and color in such a way that we can determine when we need to zoom to maintain recognition of the target as well as maintain the operator view. Future work involves extending the discrete-event model to a hybrid model to allow fine control of PTZ.