presenter: michael bowling. 2 helping the world understand data and make informed decisions...
Post on 23-Jan-2016
212 views
TRANSCRIPT
RoboticsRobotics
Presenter: Michael BowlingPresenter: Michael Bowling
2
Vision StatementVision Statement
Helping the world understand data and make informed
decisions
Potential beneficiaries:• Growing robotics and UVS sector,• Diverse industries (incl. mining, farming, service), • Society as a whole.
robots
3
MotivationMotivationDeveloping industry with high potential impact on nearly all aspects of society“Dull, Dirty, or Dangerous”
Robots are great testbeds for AI researchRobots are great for outreach
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
(Photo from AICML School Visit, 2007)
ML has a key role:Current robot systems are brittle, highly engineered
Our world is diverse, unpredictable, unstructuredChallenging problems for ML:
Complete systemData is noisy, limited, costly to gatherSafety of people, surroundings, robot itselfReal-time demands
4
Newest ThrustNewest Thrust
Concerted effort began in 2004/2005
Robotics research requires a teamDiverse talents necessarySizable software systemConsiderable engineering effort
AICML gives a distinct advantage2 full-time software developers1 robotics engineer (recent hire)Still looking for a robotics/ML PDF
5
Projects and StatusProjects and Status
1. Gait Learning (completed; poster #15)
2. Automatic Calibration (ongoing; poster #24)
3. Hybrid Car Optimization (ongoing)
4. Outdoor Navigation, (ongoing; poster #28)
5. Shodan (ongoing; poster #27)
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
6
AICML Personnel (cumulative)AICML Personnel (cumulative)
Primary PI’s: Bowling, Schuurmans, Sutton, Szepesvari
8 Software developers 1 Robotic engineer6 Grad students
7
ResourcesResources
Grants$300K AIF New Faculty GrantPortion of Rich Sutton’s iCORE chair
Robot Platforms16 Sony Aibo ERS-7s 2 Segway RMPs1 ActivMedia Pioneer 3-DX
Shodan Robotics SimulatorDeveloped and used in-houseOngoing beta-testing for release
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
8
Partners/CollaboratorsPartners/Collaborators
Early Partners:Toyota Motor CorporationUofA Prof in Computer Vision
Future Partners:CCUVS: National initiative located in Alberta
Continued discussions with a number of robotics companies
9
Early Highlights Early Highlights
Successful geocaching demonstration (Daily Planet segment)
Most efficient gait learning algorithm (Smithsonian demo)
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
10
Projects and StatusProjects and Status
1. Gait Learning (completed; poster #15)
2. Automatic Calibration (ongoing; poster #24)
3. Hybrid Car Optimization (ongoing)
4. Outdoor Navigation, (ongoing; poster #28)
5. Shodan (ongoing; poster #27)
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
Gait LearningGait Learning
Technical DetailsTechnical Details
12
The ChallengesThe ChallengesBalance and locomotion is the key problem for legged robots
Gait optimization is hard:Open loop gait may require >50 parameters
Effect of parameters involve complex interactions
Effective gaits depend upon:Walking surfaceIndividual robot characteristicsBattery level
Ideal for machine learning!
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
13
Machine Learning ChallengesMachine Learning Challenges
Training takes time; causes wear Use data
efficiently
Data is noisy Explicitly
reason about uncertainty
(Photos from CS Summer Camp, 2006)
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
14
ApproachApproach
Gaussian Process OptimizationPrior over functionsCompute posterior given observationsUse to pick next walk parameters
15
ResultsResults
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
Gaussian ProcessOptimization
Previous Best
Number of Walks Tested
Quality of Walk
16
ConclusionsConclusionsMost efficient published gait learning algorithm (IJCAI, 2007)Optimized walk in 2 hours instead of 10!Little expert knowledge requiredNo starting seed or restarts needed
Also successfully applied to finding parameters in stereo vision and NLP
Exhibited in “Alberta at the Smithsonian” in Washington, D.C., Summer 2006.
17
The FutureThe Future
Exploit the built hardware/software team
New projectsOutdoor navigation with pedestriansMobile robot manipulation
RMP + WAM armRobot minigolf (fall ‘07 grad course)
Open-source robot platform (poster #26)Pursue industrial partnerships
ToyotaCCUVS
18
QuestionsQuestions??
1. Gait Learning (completed; poster #15)
2. Automatic Calibration (ongoing; poster #24)
3. Hybrid Car Optimization (ongoing)
4. Outdoor Navigation, (ongoing; poster #28)
5. Shodan (ongoing; poster #27)
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTime™ and aTIFF (Uncompressed) decompressor
are needed to see this picture.