A Framework for Evaluating Early-Stage Human - of Marcus Hutter

information about the scene is maintained in 3D) and from
the data coming form sensors and processed by the
subconceptual area.
Macaluso 2008). In brief, the algorithm generates a cloud
of hypothesized possible positions of the robot (Fig. 8). For
each position, the corresponding expected image scene is
generated in the iconic area by means of geometric
projection operations of the corresponding knoxels in CS.
The generated image is then compared with the acquired
image (Fig. 9).
Fig. 4. The 2D image output of the robot video camera (left) and the corresponding image
generated by Figure the simulator 9. The (right). 2D image from the robot video camera (left)
Figure 7. The revised architecture with the 2D iconic area
and the corresponding 2D image generated in the iconic
and the perception loop.
area by the knoxels of CS (right).
Fig. 4 shows the 2D image S as output of the robot video camera (left) and the
Fig. 7 shows the robot architecture revisited to take into corresponding image S’ generated by the simulator (right) by re-projecting in 2D the 3D
An error measure ε of the match is computed between the
account the iconic area: in the revised architecture there information is from the current point of view of the robot.
a perception loop between the conceptual area where the The comparator
expected
block
and
c compares
the effective
the two
image
images
scene.
of Fig. 4 by
The
using
error
elastic
ε
templates
weights the expected position under consideration by
knoxels are represented, the iconic area and matching the [2]. In the current implementation, features are long vertical edges extracted
considering the distribution of the vertical edges in the
subconceptual area. This perception loop has the role from to the camera image. Spatial relations between edges’ midpoints are used to locate
generated and the acquired images (mathematical details in
adjust the match between the 2D iconic representation
each
of
edge in the simulated image and compute the relative distortion between the
expected and Chella the effective & Macaluso scene. The 2008). relative In distortion subsequent is a measure steps, only of the the error ε related
the scene obtained from the knoxels in the conceptual area,
to the differences winning between expected the positions expected that image received scene the and higher the effective weigh scene. As
and the external flow of perception data coming out from
previously are stated, taken, this while error is the sent other back ones to the are simulator dropped. in order to correct the robot
the subconceptual area.
position in the 3D simulator.
We present the operation of the revised architecture with
reference to the CiceRobot robotic project, an operating
autonomous robot performing guided tours at the
Archaeological Museum of Agrigento (Chella & Macaluso
2008).
Fig. 5. The operation of the particle filter. The initial distribution of expected robot
positions (left), and the cluster of winning expected positions (right).
Figure 10. The image acquired by the robot camera along
4 Experimental with the vertical results edges and their distribution (left). The
simulated image from CS corresponding to the hypothesis
Figure 8. The initial distribution of expected robot In order to with test the the proposed highest architecture weight. (center) we compared A simulated the operations image of the robot
positions (left), and the cluster of winning expected equipped with corresponding the described to an system hypothesis with the with operations a lesser of weight the robot (right). driven by the
positions, highlighted by the arrow.
odometric information only.
When the image generated in the iconic area matches with
In order to compare the CS content with the external
the image acquired from the robot camera, the knoxels in
environment by the iconic area, the robot is equipped with
CS corresponding to the winning image are highlighted:
a stochastic match algorithm based on particle filter (see,
they give rise to the description of the perceived scene by
e.g., Thrun et al. 2005; details are reported in Chella &
17