We used a Nomadic Scout outfitted with an on-board 1.2 GHz processor. A firewire camera with an
overhead omin-directional mirror was used to obtain range and bearing measurements to engineered
landmarks in the form of pink boxes, which were placed throughout each hallway.
The robot simultaneously constructs a local obstacle map using sonar data and a local feature map
using vision data. The local obstacle map is used to explore and navigate each GVG edge by maintaining
two-way equidistance. When the robot becomes close to a third obstacle it invokes a homing control
law which drives it to the meet point location, where the robot decides which edge to explore next.
The edge tracing for the GVG is complemented by the arc transversal median (ATM) method, which
improves the azimuth resolution of the sonar data by a factor of ten. In addition, to make
exploration more robust, we have characterized the meet points (or nodes of the GVG).
When the robot determines that a meet point has exactly two non-terminal edges it deems this meet
as "weak" and does not store it in the graph. These weak meet points often occur due to doorjambs
and corners, but can also be caused by sensor noise. The subset of the GVG with weak meet points
removed is called the reduced GVG or RGVG. The effects of weak meet points can be seen as cusped
deviations from the center of the corridor.
When a meet point is categorized as strong, the robot takes a probing step in each edge direction
to enhance the local sensor map. Then the meet point is either verified as strong or discarded as
weak. Hence spoke like motions occur at strong meetpoints.
The animation above shows the resulting RGVG of six strong nodes: three boundary nodes and three nodes
of degree three, arranged in a configuration having one cycle. For this experiment, a user prompt
"oracle" mechanism notified the system when it revisited meet point one to close the cycle in the GVG.
The autonomous topological matching procedure with the additional edge-map information is nearing
completion.
The drastic effects of noise in odometry when turning is evident in the result. As the robot successively
traces edges, the ability to calculate its global location deteriorates rapidly.
It is important to note that no data were ever recorded in a global frame of reference. Rather,
these plots were created by concatenating the reverse transformations from one frame to another.
The resulting feature-maps for this experiment are shown below. The first image shows the feature-maps embedded into a global map by tying the sub-maps to meetpoint locations. Meet point locations are shown with encircled asterisks and landmark locations are shown as triangles.
The next image shows a feature-map as it is stored in memory with the origin defined by the relative locations of the two end meetpoints. The origin of the frame is also shown in the previous image as it is embedded. Landmark locations are shown by crosses along with covariance ellipses.
Go back to: H-SLAM Experiments Page | H-SLAM page
Jump to: People | Robots | Research | Events | Projects | Education | Papers | Software | Links
© Copyright 2002 Sensor Based Planning Lab, Carnegie Mellon University. All Rights Reserved.
 |
 |
 |