|
|
Foundations of Robotics
Seminar, January 25, 2006
Time
and Place | Seminar Abstract | Speaker
Biography | Presentation Slides | Speaker
Appointments
Estimating the State
of Hybrid Systems:
HMM + Timed Automata = Context Identification
Sarjoun Skaff
Newell Simon Hall 1507
Refreshments 4:45 pm
Talk 5:00 pm
Hybrid systems such as jogging robots switch
between multiple modes of operation and are governed by different dynamics in
each mode. Estimating the state of such systems typically involves
multiple-model filters based on different dynamical models for each mode. The
combination of filters estimate which mode is in operation by comparing the
likelihoods of individual outputs, and assign the highest weight to the output
of the filter corresponding to the most likely mode.
Problems arise when the filters rely on
incomplete models, as they often discard sensor information and degrade the
accuracy of their estimates. For example, the high-order dynamics of mobile
robots such as RHex are difficult to model accurately. Therefore, a collection
of low-order models are used to approximate the dynamics, but their limited
expressiveness delays the selection of the correct model and reduces the
accuracy of the state estimate.
This talk presents a technique to quickly
and robustly identify the mode from sensor information that filters may be
unable to use. The idea is to extract the structure of sensor data and map it
to dynamics which can be modeled with sufficient accuracy. The set of
measurements corresponding to modeled dynamics is defined as a context, so
identifying the context enables the selection of the appropriate models and
prevents the use of knowingly inaccurate models.
Context identification is done with a
discrete-state estimation approach that combines Hidden Markov Models and Timed
Automata to extract the structure and temporal information from a data stream
and classify it into contexts. This combined approach avoids the complexity of
conventional techniques for capturing a system's temporal dimension (such as
Semi-Markov Procceses) and overcomes the Markov
limitation of one-step history.
Experiments suggest that this approach is
conducive to rapid and efficient extraction of the structure. Results are
presented for RHex, where the jogging and walking contexts are accurately
identified from accelerometer data.
|
Speaker Appointments |
For appointments, please contact Sarjoun
Skaff (sarjoun@cmu.edu)
The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.