Toyota AI Seminar | Place Recognition from Perceptually Ambiguous Data

The ability to recognize places is a critical component of robot mapping
and navigation systems. Unfortunately, place recognition is complicated
by two major problems. The first challenge is that the uncertainty in
the robot's position (as provided by dead-reckoning sensors) grows very
quickly. When recognizing places, this uncertainty means that the robot
must consider a large number of possible answers. The second challenge
is that current sensing systems often produce similar readings for
distinct places, making it difficult to distinguish subtly different
environments. This problem is further compounded by the need to
robustly recognize places even if the appearance of those places
changes.

In this talk, we describe several approaches that can be used to
reliably recognize places even when the prior uncertainty is large and
when the sensor data is highly ambiguous. Our approach combines data
collected over a large spatial area and uses a spectral graph method to
identify a cluster of data that is maximally self-consistent. The
resulting inlier clusters must also satisfy a
probabilistically-motivated geometrical sufficiency test. The resulting
method allows robots to recognize places more often and more robustly,
leading to improved navigational performance.

Other ongoing robotics-related research projects will be briefly described, time permitting.