|
World Modeling by Manipulation T. Debus*, P. Dupont*, R. Howe** |
 |
| MOTIVATION |  |
|
| Teleoperation is the ideal method to perform sophisticated tasks in unknown environments. In this control mode, all the operator’s actions are based on his/her interpretation of the camera images. In many cases, however, humans are not good at quantification and cannot interpret this visual information in an exact manner, prohibiting them from designing the correct strategy. | ||
 |
PROPOSED SOLUTION |
|
The proposed solution is to create a modeling
system that will use sensor information (i.e., machine perception) used
for control purposes to determine the properties of the objects and environment
as the task progresses. One practical application is a teleoperator assistant
that can provide information to the operator in real time during task
execution. This includes quantitative measurements such as the size, shape,
and orientation of objects; these properties can, for example, help determine
appropriate insertion strategies in teleoperated assembly task. |
|
| Algorithm
used to Automatically Identify Object Properties
|
To automatically identify object
properties, two basics and essential concepts are used: 1) every task can be described as a network of states 2) properties are only measurable in certain states. Based on these two concepts, a framework has been developed to reduce the overall problem to three sub-problems: |
|
Task decomposition: The operator decomposes the task into a minimal sequence of subtasks characterized by their contact states geometry and associated sets of properties. Data segmentation: Given a task decomposition and the sensor data stream, find the time intervals corresponding to each subtask. Property estimation: Given the time intervals associated with each subtask, estimate the desired properties. |
 |
 |
The Mathematical Approach: Constraint Equations and Multiple Model Estimation |
| In order to stay in contact with the environment
the robot’s wrist has to be constrained. Moreover, how the wrist is
constrained can be a unique characterization of the contact state. This
can be modeled thru physical models using constraint equations parameterized
by sensing data and object properties. Once a mathematical model is built for each contact state, a multiple regression algorithm is used to estimate all the properties associated with all the contact states. Based on the quality of the estimates an acceptance test decides if a contact is active or not. Once we know which contact is active we can keep the associated estimated properties. |
|
To get more detailed information on the techniques used
to automatically identify remote object properties, here is a list of
references.
|
| REFERENCES |
T. Debus, P. Dupont and R.D. Howe (2002). T. Debus, P. Dupont and R.D. Howe (2000). T. Debus, J. Stoll, R. D. Howe, and P. Dupont (2000). T. Debus, P. Dupont and R.D. Howe (1999). P. Dupont, T. Schulteis, P. Millman, and R. D. Howe (1999). |