Docstoc

Applications

Document Sample
Applications Powered By Docstoc
					Adaptive Robotics
COM2110

Autumn Semester 2008
Lecturer: Amanda Sharkey


                           1
      “Robots in the news”
Macho robot helps explain lizards' odd
behaviour
22:00 24 November 2008 by David Robson New
Scientist
Why does male anolis lizard perform a series of push-
ups before attempting to intimidate rivals with
colourful displays?
Terry Ord at Harvard University built a robotic lizard
that can inflate a colourful dewlap under its chin, bob
its head, and perform push-ups

                                                      2
Cameras in Puerto Rico forests recorded
when lizards turned their heads
Robot lizard‟s push-ups attracted
attention – without them lizards often
missed much of coloured dewlap
display.

                                      3
Lect 1: what is a robot? Early robots, Shakey and GOFAI, Behaviour-
based robotics
Mechanisms and robot control (and biological inspiration)
Lect 2: Grey Walter, Brooks and Subsumption Architecture.
Lect 3: Adaptation and learning
Lect 4: Artificial Neural Nets and Learning (biologically inspired, could
be used to implement robot control in behaviour-based robotics
approach, also move away from GOFAI)
Lect 5: Evolutionary Robotics (also biologically inspired – another way
of developing robot control)
Lect 6: Swarm Robotics (reasons for, biological inspiration, local control
and communication, self-organisation and emergence).
Lect 7: Biorobotics and Biological modelling.
Lect 8: Applications




                                                                         4
    “new wave” robotics
Aka
Behaviour-based robotics
Nouvelle AI
Embodied Cognition




                           5
        In opposition to:
Sense-Model-Plan-Action approach
Centralised cognition
Mind as central logic engine
Memory as retrieval from stored symbolic
database
Problem solving as logical inference
Environment – problem domain
Body – input device
Functionalism


                                           6
           Functionalism
Thinking, and other intelligent functions –
could be carried out on different hardware. -
     interested in the software.
Swiss cheese?
Physical Symbol System Hypothesis
(Newell and Simon, 1975)
Emphasis on manipulation and processing of
symbolic representations of the world
Little interest in how those representations
are related to the objects in the world.

                                                7
Review of defining characteristics of “new wave” robotics and
                        “nouvelle AI”
Simplicity – “Keep it simple”
Minimal representation
Biological inspiration
Embodiment
Situatedness
Emergence
Autonomy
Interaction with the environment
   Brain, body and world


                                                                8
Two guiding principles (from Maes,
1994)
  Looking at complete systems often
  changes the problems in a favourable way
  Interaction dynamics can lead to emergent
  complexity




                                          9
       Complete systems
Building complete systems can simplify
problem
  E.g. with sensors, easier to disambiguate
  natural language utterances because they
  are related to the objects the agent sees
  E.g. systems with sensors and actuators
  can perform tests in the environment and
  needs less modelling and inference


                                              10
       Complete systems
Since intelligent system is situated in
environment, this can be exploited
   E.g. using the environment as external memory,
   reminding which tasks remain to be done. Also
   habitat constraints, e.g. usual size of doors in
   office, can be exploited
Time: incremental solution can be arrived at
   e.g NLP and asking further questions



                                                  11
       Complete systems
Society: can look at other agents and
other solutions.
  E.g mobile robot closely following a person
  walking by, to avoid bumping into things.




                                            12
    Emergent complexity
Idea from ethology that animal‟s
behaviour can only be understood in
the context of the environment in which
it occurs.
Simon(1969) the complexity of an ant‟s
behaviour reflects the complexity of its
environment


                                       13
Emergent functionality
  E.g. Mataric‟s (1991) wall following robot.
  One module steers robot towards wall
  when distance above threshold, and one
  module steers away when distance below
  threshold – result = wall following
  Social insects following simple local rules to
  produce emergent complexity


                                              14
     Today - applications
Navlab – autonomous vehicles
  Alternative to subsumption architecture
DARPA grand challenge
Swarm robotics – pherobots
Robot sheep dog
 Leurre project: influencing the
behaviour of cockroaches

                                            15
          Application areas
Physical robots
    Household e.g. vacuuming, lawn mowing
    Autonomous vehicles e.g. Navlab
    Agricultural
    Hostile terrains e.g. underwater, space, military, bridge
    inspections, disaster (9/11)
       Urban search and Rescue (Robin Murphy)
    Entertainment e.g. toys, personal robots, exhibitions,
    games
    Companions – for the young, for the elderly
    Military
    Policing and surveillance
       Some ethical issues in involved in the above

                                                                16
               Navlab
CMU (Carnegie Mellon University) group
Robot cars, trucks, buses for autonomous
navigation
11 different Navlabs (Navlab11 on its way)
Langer, Rosenblatt and Herbert (1994) A
Behaviour-based System for Off-Road
Navigation. IEEE Trans Robotics and
Automation, 10, 6, pp 776-782


                                             17
18
Navlab 11




            19
Autonomous cross-country navigation
Rugged terrain
Processing of 1000s of images
Need to avoid failure
Simple algorithms for obstacle detection and
local map building in behaviour-based
architecture
Underlying principle: keep things simple

                                           20
Use of independent modules
  Perception module (list of untraversable
  regions)
  Local map module (maintains map of
  terrain round vehicle)
  Planning module (generates steering arcs,
  keeping clear of untraversable regions)



                                              21
22
            Perception
Takes single image as input and
produces list of untraversable regions
Terrain classification algorithm: based
on grid system
Each cell in grid corresponds to 20cm x
20cm
Each cell classified as traversable or not

                                         23
24
     Terrain classification
Strengths: simple system, each image
processed individually without terrain
matching and merging
Limitations – some misclassification – dense
vegetation can appear as an obstacle
Problems with regions with poor reflectance
e.g. water
Dependence on good sensors


                                               25
  Local Map Management
Purpose – to maintain list of untraversable
cells in region round vechicle
Module called Ganesha
Uses 2D grid-based representation of local
map
Core of system is single loop
   Read current position of vehicle, update coordinates of
   cells. Discard cells outside bounds of active regions
   Get obstacle cells and place in local map
   Update internal cell attributes
   Send list of obstacle cells to planning system

                                                             26
27
               Planning
Use map to generate commands to steer
round obstacles
Used DAMN (Distributed Architecture for
Mobile Navigation) behaviour-based
architecture
   Like subsumption architecture
   Uses specialised task-achieving modules that
   operate independently and are responsible for
   only part of vehicle control
   Some internal representation of world
   Activation selection – relies on command fusion
                                                 28
29
30
             DAMN cont.
Each behaviour votes for or against set of
vehicle actions
   Votes between –1 and +1 for each of 15
   steering commands
   Weighted sum of votes computed. Steering arc
   with maximum vote is found
   Speed also decided by voting
   Obstacle avoidance – each behaviour has list of
   current obstacles
     Votes for trajectories free of obstacles
     Votes against paths with obstacles
     Other behaviours: goals seeking, drive straight,
     maintain turn.                                     31
32
2 behaviours
  Obstacle avoidance and goal seeking
  Arbiter combines votes and issues new
  driving command every 100 ms
  Weights 0.8 for obstacle avoidance and 0.2
  for goal seeking




                                           33
            Limitations
Can‟t deal with some situations e.g dead
ends such as closed corridor with depth
greater than field of view of sensor
Limited range and speed of sensor
Non-real-time nature
Poor performance of perception on certain
types of environment.
BUT 1km traverse shows robustness

                                            34
DARPA Grand Challenge Robot Vehicle
            race 2005

Unmanned vehicles on 132 mile course in
Mojave desert
- of 23 entrants, five completed.
(previous year none got further than 11 km)
Winner: Stanley, Stanford University
6 hours, 53 minutes
(2,000,000 dollars)
2nd place: Sandstorm, CMU
See www.grandchallenge.org
                                              35
36
Route given 2 hours before competition,
in form of GPS coordinates
Teams could program routes into
vehicles
Sebastian Thrun – Stanley had vision-
based speed switch: drove faster when
it detected straight road ahead without
obstacles

                                     37
             Stanley

Used:
 GPS
 Laser Range Finder to map the road 30
 meters ahead
 Video camera to scan 80 meters ahead
 Odometry




                                         38
DARPA Grand Challenge 2007
    “Urban challenge”
“Autonomous ground vehicles executing
simulated military supply missions
safely and effectively in a mock urban
area” (DARPA press release)

Challenge completed November 2007



                                     39
  60 mile urban area course, to be completed
  in less than 6 hours.
  Rules – following all traffic regulations,
  negotiating with other traffic and obstacles.
  E.g. maintaining precedence at 4 way stop
  intersection
  11 teams given development funding.
This challenge less physically demanding, but
  involved encounters with other vehicles.

                                                  40
$2000,000 Winner: Tartan Racing
(Carnegie Mellon University)
Averaged 14 miles per hour throughout
course.
2nd: Stanford Racing (Stanford
University).



                                    41
42
    Next Grand Challenge
Where do you think that robots could
most usefully be employed?
(I.e Where should funding be put?)
What kind of robots do you think are
likely to be developed?




                                       43
     Swarm Robotics and
        applications
Swarm robotic principles
  Biological inspiration from social insects
  Simple autonomous agents
    Decentralised local control
    Minimal communication and representation
    Reactive behaviour
  Interaction with environment
  Emergence, situatedness, embodiment

                                               44
Advantages for applications
Cheap, expendable autonomous robots
Able to negotiate and exploit environment to
achieve emergent cooperative solutions to
practical problems
Redundancy and simplicity means robots can
be added, or removed, without requiring
recalibration or mission failure



                                           45
 Possible application areas

Areas that are hostile or inaccessible to
humans
e.g. clearing up toxic waste or
contaminated buildings
e.g. mine fields
e.g. planetary exploration
e.g. burning or collapsed buildings
e.g. battlefield search for survivors
                                        46
       Pheromone robotics
David Payton et al (HRL labs) (2004)
Uses „virtual pheromones‟
Imagined scenario: rescue team enters unfamiliar
building and needs to find survivors
Swarm of robots explores – one finds survivor and
emits message.
Message relayed locally among neighbouring robots
Virtual pheromone gradient propagated back to
rescue workers.



                                                    47
                   Pherobot




PalmV PDA used as main control computer   48
Virtual pheromones implemented via infrared
signals
8 radially oriented directional infrared
receivers and transmitters on each robot
Robots can transmit and receive messages
directionally relative to current orientation
Pheromone message also contains hop-count
field which can be decremented as it is
passed on – creating a pheromone gradient


                                            49
50
Augmented reality system: video
camera on users head receives
signals from robots and displays
them as arrows
“world embedded computation”
No distinct step of map
generation, but robots act as
distributed set of processors
embedded in environment.

                                51
Robot-animal interactions
 Robot Sheep Dog project
 Vaughan, R., Sumpter, N., Henderson, J., Frost,
  A., and Cameron, S. (1999) Experiments in
  Automatic Flock Control. Robotics and
  Autonomous Systems, 31, 109-117.




                                               52
Minimal generalised model of underlying
flock behaviour
Experimental system: robot, workstation and
video camera
Robot has top speed twice as fast as ducks
Position of robot and ducks determined by
processing video image
Blob detection used for flock

                                          53
Flock-control algorithm – takes in vision data
(positions of robot, flock and goal) and
returns desired vehicle trajectory.
Mimimal simulation of duck herding in
circular arena
Potential field algorithm used to generate
duck movement.
Ducks are
   attracted to each other
   Repelled from each other if too close
   Repelled from arena wall
   Repelled from robot


                                           54
- As an aside: Flocking




                          55
              Artificial Life examples
Craig Reynolds (1987) work on flocking
  behaviour
   Boids – virtual birds with basic flight capability
   3 rules
      (i) collision avoidance – avoid collisions with nearby flock-
      mates
      (ii) velocity matching – attempt to match velocity with
      nearby flock-mates.
      (iii) flock centering – attempt to stay close to nearby flock-
      mates
Each boid is a basic unit that “sees” only its nearby flock-mates
   and “flies” according to the 3 rules.

                                                                       56
Result: boids flocked and flew as a cohesive
group. When obstacles appeared in their way
they spontaneously split into 2 subgroups,
without central guidance, and rejoined after
clearing obstruction.
Boids algorithm: used to produce
photorealistic imagery of bat swarms in
“Batman returns”.


                                          57
Boids – illustrate basic principles of Alife
systems
  Large number of simple elemental units
  Units interacting with nearby neighbours
  with no central controller
  High-level emergent phenomena from low
  level interactions


                                             58
Returning to sheep dog…..
Robot is
  Attracted to flock with magnitude
  proportional to distance from goal
  Repelled from goal with constant
  magnitude




                                       59
60
Performance in real world




                            61
Leurre project




                 62
63
         Leurre project
Insbot – mixed society of cockroaches
and robots
Aim – for Insbot to aggregate with
cockroaches in shelter. Eventual aim to
influence behaviour of cockroaches.




                                      64
Chemical sensing:
  Robot covered with medium impregnated
  with cockroach pheromones
  Behaviour:
    Controlled by fused combination of basic
    behaviours
    Aggregation –more likely to stop if several
    cockroaches detected.



                                                  65
Detection of
  shelters (ambient light),
  walls (IR sensors),
  robots (IR sensors and local
  communication)
  and cockroaches (IR sensors and linear
  camera)


                                           66
67
68
Summary and conclusions
Review of main characteristics of approach
Applications
   Unmanned vehicle (and DARPA grand
   challenge)
   Swarm robotics – Pheromone robots
   Robot sheep dog
   Insbot and cockroaches
   Illustration of practical promise –
     Keep it simple approach – minimal representation
     Embodiment, interaction with environment
     Emergence, autonomy

                                                        69

				
DOCUMENT INFO