Skip navigation.


RFID-Based Localization

We are developing a system for highly accurate localization of parts that is accurate enough to be suitable for mobile manipulation. The method takes advantage of RF multipath effects and is robust to occlusion. We have demonstrated localization to an error of less than 2 cm and 4 degrees.


Failure Handling

We are looking at how robots can adeptly handle the failures that inevitably occur in the course of automated assembly.

Realtime Contextual Path Set Generation

Modern robotic motion planners typically split their work into three areas:

  1. Path sampling
  2. Path collision testing
  3. Simulation of robot/world interaction.

The path sampling process is usually completely uninformed, either sampling paths at random or using a predetermined low-dispersion sequence. By feeding back the results of earlier collision tests into the path sampling process, we can produce planners that are more both more efficient and more effective.

Path Diversity

Suppose you are designing a robot motion planner for a vehicle with highly sophisticated motion constraints. You want to precompute possible actions for the vehicle over short distances. Each action turns into a path describing where the robot moves. Although each path is a fixed length, the robot will compose paths at runtime to construct motions of arbitrary length. To stay responsive, the robot only has time to consider, say, 25 paths before committing to one alternative. What are the best 25 paths to choose from?

Personal Robotics

The goal of the Personal Robotics project is to enable robots to perform useful tasks in environments structured for humans. The project is building technologies to enable HERB, the Home Exploring Robot Butler, to function as an assistant to the elderly and the disabled. HERB can perform such tasks as opening the refrigerator, fetching a bottle of juice, and handing it to a person elsewhere in the room. I have fulfilled several roles on this project, such as providing the robot with navigation and exploring new approaches to manipulation planning.

People Prediction, Tracking, and Avoidance

Robots must behave predictably in order to be accepted in human environments. Part of predictability is that robots in turn must anticipate the motion of humans so that they can avoid causing unnecessary disturbance. For example, when approaching a person in a hallway, the robot should move to one side to allow the person to pass. In more open environments, such as the lobby of a hotel, people's paths can be harder to predict. By combining machine learning techniques with knowledge of the geometry of a given environment, the robot can form a statistical model of likely destinations for a person given their current trajectory. After constructing a time/space map of likely future person locations, the robot must then plan a path through the room that minimizes the chance interaction with the person while still reaching the robot's destination.

Mars Rover Motion Planning

Mars rovers face a hazardous environment for navigation. Although the Martian surface has been mapped, those maps are far less detailed than terrestrial maps, and the terrain on Mars is unstructured. A rover may encounter rocks, which can damage equipment or produce a high-centering condition in which the rover cannot get sufficient traction to get unstuck. The rover may also encounter loose soil. The Mars Exploration Rover, Spirit, became stuck in such soft soil and has been unable to break free. Slopes with loose soil present another kind of challenge, since the rover may slide downhill as it traverses a path.

These hazards present challenges in perception, modeling, and planning. This work focuses on the planning problem for Mars rovers. Suppose we are able to accurately detect such hazards at short range using the rover's vision system and that we can predict how the rover will interact with these arbitrarily-placed hazards? Our planner must be able to incorporate these high-fidelity terrain interaction models and generate a set of useful motions that approximate all possible motions through the observable area around the rover.

Autonomous Navigation System

The Autonomous Navigation System is capable of autonomously controlling any of several vehicles designated by the Army, including the Multi-functional Utility Logistics Equipment (MULE) platform, the Armed Reconnaissance Vehicle (ARV) and Manned Ground Vehicles. The ANS program provides navigational, perception, path-planning and vehicle-following algorithms, as well as the requisite on-board sensor package for autonomous mobility.

Syndicate content