Spatial Understanding


If a vehicle spatially understands its surroundings, it significantly improves the quality of task performance. This page demonstrates the Swan robot’s useful performances with accurate geometric environmental maps. MotionLab’s Symmetric Geometry can create accurate geometrical maps of Swan’s surroundings.

[1] Vacuuming-area teaching and vacuuming-execution

Suppose a “vacuum-cleaning robot” has to clean a polygonal area in a human-made structure, as shown here:


To generate the map of the polygonal area, MotionLab adopts a method in which a human teacher teaches the map to the robot.

Andrew Harding teaches the area to the Swan,
while Swan follows him,
records vertex positions at each halt, and
extracts geometrical features using side sonars.

His hand signal tells the robot
that it comes back to the start point.

As a result, the Swan successfully constructs a vacuum-area map.

Using the map,
the Swan robot makes an inward spiral
to perform an “area-cleaning” task.

[2] Map-Based Navigation

We constructed a geometrical map of the MotionLab laboratory using the follow-the-teacher procedure. (These teaching steps are not shown here.) The map contains positions and names of several featured points: “home,” “sf,” “library,” “ny,” and “copier,” as shown in the diagram below. This entire set of information embedded in a graph supports the Swan robot as it navigates reliably.


We initially place the Swan at “home.”
Then, it is instructed to go to “library,”
next to “copier,”
and finally to come back “home.”

Dijkstra’s algorithm finds the minimum-cost
path using 
the map embedded in the graph.
The map governs each step of navigation.

During this navigation, the Swan dynamically localizes its frame (position and direction) to make the navigation reliable and precise. Having no absolute position sensors, the Swan robot must perform localization even for this apparently straightforward navigation.

[3] Human-assisted map generation for driverless cars

When a taxi driver moves to a new city, for example, Newtown, he or she will study streets and featured places such as the government offices, the airport, hospitals, parking lots, and grocery stores in Newtown.

On the other hand, if one wants to use a driverless car for taxi business in Newtown, the car may want a digital street map with additional data of Newtown as well.

To generate such a map, MotionLab proposes that the car owner adopts a similar teaching-learning procedure as shown in [1] and [2] above. However, in this case, he or she sits inside the car to command the car’s movement during the session. This method has the following crucial advantages:

  • The owner can help the car acquire necessary and sufficient data for the purpose because the teacher genuinely understands the purpose of the target task.
  • This digital map is ideal for the car’s purpose because the same car sensors are used both for map generation and taxi-business execution.

It is not realistic to expect the car to generate the map by itself without a human’s assistance. No computer software is smart enough to perform such a complicated, situation-sensitive teaching task.