-
Problems with planning
-
An alternative view: behavioral robotics
-
Emergence
-
Back to Brooks and his behavioral approach
-
How behavior-based robotics works
-
Examples of robots with behavior-based architechtures
-
Making Genghis walk
-
More complex behaviors
-
Making Genghis walk with a classical planning architecture
Problems with planning
- Planning algorithms can be slow.
- Planning requires that information about the world be explicitly encoded.
- Plans fail; the world is dynamic and imperfect.
NB: We say that robot architectures that rely on planning follow the
classical view. i.e.,they are based on representation and
manipulation of explicit knowledge.
An alternative view: behavioral robotics
Questioning of the classical view began in the mid 1980s. We'll look
at just one approach that was proposed: Behavioral robotics
(Rodney Brooks).
Inspiration for behavioral robotics includes the following:
- Situatedness - the robot is situated in the real world. It
does not operate on an abstract representation of reality, but on
reality itself.
- Embodiment - a robot has a physical presence. Its dynamic
interactions with the world cannot be simulated faithfully.
- Emergence - intelligence arises from the interactions of
the robotic agent with its environment.
The theory behind the work of Rodney Brooks et al:
The world is its own best model. It is always exactly up to date.
It always contains every detail there is to be known. The trick is to
sense it appropriately and often enough.
Emergence
Notion that complex phenomena arise from the (serendipitously) coordinated
activity of simpler phenomena. Some examples from nature include:
- Moths' bat evasion tactics
- Termite colonies
Philosopher John Searle has argued that consciousness is an emergent
property of the brain.
We all know (I hope) that bats use echolocation to detect prey. The prey
have techniques of their own, however to elude the bats. [The following is
taken from Barry Sinervo, 1997.]
Moths receive ultrasonic bat vocalizations with two ears on each side
of the thorax. When pressure waves from the high energy bat vocalizations
strike the ears and vibrate the membranes of the moth ears, two sensory
receptors (A1 and A2) can fire depending on the energy of the sound.
The sensory neurons trigger an action potential in the sensory interneurons
which conduct the electrical impulses to the next synapse. The next neuron
in the chain after the first synapse is triggered by neurotransmitters which
are released and cross the synaptic junction and trigger a new action potential.
The impulse can travel to the brain in this manner, or to ganglia in the thorax.
Neurons in the ganglia or brain can integrate the information and send an
action potential on to motor neurons that cause muscles to fire. The
differential senstivity of the A1 and A2 sensory neurons leads to a stimulus
filtering of the bat sounds that gives the moth two options:
-
long distance evasion tactics when the bat is far away
-
short distance evasion tactics when the bat is at extremely close range.
The A1 cell is sensitive to low energy sounds (e.g., distant bat calls), and
the A2 is sensitive to high energy sounds (e.g., close bat calls).
We will focus on A2:
If the moth's long- to medium-distance evasive maneuvers fail, and the bat
is about to collide, the A2 neuron begins firing because of the high
energy reaching the moth ear. The A2 cells send a message to the thoracic
ganglia, and this seems to shut down wing beats or cause them to fire
erratically. This leads to erratic flight which may be a last ditch attempt
to elude the ranging and speed computing neurons of the bat's brain.
Back to Brooks and his behavioral approach
Brooks argues that the number of years that evolution has spent on
"higher level" function such as language, planning, intellectual enterprise,
etc., is relatively tiny. The harder problems (behaviors) are the simple
ones.
How behavior-based robotics works
- decompose design not into functional components such as
perception and planning, but into behaviors such as obstacle avoidance,
wall-following, and exploration
- each behavioral module accesses sensor inputs independently to extract
just the information it needs
- each behavioral module sends its own signals to the effectors
- behaviors are arranged into a hierarchy
- higher level behaviors can override lower level behaviors
Roboticists have been quite successful in demonstrating that many basic
competences in the physical world can be achieved using simple,
inexpensive mechanisms.
Examples of robots with behavior-based architectures
Making Genghis walk
Making Genghis walk on flat terrain involves only the following simple
behaviors and a bit of coordination:
- Each leg moves in cycles:
- on the ground
- off the ground
- Coordinate so that 3 of the 6 legs (not all on the same side) are
on the ground at any one time.
Walking on rugged terrain is obviously more complicated -- but not
much more. Just a simple adjustment needs to be made. When a leg's
forward motion is blocked, retract it, lift it higher, and try again:
The diagram above, which specifies simple action sequences, is called
an AFSM (Augmented Finite State Machine). Timers
control the amount of time it takes to traverse each arc in the AFSM.
Note that this is model-free. That is, it requires no model of the
environment. It does not deliberate or use search for
generating the controls.
More complex behaviors
The Subsumption Architecture offers the ability for synchronizing
AFSMs. This enables the programmer to compose increasingly complex
controllers.
Making Genghis walk with a classical planning
architecture
Say we wanted to make Genghis walk by implementing a classical planning
architecture that makes use of a configuration space representation:
- To do this, we woould need an accurate model of the terrain.
- The configuration space would be 18-dimensional! (12 dimensions
for the legs; 6 for the location and orientation of the robot relative
to its environment.)
- Even if a plan could be devised in a reasonable amount of time,
there may be no graceful way to recover from plan failures.