Upload
minerva-stephens
View
36
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Probabilistic Robotics. SLAM. The SLAM Problem. A robot is exploring an unknown, static environment. Given: The robot’s controls (U 1:t ) Observations of nearby features (Z 1:t ) Estimate: Map of features (m) Pose / Path of the robot (x t ). Why is SLAM a hard problem?. - PowerPoint PPT Presentation
Citation preview
Probabilistic Robotics
SLAM
2
Given:
• The robot’s controls (U1:t)
• Observations of nearby features (Z1:t)
Estimate:
• Map of features (m)
• Pose / Path of the robot (xt)
The SLAM Problem
A robot is exploring an unknown, static environment.
3
Why is SLAM a hard problem?
• In the real world, the mapping between observations and landmarks is unknown
• Picking wrong data associations can have catastrophic consequences
Robot poseuncertainty
Nature of the SLAM Problem
4
Continuous Location of objects in component the map
Robot’s own pose
Discrete Correspondence component
Object is the same or not
5
SLAM: Simultaneous Localization and Mapping
•Full SLAM:
Estimates Entire pose (x1:t) and map (m)
Given Previous knowledge (Z1:t-1, U1:t-1) Current measurement (Zt, Ut)
),|,( t:1t:1t:1 uzmxp
Estimates entire path and map!
6
Graphical Model of Full SLAM:
),|,( :1:1:1 ttt uzmxp
7
SLAM: Simultaneous Localization and Mapping
•Online SLAM:
Estimates Most recent pose (xt) and map (m)
Given Previous knowledge (Z1:t-1, U1:t-1) Current measurement (Zt, Ut)
),|,( t:1t:1t uzmxp
Estimates most recent pose and map!
8
Graphical Model of Online SLAM:
121:1:1:1:1:1 ...),|,(),|,( ttttttt dxdxdxuzmxpuzmxp Integrations typically done one at a time
9
SLAM with Extended Kalman Filter
•Pre-requisites
• Maps are feature-based (landmarks)
small number (< 1000)
• Assumption - Gaussian Noise
• Process only positive sightings
No landmark = negative
Landmark = positive
10
EKF-SLAM with known correspondences
Correspondence Data association problem
Landmarks can’t be uniquely identified
Correspondence variable (Cit)
between feature (fit) and real landmark
Titit
it
it Srf
True identity of observed feature
11
EKF-SLAM with known correspondences
Signature Numerical value (average color)
Characterize type of landmark (integer)
Multidimensional vector
(height and color)
12
EKF-SLAM with known correspondences
Similar development to EKF localization
Diff robot pose + coordinates of all landmarks
Combined state vector
TNyNxNyxt
t SmmSmmyxm
xy ,,1,1,1 ...
(3N + 3)
),|( :1:1 ttt uzyp Online posterior
13
Motion update
Mean
Covariance
Iteration through measurements
Test for new landmarks
Initialization of elements
Expected measurement
Filter is updated
EKF-SLAM with known correspondences
Observing a landmark improves robot pose estimate
eliminates some uncertainty of other landmarks
Improves position estimates of the landmark + other landmarks
We don’t need to model past poses explicitly
14
EKF-SLAM with known correspondences
15
Example:
EKF-SLAM with known correspondences
16
Example:
• Uncertainty of landmarks are mainly due to robot’s pose uncertainty (persist over time)
Estimated location of landmarks are correlated
EKF-SLAM with unknown correspondences
• No correspondences for landmarks
• Uses an incremental maximum likelihood (ML) estimator
17
Determines most likely value of the correspondence variable
Takes this value for granted later on
EKF-SLAM with unknown correspondences
18
Motion update
Hypotheses of new landmark
Mean
Covariance
19
General Problem
Gaussian noise assumption Unrealistic
Spurious measurements Fake landmarks
Outliers
Affect robot’s localization
20
Solutions to General Problem
Provisional landmark list
New landmarks do not augment the map
Not considered to adjust robot’s pose
Consistent observation regular map
21
Solutions to General Problem
Landmark Existence Probability
Landmark is observed
Observable variable (o) increased by fixed value
Landmark is NOT observed when it should
Observable variable decreased
Removed from map when (o) drops below threshold
22
Problem with Maximum Likelihood (ML)
Once ML estimator determines likelihood of correspondence, it takes value for granted
always correct
Makes EKF susceptible to landmark confusion
Wrong results
23
Solutions to ML Problem
Spatial arrangement
Greater distance between landmarks
Less likely confusion will exist
Trade off:
few landmarks harder to localize
Little is known about optimal density of landmarks
24
Signatures
Give landmarks a very perceptual distinctiveness
(e,g, color, shape, …)
EKF-SLAM Limitations
• Selection of appropriate landmarks
• Reduces sensor reading utilization to presence or absence of those landmarks
Lots of sensor data is discarded
• Quadratic update time
Limits algorithm to scarce maps (< 1000 features)
• Low dimensionality of maps harder data association problem
25
EKF-SLAM Limitations
• Fundamental Dilemma of EKF-SLAM
It might work well with dense maps (millions of features)
It is brittle with scarce maps
26
BUT
It needs scarce maps because of complexity of the algorithm (update process)
27
SLAM Techniques
•EKF SLAM (chapter 10)
•Graph-SLAM (chapter 11)
•SEIF (chapter 12)
•Fast-SLAM (chapter 13)
28
Graph-SLAM
• Solves full SLAM problem
• Represents info as a graph of soft constraints
• Accumulates information into its graph without resolving it (lazy SLAM)
• Computationally cheap
• At the other end of EKF-SLAM
Process information right away (proactive
SLAM)
Computationally expensive
29
Graph-SLAM
• Calculates posteriors over robot path (not incremental)
• Has access to the full data
• Uses inference to create map using stored data
Offline algorithm
30
Sparse Extended Information Filter (SEIF)
• Implements a solution to online SLAM problem
• Calculates current pose and map (as EKF)
• Stores information representation of all knowledge (as Graph-SLAM)
Runs Online and is computationally efficient
• Applicable to multi-robot SLAM problem
31
FastSLAM Algorithm• Particle filter approach to the SLAM problem
• Maintain a set of particles
• Particles contain a sampled robot path and a map
• The features of the map are represented by own local Gaussian
• Map is created as a set of separate Gaussians
Map features are conditionally independent given the path
Factoring out the path (1 per particle)
Map feature become independent
Eliminates the need to maintain correlation among them
32
FastSLAM Algorithm
• Updating in FastSLAM
Sample new pose update the observed features
• Update can be performed online
• Solves both online and offline SLAM problem
• Instances Feature-based maps
Grid-based algorithm
33
• Local submaps [Leonard et al.99, Bosse et al. 02, Newman et al. 03]
• Sparse links (correlations) [Lu & Milios 97, Guivant & Nebot 01]
• Sparse extended information filters [Frese et al. 01, Thrun et al. 02]
• Thin junction tree filters [Paskin 03]
• Rao-Blackwellisation (FastSLAM) [Murphy 99, Montemerlo et al. 02, Eliazar et al. 03, Haehnel et al. 03]
Approximations for SLAM Problem
34
Thanks !!