Autonomous Robotics: Action, Perception and Cognition: Analogies

Autonomous Robotics:
Action, Perception and
Cognition:
Analogies with nervous
systems
Gregor Schöner
Institute for Neural Computation
Theory of Cognitive Systems
[email protected]
Human movement
Human movement
“homo faber”: we are the skillful species who
make things
using fluent sequences of movement, linked online to sensory information
multiple motor skills which can be adapted and
be performed concurrently
excellent fast scene perception
fine (compliant) manipulation skills
Human movement generation
posture/balance
involuntary... automatic
locomotion
navigating: moving through space
stepping
rhythmic (dance, music)
voluntary
whole body skills, sports
reach, grasp, manipulate
speech articulatory movement
object oriented
What is “motor control”?
... the neural processes underlying movement
generation of organisms…
“movement generation”
[neurally controlled movement…
not the tropisms and transport phenomena of biological
mobility in plants, amoebae, etc.
not falling from a tower ;=)
What is entailed in generating an
object-oriented movement?
scene and object perception
movement preparation
movement
preparation
movement initiation and
termination
movement timing and
coordination
control
degree of freedom problem
control
timing
What is entailed in generating an
object-oriented movement?
sequential organization of
movement: behavioral
organization
movement
preparation
goals, motivation, problem
solving
memory, spatial maps
adaptation
skill learning
control
timing
What is entailed in generating an
object-oriented movement?
=> spans perception, cognition
and control
movement
preparation
… difficult to isolate any
individual process, which is
why movement is hard to
study
critical to understand
integration
control
timing
Neural computation: learning
from analogies with the human
nervous system
(1) visual cognition
feature maps, within neural activation
controls the attentional foreground
and that are continuously linked to sensory
input …
sequences of attentional selection decisions
generate scene representations
purely spatial
y field, which
the with
number
ject
theof objects in the scene,
tional
to ensure that each
s then mechanism
passed
nce.
connected
to is driven by scene
triggers the
esentation in(2)
bodymovement
eference frame and object
planning
representation
scene
into
cessfully representation
reon
bel cue. The
Fig. 4. Label query. In this figure, two camera images in the top row indicate,
red
frame
to
extract
a
on
which
retinal
position
the system is currently fixated with its gaze. The
the saccade
image on the left shows the eye position after the scene is scanned, sitting
movement
plan:
ch is roughly on the
m of a given label
last scanned object. The right image shows the result
ov
slice
t. This states cue for the tube of sun screen. This cue pre-activates thee corresponding
em
x
extent and direction
query
en
epresentation (blue plane), producing a supra-threshold peak (red) int thetespace-label
ning of each
d in constant
g each object
nt int the bottom
field. The saccade movement is triggered by this selection, seen
row before and after the cue input.
epresentation of (initial)
of hand in space
re, two camera images in the top row indicate,
movement
direction
activation
… based on neural activation fields
1
0.5
0
movement
direction
ms]
it me [
3
2
1
6
mo 5
v
dire eme 4
ctio nt
n
250
precue
750
RS
response
signal
activation
complete
precue
PS
500
6
5
complete
precue
4
ms]
time [
3
mov2 1
6
e
dire ment
tion
precue
response
signal
n
Bastian, Riehle, Schöner, European Journal of Neuroscience,Vol. 18, 2047-2058 (2003)
… online updating
(3) timing and coordination
all movement is coordinated, timed, and
multi-sensory…
(4)does
muscles
his posture our elbow
not
have like a passive mechanical
action
tem withagonist-antagonist
a free joint at the
Q
(dampened)
ow: are tunable
¨
J =0
springs
ere J is inertial moment of
earm (if
upper
armper
is held
fixed)
one
lambda
muscle
force applied
tead, the
elbow
when
tested
onresists,
muscles
hed =>
there is at
active
detached
onecontrol=
end
bilization of the joint
co-constraction
controls stiffness
antagonist
L
agonist
L
=>experiment
Q
L
tol Feldman
figured out,
at the
roscopic
muscle activation
and force
cription
of result
this
from spatial
bilization
is
command
invariant
racteristic
force applied
L
Q
L
research
active elastic actuators
Fig. 5. Exploded 3D view of VSA-CUBE with basic components highlighted (a). Two configurations with different values of stiffness: minimum
stiffness (b) and maximum stiffness (c) with no external loads.
A. Mechanical design
[Antonio Bicchi]
The actuator external shape is a cube with a 55 mm edge,
(see Fig. 1,5). Small square grooves are machined along each
edge to permit the interconnection of the unit with other
components to form multi-DOF systems, as described in Sec.
(a)
(b)
why autonomous robots?
autonomous robots as demonstrators
of neural function
to show that neural process models are capable of
generating the modeled behavior based on real
sensory information …
such proof of function as a source of heuristics
discover problems that are often overlooked (e.g., coordinate
transforms)
discover non-problems that need not be solved to achieve a function
confront the problem of synthesis or integration…