►
From YouTube: Peer Neubert Interview with a Roboticist
Description
Thanks to Peer for being the first roboticist to be interviewed on HTM School. We talked about superpixels, hypervectors, long term robots, and how grid cells can be used for localization (ratslam). Here are links to some of the things we mentioned:
RatSLAM Video: https://www.youtube.com/watch?v=t2w6kYzTbr8
OpenRatSLAM Code: https://code.google.com/archive/p/ratslam/
Flamingo Superpixel image from this blog: http://popscan.blogspot.com/2014/12/superpixel-algorithm-implemented-in-java.html
A
Oh
so
I
have
been
doing
interviews
with
a
neuroscientist,
but
today
I
have
a
roboticist
in
the
new
mental
offices
with
us
and
because
I'm
an
opportunist
I
decided
to
ask
if
he
would
be
interviewed
because
he's
doing
some
really
interesting
thing
in
robotics,
so
I'll
just
go
ahead
and
introduce
you
now.
This
is
dr..
Peter
Norbert
from
the
chemnitz
Institute
of
Technology
in
Germany
right
I
get
that
right.
A
B
Maybe
that's
that's
part
of
the
story,
so
the
time
when
I
had
really
decided
or
got
into
robotics
terrorists
was
studying.
Computer
science
in
Chemnitz
and
I
visited
hands
on
the
course
of
Peter
Watson,
where
we
had
to
program
small
robots
to
solve
a
navigation
task
in
the
maze
and
straight
forward.
From
from
this
hands-on
course,
I
visited
a
course,
but
we
have
to
read
a
scientific
paper
and
disgustedly
is
a
scientific
paper,
and
that's
really
amazing
to
to
see
how
close
science
is
to
this
practical,
actually
moving
thing
in
this
little
maze,
I
really.
A
A
C
A
C
C
C
B
C
B
An
object,
so
it's
all
about
avoiding
a
pre-major
hard
decision
for
very
long
time
to
detect
some
object
in
an
image
to
see.
Okay,
it's
a
cat.
The
task
was
to
for
the
general
pipeline
was
to
first
segment
out
the
foreground,
object
from
the
background
and
then
decide
what
this
foreground
object
is,
and
this
is
very
often
a
very
hard
problem.
They
a
very
cluttered
desk
with
a
lot
of
stuff
on
it
and
I
asked
you
okay.
What
is
the
object
in
this
image?
Yeah.
B
I'm
talking
about
the
whole
desk
or
the
keyboard,
or
the
mark
or
whatever,
and
at
this
point
in
time
it's
hard
to
you
to
segment
out
this
thing
in
the
foreground
object
in
particular,
if
you
think
of
some
occluded
parts
of
the
boundary
of
some
object,
it's
really
hard
to
get
them
from
from
pixels
bottom.
If.
A
A
B
C
C
B
B
A
B
A
This
leads
me
to
another
question
because
we're
talking
about
computer
image,
computer
vision
right
and
how
camera
footage
or
what,
however,
we're
recording
you
know
light.
Essentially,
how
do
we
get
that
encoded
into
a
system
that
a
robot
can
understand
and
make
sense
of,
and
this
is
just
one
way
or
one
tactic
right,
yeah
I
mean
in
the
end
it's
turned
into
bits
and
bytes
that
the
robot
has
to
understand.
What
is
that
encoding
generally
in
robotics
for
sensors
to
give
to
the
robot
I.
B
Think
if
it's
possible
to
talk
about
a
general
encoding
of
sensor,
information
in
robotics,
then
I
think
it's
in
terms
of
coordinate
frames
to
say,
I
have
something
like
robot,
centroid,
coordinate
frame
and
all
my
sensor.
Information
is
relative
to
this
coordinate
frame
so
happens
to
be
3d
points
or
some
other
three-dimensional
structures
in
the
world
or
if
the
robot
is
just
moving
on
the
ground
plane
on.
C
B
B
A
A
C
B
B
Localization
of
mapping
I'll
put
it
right
there
and
imagine:
a
robot
enters
a
new
unknown
place
in
the
world,
and
it
want
to
create
a
map
of
this
environment
and
simultaneously
know
where
it
is
in
the
map.
So
if
you
have
a
good
estimate
of
your
your
position,
it's
easy
to
take
all
your
sensor,
measurements
and.
B
A
map,
and
if
you
have
a
map,
it's
very
it's
more
or
less
easy
to
to
know
where
you
are
in
the
map
based
on
your
sensor
information.
But
if
you
have
neither
of
them,
it's
really
a
challenging
task.
Yeah
and
it
has
been
a
very
hard
problem
and
well
research
problem
in
robotics
for,
like
two
decades
yeah.
C
A
One
of
the
most
interesting
methods
of
solving
that
problem
is
in
this
rat
slam
paper,
which
I'll
link
in
the
description
and
I'll
bring
some
graphics
off,
and
that's
what
I
wanted
to
talk
about
because
they
use
these
grid
cell
modules
are
sort
of
the
idea
of
how
grid
cells
work
in
the
internal
cortex
to
help
a
robot
navigate
through
the
world,
which
is
really
interesting.
Experience.
B
A
B
C
A
B
C
B
A
I
think
about
you
know
raw
robots
that
we
shoot
to
other
planets
to
to
explore
and
occasionally
come
back
and
report
on
what
they
found.
You
know
we
can't
really
do
that
very
well.
Today,
I
reminded
of
some
discussion
we
had
about
the
curiosity
on
Mars.
How
does
that
work
today?
How
much
automation
goes
on
in
that
long-term
robot.
B
B
C
B
A
B
A
B
A
Do
when
you
think
about
it,
but
even
construction
robots,
something
that's
just
continually
moving
supplies
where
they
need
to
be
removing
debris.
You
know
collecting
garbage
trash
whatever
you
know
something
you
can
put
out
there
and
say:
do
your
job
and
come
back
to
me
in
a
few
years,
I'll
be
great,
yeah,
okay!
Well,
we
have
one
more
topic
that
they
can
go
over,
and
this
is
the
this
is
the
hardest
topic
for
me
as
hyper
vectors.
We're
really
talking
about
is
high,
dimensional
vectors
right.