Practical EnhancementsDistance-Based UpdatesLandmark HistoriesExtended Motion ModelEmpirical ResultsPhysical Robot ExperimentsSimulation ExperimentsSummaryPractical EnhancementsEmpirical ResultsSummaryPractical Vision-Based Monte CarloLocalization on a Legged RobotMohan Sridharan Gregory Kuhlmann Peter StoneLearning Agents Research GroupDepartment of Computer SciencesThe University of Texas at AustinIEEE International Conference on Robotics and Automation,2005M. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryThe ProblemMobile Robot LocalizationMaintain estimate of global position and orientation over timeGiven map of fixed landmark locationsNot SLAMM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryThe ProblemMobile Robot LocalizationMaintain estimate of global position and orientation over timeGiven map of fixed landmark locationsNot SLAMM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryChallenging PlatformTypical PlatformWheeled robotRange-finding sensorsSony Aibo ERS-7Color CMOS Camera in noseNarrow field-of-view (56o)30 YCrCb frames per secondQuadruped576MHz processorAll on-board processingM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryChallenging PlatformOur PlatformLegged robotVision-based sensorsSony Aibo ERS-7Color CMOS Camera in noseNarrow field-of-view (56o)30 YCrCb frames per secondQuadruped576MHz processorAll on-board processingM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryGoalDesiderataNavigate to specific point quicklyRemain localized while collidingRecover quickly from kidnappingsApproachBegin with baseline MCL algorithmAdd set of practical enhancementsLarge improvement over baselineM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryGoalDesiderataNavigate to specific point quicklyRemain localized while collidingRecover quickly from kidnappingsApproachBegin with baseline MCL algorithmAdd set of practical enhancementsLarge improvement over baselineM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryGoalDesiderataNavigate to specific point quicklyRemain localized while collidingRecover quickly from kidnappingsApproachBegin with baseline MCL algorithmAdd set of practical enhancementsLarge improvement over baselineM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryMethod: Particle FilteringEstimate p(hT|oT, aT−1, oT−1, aT−2, . . . , a0):Distribution of poses given observations and actionsRepresented by finite set of samples: particlesEach is a hypothesis: hhx, y, θi , piAverage to get single estimate of pose and confidenceM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryMethod: Particle FilteringEstimate p(hT|oT, aT−1, oT−1, aT−2, . . . , a0):Distribution of poses given observations and actionsRepresented by finite set of samples: particlesEach is a hypothesis: hhx, y, θi , piAverage to get single estimate of pose and confidenceM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryMethod: Particle FilteringEstimate p(hT|oT, aT−1, oT−1, aT−2, . . . , a0):Distribution of poses given observations and actionsRepresented by finite set of samples: particlesEach is a hypothesis: hhx, y, θi , piAverage to get single estimate of pose and confidenceM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryOutline1Practical EnhancementsDistance-Based UpdatesLandmark HistoriesExtended Motion Model2Empirical ResultsPhysical Robot ExperimentsSimulation ExperimentsM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryDistance-Based UpdatesLandmark HistoriesExtended Motion ModelOutline1Practical EnhancementsDistance-Based UpdatesLandmark HistoriesExtended Motion Model2Empirical ResultsPhysical Robot ExperimentsSimulation ExperimentsM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryDistance-Based UpdatesLandmark HistoriesExtended Motion ModelBaseline: Observation UpdateNeed sensor model: p(o|h)Predicts observations given pose hypothesis using mapUpdate each particle when robot sees somethingCompute similarity for each observed landmark in frameUse angles only [Rofer and Jungel, 2003]Measured and expected angle differenceCompute product of similaritiesAdjust probability closer to new valueM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryDistance-Based UpdatesLandmark HistoriesExtended Motion ModelBaseline: Observation UpdateNeed sensor model: p(o|h)Predicts observations given pose hypothesis using mapUpdate each particle when robot sees somethingCompute similarity for each observed landmark in frameUse angles only [Rofer and Jungel, 2003]Measured and expected angle differenceCompute product of similaritiesAdjust probability closer to new valueM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged RobotPractical EnhancementsEmpirical ResultsSummaryDistance-Based UpdatesLandmark HistoriesExtended Motion ModelBaseline: Observation UpdateNeed sensor model: p(o|h)Predicts observations given pose hypothesis using mapUpdate each particle when robot sees somethingCompute similarity for each observed landmark in frameUse angles only [Rofer and Jungel, 2003]Measured and expected angle differenceCompute product of similaritiesAdjust probability closer to new valueM. Sridharan, G. Kuhlmann, and P. Stone – UT Austin Practical Vision-Based MCL on a Legged
View Full Document