tddc17 percetion2 2010 presentation

Mariusz WzorekArtificial Intelligence & Integrated Computer Systems DivisionDepartment of Computer and Information ScienceLinköping University, Sweden

Robotics/Perception II

Mariusz WzorekArtificial Intelligence and Integrated Computer Systems Division (AIICS)


2

Outline

• Sensors - summary• Computer systems• Robotic architectures• Mapping and Localization• Motion planning• Motion control


3

Some of the key issues in robotics• Motion

– object manipulation– locomotion: holonomic ↔ non-holonomic

• Navigation– localization (where am I, where are objects relative to me?)– planning (how do I get there?)

• Mapping– building an environmental map from sensor inputs– for navigation purpose– for exploration (where haven’t I been)

Related areas: perception, control, situation awareness, monitoring, planning, reasoning, communication, software architectures, power systems

Video


4

Anatomy of Robots

Robots consist of:• Motion mechanism

– Wheels, belts, legs,propellers, rotors

• Manipulators – Arms, grippers

• Sensors• Computer systems

– Microcontrollers, embedded systems• Body/Frame


5

Sensors

Summary of most commonly used sensor for mobile robots.

Any weather

Any light

Detection in 15m

Fast response Weight Affordable

CCD Camera/stereo/Omnidirectional/o. flow C C C C

Ultrasonic C C C C

Scanning laser C C C C

3D Scanning laser C C C*Millimeter Wave Radar C C C C

* scanning laser on a tilting unit


Sensors cont’d - common interfaces

• analog - voltage level• digital:

– serial connections e.g.:• RS232• I2C• SPI• USB

– pulse width modulation - PWM–

6


Computer systems

Many challenges and trade-offs!– power consumption– size & weight– computational power– robustness– different operational conditions - moisture,

temperature, dirt, vibrations, etc.

7


Computer systems cont’d

PC104 - standardized form factor• industrial grade,• relatively small size,• highly configurable,• >200 vendors world-wide,• variety of components• variety of processors

8


Computer systems cont’d

Custom made embedded systems• with integrated sensor suite• micro scale, light weight• fitted for platform design

9

• application specific


Computer systems cont’dExample system design:

10

!"#$• !"#$!%&'!()*+,-./!

• !"!%0!12/!

• !"!%0!3456!789:)!

%&#$• !;<<!/&'!()*+,-!===!

• !>?@!/0!12/!

• !?">!/0!3456!789:)!

'%#$• !;<<!/&'!()*+,-!===!

• !>?@!/0!12/!

• !?">!/0!3456!789:)!

AB6)8*)B!

CD9BE6!

FFG!FHIH8!

F4-)84!

J6)8-4I!

F4-)84!

(4*!

J9IB!

K*9B!

L4-464!1/2M!NL2CO!L2FCP!

048H-)B89E!(8)55,8)!2I+-)B)8!

%(C!1)E)9:)8!

Q45)8!14*R)!S9*7)8!

/H7,I)!

AB6)8*)B!T97)H!C)8:)8!

U<>#""V!W98)I)55!0897R)!

%C/!/H7)-5!

()8E)X+H*!C)*5H8!C,9B)!

FH--,*9E4+H*!/H7,I)!

(SF!C)*5H8!C,9B)!

CB)8)H.T959H*!

/H7,I)!

(HD)8!/4*4R)-)*B!K*9B!

AB6)8*)B!

1C>Y>!

2*4IHR!

S98)D98)!

/9*9GT!

1)EH87)85!


11

Outline



12

Telesystems/Telemanipulators

Video

Delay


13

TelepresenceExperience of being fully present at live eventwithout actually being there• see the environment through robot’s camera• feel the surrounding through robot sensors

Realized by for example: • stereo vision,• sound feedback,• cameras that follow the operator’s head movements,• force feedback and tactile sensing


14

Reactive systems

Robot responses directly to sensor stimuli.Intelligence emerges from combination of simple behaviors.Subsumption architecture (Rodney Brooks, MIT):• concurrent behaviors• higher levels subsume behavior of lower layers• no internal representation• finite state machines

SENSE ACT

Video


15

Limitations of Reactive Systems• Agents without environment models must have sufficient

information available from local environment• If decisions are based on local environment, how does it

take into account non-local information (i.e., it has a “short-term” view)

• Difficult to make reactive agents that learn• Since behavior emerges from component interactions plus

environment, it is hard to see how to engineer specific agents (no principled methodology exists)

• It is hard to engineer agents with large numbers of behaviors (dynamics of interactions become too complex to understand)


16

Hierarchical Systems

SENSE PLAN ACT

extractfeatures

combine features into model

plantask

executetask

controlmotors

Sensors

Actuators


17

Hierarchical Systems (Shakey Example)Shakey (1966 - 1972)• Developed at the Stanford Research Institute • Used STRIPS planner (operators, pre and post conditions)• Navigated in an office environment, trying to satisfy a goal given to it

on a teletype. It would, depending on the goal and circumstances, navigate around obstacles consisting of large painted blocks and wedges, push them out of the way, or push them to some desired location.

• Primary sensor: black-and-white television camera• Sting symbolic logic model of the world in the form of first order

predicate calculus• Very careful engineering of the environment

Video


18

Hybrid systems (three-layer architecture)

SENSE

PLAN

ACT


19

Hybrid systems (Minerva Example)Tour guide at the Smithsonian's NationalMuseum of American History (1998)

Video


20

Outline



21

Laser Range Finder


22

Laser Range Finder - the principle

0°

View from top

Video


Laser Range Finder - the principle cont’d

23


24

0°

40° 140°

α

β

View from top

α=180°β=100°

SICK LMS Resolution Modes Angular range

Angular resolution

0°..100° 1°0°..100° 0.5°0°..100° 0.250°..180° 1°0°..180° 0.5°

Mode Measurement Range

mm mode 0..8191mm = 8.191meterscm mode 0..8191cm = 81.91meters


25


26

Mapping

Based only on odometry + range data


Mapping cont’d

27

Based only on odometry + range data + matching 3D Mapping video


28

Mapping cont’d

Based only on odometry + range data + matching


29

Robotic perceptionRobot perception viewed as temporal inference from sequences of actions and measurements → dynamic Bayes network of first order Markov processExample: localization - holonomic robot with range sensor, estimate pose while moving

next belief state? → Bayesian inference problem


30

Recursive filtering equation

using Bayes’ rule, Markov assumption, theorem of total probability

Motion model: deterministic state prediction + noiseSensor model: likelihood of making observation ot when

robot is in state xt


31

Motion and sensor model

Assume Gaussian noise in motion prediction, sensor range measurements


32

Localization algorithmsParticle filter (Monte Carlo localization):• belief state represented by collection of particles that correspond to

states• belief state sampled, each sample weighted by likelihood it assigns

to new evidence, population resampled using weightsKalman filter:• belief state represented as single multivariate Gaussian• each step maps a Gaussian into a new Gaussian, i.e. it computes a

new mean and covariance matrix from the previous mean and covariance matrix

• assumes linear motion and measurement models (linearization →extended KF)


33

Global Localization (Particle Filter Examples)


34

Video_sonar

Video_laser

Global Localization (Particle Filter Examples; sonar vs laser)


35

Extended Kalman Filter Examples


36

More Localization Examples


37

SLAMLocalization: given map and observed landmarks, update pose

distributionMapping: given pose and observed landmarks, update map

distributionSLAM: given observed landmarks, update pose and map

distributionProbabilistic formulation of SLAM:add landmark locations L1, . . . , Lk to the state vector, proceed as for

localizationProblems:• dimensionality of map features has to be adjusted dynamically• identification of already mapped features

Laser-based SLAM Video Distributed SLAM Video


38

Outline



39

One DOF for each independent direction in which robot can move

Configuration of robot specified by 6 numbers => 6 DOFsNon-holonomic: number of effective DOFs less than controllable DOFs

Car has 3 effective DOF and 2 controls =>non-holonomic

Degree of freedom (DOF)


40

Motion PlanningMotion types:• point-to-point• compliant motion (screwing, pushing boxes)Representation: configuration space vs workspaceKinematic state: robot’s configuration (location, orientation, joint

angles), no velocities, no forces

Path planning: find path from one configuration to anotherProblem: continuous state space, can be high-dimensional


41

WorkspacePhysical 3D space (e.g. joint positions)Robot has rigid body of finite sizeWell-suited for collision checkingProblem: linkage constraint (not all workspace coordinates

attainable) makes path planning difficult in workspaceKinematics: given kinematic robot state in configuration

space, calculate workspace coordinates by applying chain of coordinate transformations (simple, well-posed problem)


42

Configuration SpaceSpace of robot states (e.g. joint angles)Robot is a point, obstacles have complex shapeProblem: tasks expressed in workspace coordinates, obstacle

representation problematicInverse kinematics: given workspace coordinates, calculate

kinematic robot state in configuration space (often hard and ill-posed problem)

Free space (attainable configurations) ->occupied space (not attainable configurations, obstacles)

Planner may generate configuration in configuration space, apply inverse kinematics and check in workspace for obstacles


43

Workspace vs. Configuration Space

ϕelb

ϕshou


44

Workspace vs. Configuration Space


45

Path PlanningBasic problem: infinite number of states ->convert to finite state

spaceCell decomposition:• divide up space into simple cells,• each of which can be traversed “easily”Skeletonization:• identify finite number of easily connected points/lines• that form a graph such that any two points are connected by a

path on the graphGraph search and coloring algorithmsAssumptions: motion deterministic, localization exact, static scenesNot robust with respect to small motion errors, doesn’t consider

limits due to robot dynamics


46

Cell DecompositionProblem: may be no path in pure free space cells

Soundness(wrong solution if cells are mixed)

vs.Completeness

(no solution if only pure free cells considered)

Solution: recursive decomposition of mixed (free+obstacle) cells or exact decompositionDoesn’t scale well for higher dimensions


47

Cell Decomposition

Grayscale shading - cost from the grid cell do the goal


48

SkeletonizationVisibility graphs: find lines connecting obstacle vertices through

free space, build and search graph; not for higher dimensionsVoronoi graphs: find all points in free space equidistant to two or

more obstacles, build and search graph; maximizes clearance, creates unnecessarily large detours, doesn’t scale well for higher dimensions

Probabilistic roadmaps:• generate randomly large number of configurations in free space,

build graph (construction phase) • search graph (query phase)scales better to higher dimensions but incomplete


49

Visibility and Voronoi Graph

qinit

qqoal


50

Probabilistic Roadmaps(PRM planning procedure example)


51

Probabilistic Roadmaps(PRM planning procedure example: Construction phase )

Generate randomconfigurations

Make connections


52

Probabilistic Roadmaps(PRM planning procedure example: Query Phase )

start

goal

start

goal

Curve Replacement & Smoothing

Add start andgoal configurationsto the roadmap

start

goal

A*


53

Probabilistic Roadmaps


54

Potential FieldsAdditional cost function in connectivity graph:potential grows when closer to obstaclesMaximizes clearance from obstacles while minimizing path length:• creates safer paths

Problem: choosing appropriate weight


55

Motion Control• Path following involves forces: friction, gravity, inertia,• Dynamic state: kinematic state + robot’s velocities• Transition models expressed as differential equations• Path planner assumes robot can follow any path• Robot’s inertia limits maneuverability

Problem: including dynamic state in planners makes motionplanning intractableSolution: simple kinematic planners + low-level controller for forcecalculationOther solution: motion control without planning: potential field andreactive control


56

Controllers• Techniques for generating robot controls in real time

using feedback from the environment to achieve a control objective

• Reference controller: keep robot on pre-planned path (reference path)

• Optimal controller: controller that optimizes a global cost function, e.g. optimal policies for MDPs

• Stable: small perturbations lead to bounded error between robot and reference signal

• Strictly stable: able to return to reference signal


57

Closed-loop control

Performance of controller:• stability• overshoot ← inertia• steady-state error ← friction• rise time• settling timeP controller:


58

Closed-loop controlPD controller:• decreases overshoot• decreases settling timePID controller:

• eliminates steady-state error


59

Closed-loop control


60

Potential Field ControlMotion planning: pot. field as additional cost functionGenerating robot motion directly: superpose potential field of

obstacles with field generated by an attractive force, follow the gradient

Efficient calculation possible, especially compared to path planningProblem: local minima, only kinematic method

Video


61

Potential Field ControlRobot is treated as a point under the influence

of an artificial potential field.Generated robot movement is similar to a ball

rolling down the hill• Goal generates attractive force• Obstacle are repulsive forces


62

Reactive ControlBefore: model of environment required (reference path, potential

field)Problems with model based approaches:• accurate models are difficult to obtain, especially in complex or

remote environments• computational difficulties, localization errors → reactive controlEvent driven systems (↔transformational systems) →finite state

machinesEnvironment feedback plays crucial role in the generated behaviorReactive controller are an implementation of a policy for a MDP,

policies can be learned (e.g. Q-learning)


63

Reactive Control

Emergent behavior: behavior that emerges through interplay of (simple) controller and (complex) environment

Characteristic of great number of biological organisms

Video

tddc17 percetion2 2010 presentation

Documents

Transcript of tddc17 percetion2 2010 presentation