My work in robotics spans the areas of manipulation, mobile robotics, and mobile manipulation. Within this broad range of areas, I have focused on problems where an algorithmic and minimalist approach yields a theoretical completeness or optimality result and provides a practical algorithm that we demonstrate on real robot hardware.

The end results of my research are algorithms that incorporate aspects of motion planning, mechanics, computational geometry, and feedback control. A minimalist approach brings focus to what capabilities are essential to accomplish a task. Sometimes this may be a self-imposed restriction, e.g., how can a general purpose manipulator perform a positioning task only using pushing or tapping; but in most cases this is directly influenced by using hardware that is as simple as possible.

• Topological mapping
• Mobile manipulation
• Coverage planning
• Manipulation
• Future directions

Our work in this area focuses on enabling a team of sensing-limited mobile robots to cooperatively map an indoor environment. The robots we consider have sparse and limited-range sensing; we have built robots that use only five infrared rangefinders with a maximum range of 80 cm. These sensing limitations can arise from the need to build inexpensive robots, e.g., disposable robots to evaluate a contaminated building. They also serve to place a focus on what is absolutely necessary for mapping.

We have developed strategies for a single such robot to map an environment [WAFR 2004, ICRA 2005]. One fundamental problem is “closing the loop,” i.e., deciding when the robot has returned to a place it has already visited. With limited-sensing, it is not possible to uniquely recognize a place based on sensor readings from one location, so our solution uses odometry almost exclusively. Our method is the first evidential approach to making loop-closing decisions. Another problem we have addressed is mapping open spaces, commonly considered a weakness of traditional topological maps. We have solved this problem by using wall-following to map the perimeter and “forays” through open space to create connections across the interior. Our goal is optimal navigation between explored areas of the environment by combining forays with a wall- and hall-following map to create the subset of the visibility graph navigable by a robot with sensing limitations. This work is among the first to address exploration, mapping, and navigation with sparse sensing. We have demonstrated our approach to basic mapping both in simulation and on a real robot.

For multiple robots to cooperatively map an environment, they should explore separate but overlapping regions. They must then be able to merge their individual topological maps into a single global map. We have developed a merging algorithm [DARS 2004, IJRR submitted] that combines ideas from graph matching and image registration to solve this problem. While not the first to address this problem, our algorithm is the most general and robust algorithm published to date. Our experiments have shown our algorithm to be both accurate and efficient for simulated and real-world maps.

Mobile manipulation

We built a set of differential-drive robots that have a 2-DOF manipulator; the end-effector is a “palm,” here, simply a flat plate. Since this manipulator cannot grasp objects, it must work in concert with fixtures in the environment or with another robot in order to lift or carry objects.

We developed an offline planning algorithm for a single such robot to pick up an object by pushing it against a wall, sliding the palm underneath the object, sliding the object up the wall, and tilting it away from the wall so it rests entirely on the palm. This planner incorporated a model of frictional contact in order to plan a sequence of motions for the robot which were then executed open loop. We implemented and demonstrated this planner on one of our robots [IROS 2001].

Next, we turned our attention to the problem of carrying an object with two robots, where each robot supports one end [IROS 2003]. As the robots move, errors in their motion will cause the object to slip on the palms and eventually fall. The robots must make corrective actions to maintain a stable “carry.” We first formulated a centralized policy that determines the robots' actions based upon the current sensor readings from a simple tactile sensor on each palm. We showed that this policy is correct through analysis of the task mechanics which include frictional contact at each palm. We then found that this policy could be transformed into a distributed policy where some errors can be corrected independent of the other robot and other errors require cooperation. The two robots communicate only when necessary to coordinate their actions. In our experiments, the robots were able to carry different sized objects without modification to the algorithm [TRO, submitted]

In the interest of exploring more versatile mobile manipulators --- manipulators in which there is little separation between manipulation and mobility --- we studied a novel robot that uses its wheels for both. This robot is a car-like robot with a chassis short enough that an object can be placed on top of its wheels. This robot can perform maneuvers such as parallel parking and three-point turns to position the object locally, but we have focused on handing off the object from one robot to another. As the robot drives forward, the object rolls off the front of the first robot, so we place a second robot there to receive the object. We have found that we can plan a path for the object as though it were a car-like robot, and then decompose that path into segments for individual robots that alternately hand off the object to each other [ICRA 2003]. Our current work with this system is developing feedback control to minimize errors in the object configuration that accumulate with handoffs.

Coverage planning

My Ph.D. work studied tapping for positioning planar parts on a support surface: first striking the part to give it some initial velocity and then letting it slide until it comes to rest due to friction. While others had previously built robotic systems that manipulated parts by tapping, I was the first to do a formal analysis and develop a planning algorithm and feedback control strategies for this mode of manipulation. My planner uses models of friction and impact to determine how a part should be struck to send it to a desired goal configuration. However, since it is not possible to configure an part arbitrarily with a single tap, I developed feedback control strategies that plan a sequence of taps to reach a goal and compensate for actuation errors during plan execution.

One of the main conclusions of my dissertation was that tapping can be used to precisely position a part, even if the actuator is not positioned accurately. This result led to my first research project at RPI: a tapping micropositioning cell. Here, a small number of fixed position (and orientation) tapping actuators are placed about the perimeter of the cell. A part is placed in the cell, and by repeatedly measuring the part configuration and firing one of the actuators, the part can be positioned very accurately. Such a system could be used for parts feeding on assembly lines or for part alignment in manufacturing processes. I formulated an algorithm to demonstrate that a system with only three actuators positioning a circular part is controllable [ICRA 2000]. In subsequent work [ISATP 2003], I built this system and successfully demonstrated it.

I have applied my tapping mechanics to a related research topic: vibratory manipulation. Many real parts feeding systems make use of vibration to manipulate parts. The interaction between the part and the manipulator is complex because the shape and motion interact to affect the resulting impacts. My work is some of the first analysis of this mode of manipulation, characterizing gross object motion in terms of manipulator shape and vibration parameters [IJRR submitted].

An earlier project (in collaboration with S. Akella, K. Lynch, and M. Mason) studied the use of a single joint manipulator, positioned over a constant velocity conveyor belt, to position and orient parts for parts feeding applications [Algorithmica 2000, WAFR 1996, ISRR 1996]. The manipulator consists of a rotating fence that can push a part up the conveyor and “catch” it again when it reaches the fence. We proved the ability to feed a part from an arbitrary configuration on the conveyor, implemented a planner for this system, and demonstrated an experimental implementation. This work showed that the mechanics of pushing could be combined with an extremely simple manipulator to produce a flexible and provably complete robotic system.

Future directions

Sensing-limited robots will have to work cooperatively in order to compete with their rich-sensing endowed cousins. To this end, I am developing collaborations to work on (active) sensor networks. There are numerous applications that require mobile sensor nodes or that would benefit from having robots that can move among and interact with stationary nodes. I expect to extend our research in topological mapping and coverage to work with communication constraints or to enhance the functionality of a sensor network.

I will also continue to work on mobile manipulation. While the boom in sensor networking research has brought attention to monitoring and surveillance tasks, the fundamental use of mobile robots will be to go places and do things that require manipulation --- from cleaning up your house to building a base on another planet. Versatile manipulation from a mobile platform, combined with mapping, navigation, and communication capabilities, will make all this possible.