Robocup an application domain for distributed planning and sensoring in multi robot systems
Download
1 / 78

RoboCup: An Application Domain for Distributed Planning and Sensoring in Multi-robot Systems - PowerPoint PPT Presentation

Enrico Pagello President of the International IAS-Society RoboCup: An Application Domain for Distributed Planning and Sensoring in Multi-robot Systems IAS-Lab Intelligent Autonomous Systems The University of Padua Presentation Outline What a Cooperative Multi-Robot System should be

loader
I am the owner, or an agent authorized to act on behalf of the owner, of the copyrighted work described.
capcha

Download Presentation

RoboCup: An Application Domain for Distributed Planning and Sensoring in Multi-robot Systems

An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -

Presentation Transcript


Enrico Pagello

President of the International IAS-Society

RoboCup: An Application Domain for Distributed Planning and Sensoring in Multi-robot Systems

IAS-Lab

Intelligent Autonomous Systems

The University of Padua

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Presentation Outline

  • What a Cooperative Multi-Robot System should be

    • T. Arai, E. Pagello, L. Parker. Editorial: Advances in Multi-Robot Systems.

      IEEE/Trans. On R&A, Vol. 18, No. 5, pp 655-661, October 2002

  • Scientific perspective in RoboCup with respect to Cooperation

  • Research on RoboCup at IAS-Lab, The University of Padua

    • Distributed Sensoring: An Omnidirectional distributed vision sensor

      • E. Menegatti, A. Scarpa, D. Massarin, E. Ros, E. Pagello: Omnidirectional Distributed Vision System for a Team of Heterogenueous Robots. Proc. of IEEE Workshop on Omnidirectional Vision (Omnivis’03), Praga June 2003

      • E. Menegatti, A. Pretto, and E. Pagello Testing Omnidirectional Vision-based Monte-Carlo Localization under Occlusion. Proc. Of IROS-2004, Sendai (Japan), Sept 29 - Oct 2, 2004

    • Cooperative Robotics: An Hybrid Architecture a MSL Team

      • A. D’Angelo, E. Menegatti, and E. Pagello: How a cooperative behavior can emerge from a robot team. Proc. of DARS’04, Toulouse June 2004

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Why Multi-Robot Systems (MRS) have been so successful ?

  • In challenging application domains, MRS can often deal with tasks that are difficult, if not impossible, to be accomplished by an individual robot.

  • A team of robots may provide redundancy and contribute cooperatively to solve the assigned task, or they may perform the assigned task in a more reliable, faster, or cheaper way beyond what is possible with single robots.

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


What a Cooperative Multi-Robot System is ?

  • Cooperative Robotics research field is so new that no topic can be considered mature

  • Early research goes to

    • Cellular Robotics by [Fukuda, IECON 1987] and Cyclic Swarm by [Beni, Intelligent Control 1988]

    • Multi-Robot Motion Planning by [Arai, IROS 1989]

    • ACTRESS Architecture by [Asama, IROS 1989]

  • [Dudek, Autonomous Robots 1996] and [Cao, Autonomous Robots 1997] gave a taxonomy

  • In [Arai, Pagello, & Parker, IEEE/Trans. 2002]we identify several primary research areas

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Research roots for Cooperative Multi-Robot Systems

  • Cooperative mobile robotics research began after the new behavior-based control paradigm

    • Brooks 1986, Arkin 1990

  • Since behavior-based paradigm is rooted in biological inspirations, many researchers found it instructive to examine the social characteristics of insects and animals

  • The most common application is using simple local control rules of various biological societies, like ants, bees, and birds, for similar behaviors in MRS

    • MRS can flock, disperse, aggregate, forage, and follow trails, etc.

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


New and interesting research issues

  • The dynamics of ecosystems has been applied to MRS to demonstrate Emergent Cooperation

  • Cooperation in higher animals, such as wolf packs, has generated significant study in Predator-Prey Systems

    • Pursuit policies relay expected capture times to the speed and intelligence of the evaders and the sensing capabilties of the pursuers

  • Competition in MRS, such as in higher animals including humans, is being studied in domains such as multi-robot soccer.

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Inherently cooperative tasks

  • A particular challenging domain for MRS is the one

    whose tasks are inherently cooperative, that is,

    tasks in which the utility of the action of one robot

    is dependent upon teammates’ current actions

    • Inherently cooperative tasks cannot decomposed into

      independent sub-tasks to be solved by a DARS

    • Team success throughout task execution is measured by

      the by the combined actions of the robot team, rather than

      by individual actions

  • More recently identified biological topics of relevance are:

    • Imitation in higher animals to learnnew behaviors

    • Physical Interconnectivity by insects such as ants,

      to enable collective navigation over challenging terrains

    • How to maintain Communication in a distributed animal society

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Communication versus Cooperation

  • Communication issue in MRS started since the inception of Distributed Autonomous Robots Systems (DARS) research.

  • Distinctions between Implicit and ExplicitCommunication

    are usually made:

    • Implicit communication occurs as a side-effect of other actions,

      or “through the world”

    • Explicit communication is a specific act designed solely

      to convey information to other robots on the team.

  • Communication affects the performance of MRS in

    a variety of tasks

    • even a small amount of information can lead to great benefit

  • The challenge is to maintain a reliable communication even when connections between robots may change dynamically and unexpectedly

    • setting up and maintaining distributed network

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Architecture and Task Planning

  • Research in DARS has focused on the development of architectures, task planning capabilities, and control addressing the issues of:

    • action selection

    • heterogeneity versus homogeneity of robots

    • achieving coherence amidst team actions

    • resolving conflicts, etc.

  • Each architecture focuses on providing a specific type of DARS capability:

    • fault tolerance

    • swarm control

    • role assignment, etc.

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Architecture and Task Planning, Localization and Mapping

  • Research in DARS has focused on the development of architectures and task planning capabilities, where each architecture focuses on providing a specific type of distributed capability

  • Initially, most of the research took an existing algorithm

    developed for single robot mapping, localization, or exploration,

    and extended it to MRS

    • [Fox et al., Autonomous Robots 2000] took advantage of a MRS to improve positioning accuracy beyond single robot to develop a colaborative multi-robot exploration

  • Only more recently, researchers have developed new algorithms

    that are fundamentally distributed, to take advantage from MRS

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


RoboCup Soccer :The oldest RoboCup standard problem

  • Middle-size League

    • Building, maintaining, and programming a team of fully autonomous robots

    • High speed moving (>2m/s)

    • Large field (12m X 8 m)

    • Sensing the environment

    • Cooperation abilities

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


RoboCup Soccer :From simple moves towards complex actions

  • Middle-size League: progresses from 1997 to 2003

    USC (USA) - Osaka Univ. (Japan) Nagoya 1997 Isfahan Univ (Iran) - AIS (Germany) Padua 2003

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Middle-size League:RoboCup2003 : Vision and Localization

  • Vision is still a key research issue in MSL

    • All teams used color information

    • Half of teams use shape detection

    • Even less can make edge detection

    • Auto-color calibration is a hot topic,

      especially to relax lightning condition

  • Robot Self-Localization is mainly based

    on Visual Landmarks

    • Most teams detect corner posts

    • Half of teams detects also field lines

    • Several teams use statistical methods

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Middle-size League:RoboCup2003 : Control Architectures

  • One half of teams use

    reactive control architectures

    (behavior based robotics)

  • One third of teams use their

    own architectures like:

    Dual Dynamics, two-level FSMs,

    Fuzzy Approaches, etc.

  • Several teams develops advanced robot skills using learning

  • Only a few teams extends reactive motion control with path planners based mainly on potential field methods or similar

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Research on RoboCup topics@ IAS-Lab, Dept. of Information Engineering, The University of Padua

  • Soccer-robot design

  • ODVS (Omnidirectional Distributed Vision System)

  • MonteCarlo Localization

    using omni-vision

  • Coordinated behaviors

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Evolving the Artisti Veneti Team

  • First platform for MSL was designed on 1998 over a Pioneeer1 base

  • Second and third platforms evolved from a Pioneer1 to a Pioneer2 base

  • Third platform is a Golem robot

  • We shifted from 2-wheeled robot, with a directional camera, towards omnidrive and omnivision platforms

  • Fourth platform ehnance the circular movement of original goalkeeper

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Convex Mirror

Perspective camera

Perspex Cylinder (support)

Omnidirectional Sensor

Mirror

Camera

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Vertex

Vertex

Vertex

Vertex

Vertex

Vertex

Vertex

Vertex

Vertex

Vertex

x

y

Pin Hole

Pin Hole

Pin Hole

Pin Hole

Pin Hole

Pin Hole

Pin Hole

Pin Hole

Pin Hole

Pin Hole

P

dMax

dMax

dMax

dMax

dMax

dMax

dMax

dMax

dMax

d1

d1

d1

d1

d1

d1

d1

d1

d1

X

X

X

X

X

X

X

X

X

X

DMin

DMin

DMin

DMin

DMin

DMin

DMin

DMin

DMin

DMax

DMax

DMax

DMax

DMax

DMax

DMax

DMax

DMax

D1

D1

D1

D1

D1

D1

D1

D1

D1

How to design a mirror

Made by F. Nori at IAS-Lab

Mirror profile construction

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Mirror’s three parts:

Measurement Mirror

Marker Mirror

Proximity Mirror

Our robot mirrors

The task determines the mirror profile

Mirror Profile

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


A mirror designed forAIS – Fraunhofer Institut (Germany)

  • Three-parts mirror

  • Tailored on their mobile robot

  • Satisfing customer requirements

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


For Goalie:

Locate the ball

Identify the markers

See the defended goal

For Attacker:

Locate the ball

Identify the markers

See both goals

Lighter mirror

In the case of Soccer RobotsRequirements and profile

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Characteristics:

Chassis shaped for omnidirectional vision

Mirror profile designed for the robot’s task

Mirror

Camera

Heterogeneous Robots

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Heterogeneous Vision Systems

Peripheral vision

Foveal vision

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Heterogeneous Vision Systems

OVA’s view

PVA’s view

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Features and Events for Omnivision

Events:

  • A new edge

  • A disapearing edge

  • Two edges 180° apart

  • Two pairs od edges 180° apart

Features:

  • Vertical edges

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


P1

P2

P5

P4

P3

Omnidirectional Vision and Mapping

  • It simplifies data interpretation:

    • Discriminate b/t “turns” and “travels”

    • Simplify “Exploring around the block”

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Experimental Results

  • Correct tracking of edges

  • Recognition of actions

  • Calculation of the turn angle

The path segmentation

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Single Robot Mapping Strategy

  • Use an omnidirectional vision sensor

  • Detect topologically meaningful features in the environment

  • Use Spatial Semantic Hierarchy of Kuipers (SSH)

  • Build a topological map

  • Use the map to explore the environment

E. Menegatti, E. Pagello, M. Write Using Omnidirectional Vision within the Spatial Semantic Hierarchy IEEE/ICRA2002,Washington, May 2002

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Multi-robot mapping strategy

  • Every robot builds its own local map

  • When two robots can see each other, they share their local maps by matching their current views:

    • Identifying the objects seen by both robots

    • Estimating their relative distance and orientation

  • If the match is successful, they transmit their own local map to the teammate

  • Each robot connects this new local map to its local map

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Some hints

  • Every robot carries on an independent exploration by using use a misanthropy robotstrategy i.e.

    • Follow a direction of exploration that increases the distance

      from the visible teammates

  • Use redundacy of the observers and observation to improve the map

  • Exploit the heterogeneity of the robots more deeply in tasks too expensive (or not achievable) for homogeneous robots

  • Use maps of non previoulsy met robots to navigate. The bridge is the common starting location.

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


ODVS for Navigation

We realised a network of smart uncalibrated sensors able to learn how to navigate a blind service robot in an office like environment

The sensors learn by observing the robot motion.

The first stage is supervised, then the knowledge is propagated autonomously exploiting the overlapping field of view of the sensors

VA1

VA2

E. Menegatti, E. Pagello, T. Minato, T. Nakamura, H. Ishiguro

“Toward knowledge propagation in an omnidirectional distributed vision system”

Proc. of 1st Int. Workshop on Advances in Service Robotics (ASER'03),

Bardolino (Italy), March 2003

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Implicit Communication

  • VA1 learns its own mapping

  • VA1 moves the robot in the field of view of VA2

  • VA2 observes the robot

  • VA2 receives from VA1 the motor commands sent to the robot

  • VA2 trains its own neural nets to build its own mapping

VA1

VA2

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Why Monte-Carlo Localization

  • Monte Carlo Localisation (MCL) as a very successful approach

  • Applying MCL to omnidirectional vision used as a range finder

  • An experimentally generated sensor model

  • The fusion of sensor data for pose likelyhood calculation

  • A global localization experiment in a RoboCup Environment

  • Robustness to occlusion

  • An application to a non-roboCup Environment

    E. Menegatti, A. Pretto, E. Pagello

    A New Omnidirectional Vision Sensor for Monte-Carlo Localization

    Proc. of 8th RoboCup Int. Symposium, Lisbon (Portugal), July 2004

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


MCL (Monte Carlo Localisation) in one page

  • MCL is a probabilistic technique to estimate the robot’s positions

    from the odometric and sensor data

  • We calculate the probability density of robot positions (the belief)

    by a set of weighted samples

  • The samples are localisation hypothesis

  • When the robot moves, everytime a new image is processed,

    the samples are moved in accordance with the motion model

  • To every sample is associated a weight proportional to the probablity

    that the robot is occuying that position

  • When the robot grasps new data, the sample weights are updated

    according to the sensor model

  • At every step a resampling eliminates the less probable positions

E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


Our approach to MCL

  • Starting from the work of [Kröse, IVC 2001] and [Burgard, ICRA 2002] , we realised an omnidirectional image-based Monte Carlo Localisation system for a large office environment [Menegatti, RAS 2004]

  • We decided to port a similar approach in RoboCup, but image-based localisation is not suited due to:

    • (i) many occlusions

    • (ii) an high dynamical environment

    • (iii) high computational costs for processing the whole image

  • Previous works in RoboCup implemented MCL using complex method for landmark or feature detection, and need to cope with dynamic occlusions

    [Utz, RoboCup-IV 2001], [Enderle, IAS2000]

  • We fell back on range-scanner, like [Fox, JAIR 1999][Thrun, AI 2000]

  • E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Our omnidirectional enhanced range finder

    • We detect colour transitions of interest:

    • G- W, G -Y, G - Blue

    • We detect occlusion:

    • G - Black

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Probability distribution of the robot’s pose

    • The scan of every colour transition of interest (here Green-White)

    • gives a probability distribution in the whole field.

      • Black dots = high probability , White dots = low probability

    • Note the symmetry in the environment

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Sensor Model.1 - Calculating p(o|l)

    • p(o|l) is the probability to have a the scan o at the location l

    • oi is the measurement along the single ray i of the scan

    • Omni-Scan:

    • One scan per colour transition of interest

    • Every scan has 60 rays (one every 6°)

    • Every ray has one receptor every 4 cm from 10 cm to 4 meters

    • When a transition is found the ray is not searched anymore

    {i = 1:60}

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Expected and Real Scans

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Sensor Model.2 – estimating p(oi|l)

    • Taking 2000 images in different positions in the field

    • For every ray of the 2000 scans

      • Computing theactual distance of the colour transition (here Green-White)

      • Estimating the distance of the colour transition with the vision software

      • Running the Expectation Maximisation (EM) to fit the experimental data separately for every colour transition

    Expected Distance

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Sensor Model.3 –Results

    • The resulting probability density calculated for every colour transition is the sum of three components:

    • Erlang distribution (accounting for image noise and imperfect colour quantization)

    • Gaussian distribution centered around the expected distance

    • Discrete density (accounting for missing the transition)

    Gaussian

    Discrete

    Erlang

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Combining the three probability distributions

    Probability distribution for the green-white ToI

    Probability distribution for the green-blueToI

    Probability distribution for the green-yellow ToI

    Resulting Probability distribution for the robot’s pose

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Global Localisation

    Step 0

    Step 4

    Step 6

    Step 18

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Sensor Occlusion.1

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Sensor Occlusion.2

    Our system is able to recognise occlusion by other robots as a Green-Black ToI along a ray

    These rays are labeled as FAKE_RAY (f)

    and discarded from the calculation of p(o|l)

    We called this process ray discrimination

    Our system scans with less rays (so less information), but keeps

    the usable information and avoids using expensive algorithm as distance filters.

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Sperimentation at University Building

    • We need uniformly colored surfaces, clear color gaps, and uniform light

    • Red floor, white walls, and gray furnitures

    • New color transitions:

    • Red - White, Red - Gray

    • The omnidirectional image is scanned with 60 rays, one every 6 degrees

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Ideal scans and probabilities

    in real environments

    • The ideal scan was different from the real one:

    • Robot shadow

    • Mirror deformation

    • Error in color detection near the door

    • In the probability map of the environment, there are dark zones everywhere the probability to have the observation is higher:

    • All cornered zones are darker

    • The samples closer to the real pose have a higher weight

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Extending the limits of the sensorial horizon of the single agent

    • The first step: using omnidirectional vision (RoboCup is an example of this)

    • But, RoboCup proved omnidirectional vision

      is not enough for highly dynamic environments:

      • cannot see occluded objects

      • cannot see very distant objects

    • To realise a Distributed Vision Systemwe need to share information between the agents of a team

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Omnidirectional Distributed Vision System (ODVS)

    Tracking multiple moving objects in highly dynamic environments

    by sharing the information gathered by every single robot

    • Requirements:

    • Robots’ only sensor: omnidirectional vision

    • No use of external computer

    • Every robot shares its measures

    • Every robot fuses all measures received

    • by teammates

    • Measures can refer to different instants

    • in time

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Enhancing the ODVS by fusing multiple observations

    E. Menegatti, A. Scarpa, D. Massarin, E. Ros, E. Pagello

    Omnidirectional Distributed Vision System for a Team of Heterogenueous Robots

    Proc. of IEEE Workshop on Omnidirectional Vision (Omnivis’03), Praga June 2003

    • Fusing Multiple Observations from Single Measurements

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Inspiring Works

    • Stroupe et al. [ICRA 2001]

      • Perspective cameras

      • They fused measurements made at the same instant in time

      • No tracking, just recognition

    • Gutmann et al. [IROS 2001]

      • Laser Range Finders + Perspective cameras

      • External Global Sensor Integrator

      • Robot uses external information only for unseen objects

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Problems for ODVS

    • Our Robots are heterogeneous:

      • Omni-sensors are different

      • On-board processing power is different

      • Robots’ platforms are different

    • The robot need to share:

      • the same spatial frame of reference

      • the same temporal frame of reference

    • The system must be robust to failure of single robots

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Single Sensor

    Architecture of the Perception Module

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Single Measures

    • A 2D Gaussian is associated to every measure

    • The Gaussian represents the probability that

    • the object is actually located at that point

    • Gaussian widths are determined

    • experimentally for every single robot

    • To share the measurements with other robots:

    • Measure is transformed in the absolute reference frame of play field

    • A time stamp is associated to every measure

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Fusing Multiple Observations (1)

    • Measures come from:

      • the vision system of the single robot

      • the vision system of the teammates

    • Measures can refer to:

      • Different objects

      • The same object(this is the most frequent case because of omnidirectional vision)

    • They are processed in the same way:

      • They are fused using a Kalman filter

      • They are stored in ‘tracks’

      • Multiple tracks allowed for single object (Multi-modal distribution)

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Fusing Multiple Observations (2)

    • Two measurements (i.e. two Gaussians) are fused with a Kalman filter:

    Mean Object Position

    Associated Variance

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Tracks management

    • Every new measure is compared with existing tracks:

      • If compatible the measure is added to the track

      • If NOT compatible a new track is created

    • When a track is not updated the associated variance increase

    • Over certain threshold the track is deleted

    • We allows more tracks for the same object

    • Real position assumed to be the one of smallest variance

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Measures from Different Robots

    Problems:

    • sharing the same spatial frame of reference

    • sharing the same temporal frame of reference

    • Trusting teammates

    • Managing ‘old’’ measurements

      Adopted solutions:

    • Robust self-localisation thanks to omni-vision

    • Internal clock synchronised via Network Time Protocol (NTP)

    • Variance of measures from teammates are doubled

    • The state of the object is recalculated

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Old Measurements

    • Old measurements cannot be thrown away

      For example: Very slow vision system reporting

      very accurate measurements

    Image processing time

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Experiment (1)

    Ball moving between steady robots

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Experiment (2)

    Moving ball and ‘blind robot’

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Experiment (3)

    Kidnapped ball

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Discussion

    • We implemented an Omnidirectional Distributed Vision System

    • The system is robust to failure of the single robots

    • The system exploits:

      • the heterogenity of the sensors

      • The redundancy of the observations

    • We presented experiments in real game scenarios

    • The system requires fine tuning of the parameters:

      • Variance associate to every measure

      • Rate of growth of variance when track not updated

      • Variance of teammates’ observations

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    A Hybrid Architecture for MRS

    • We suggest to use an Hybrid architecture where

      the Deliberative part and the Reactive part

      can take mutual advantages.

    • We introduced Robot Schemas at the low level,

      as building blocks to grow-up complex behaviors

      from simple ones, according to Arbib and Arkin :

      • Behaviors are chunks of basic knowledge of

        how to act and perceive.

    • Each behavior is implemented with a schema composed by

      • a motor schema, representing the physical activities

      • a perceptual schema which includes the sensing

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    The Perceptual/Motor Schema

    • At each level, the primitive control component is

      a behavior built by perceptual and motor schemas only.

    • The lower reactive level uses only information coming from sensors, and feeds the motors with appropriate commands.

      • It can elaborate on some perceptual patterns generated by other individual robots, both opponents and temmates.

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    An Abstract Architecture

    • Compound behaviors appear only at higher level, when they may receive more structured information about the environment.

    • Only the higher deliberative levels refer to cooperative capabilities that any robot could exhibit as a teammate, while a cooperative behavior is going to emerge.

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    The Layered Levels of Control

    • By releasing a behavior, we fire an activation-inhibition mechanism, built on some given evaluation condition rule, at some level of abstraction.

      • Simple Behaviors like defendArea, or carryBall, are implemented as motor schemas accessing directly the robot effectors.

      • Basic Behaviors, like playDefensive, and chaseBall, are obtained by simply appending two perceptual schemas seeBall and haveBall.

        • playDefensive : seeBall --> defendArea

        • chaseBall : haveBall --> carryBall

    • Since a primitive behavior results in appending just one perceptual schema to one motor schema, at the reactive level we obtain sensori-motor coordinations

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Software architecture:ArtiFACT (Artisti Fuzzy Agents Control Toolkit)

    • We designed a new hybrid deliberative reactive architecture.

    • The classic deliberative paradigm (Sense-Reason-Act) has been evolved reinforcing reactive behaviors.

    • A direct link between sense and act has been introduced to speed-up the reactive response of the robot

    • Thus, deliberative conditions can be bypassed for certain inputs which need more reactive behaviors

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    A Functional Architecture

    • The architecture of each single robot shows

      • An inner loop, for close feedback,

      • An outer looop, for high level reasoning.

    • To allow cooperation with teammates, twosensorial sources can input asynchronously both

      • Environment constraints (the “Ruler”)

      • Information about teammates (the “Teamplay”)

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    On Role Allocation in RoboCup

    • Inspired by Stone and Veloso’s pioneering work, many teams employ role-based coordination, in which robots can take on different static roles within the team

    • Although it would be possible to statically assign roles once forever, most teams switched to dynamic role allocation, by solving an iterated assignement problem, where the current allocation is re-evaluated periodically 10 times for each second

      • Given n robots, n prioritized (weighted) single-robot roles, and some estimates of how well each robot can be expected to play each role, assign robots to roles so as to maximize the overall expected performance

    • Gerkey and Mataric [Springer Book on RoboCup2004] showed that this technique is an instance of the canonical Greedy algorithm for Optimization theory

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    RoboCup Team solutions adopted

    • RoboCup role allocation problem is similar to task allocation problem for MRS in order to cooperatively achieve the goal, where a time-extended role concept replace that of a transient task

    • CS Friburg Team used a distributed role allocation mechanism in which two robots may exchange roles only if both want to do it, both moving to a higher-utility role for themselves.

    • ART Team, as well as early, Artisti Veneti Team, ordered the roles in a descending priority, and then assigned each to the available robot with the highest utility.

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Utility Functions

    • Multi-robot role allocation is a dynamic decision problem,

      that varies in time, according to the environmental changes,

    • Utility concept rely on the fact that each individual robot can somehow internally estimate the value (i.e. the cost) of

      executing an action

      • In RoboCup it is common to compute utility as the weighted sum of factors like distance to target, distance to ball, defence-offense coonfigurations, etc.

      • The computation is affected by sensor noise, general uncertainties, and environmental changes

      • Given the utility value Uij of each robot i for each role j, find the highest utility Uij, assign robot i to role j, and iterate

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Dynamic role assignment in practice

    • We developed an enhanced reactive approach starting from

      behavior-based hand-coded software

    • Dynamic role assignments among attacker, supporter and defender,

      were managed by considering collision avoidance issues and

      competitive behaviors

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Coordinating the Master/Supporter Roles

    • Consider the coordination between two robots carrying the ball towards the opponent’s goal:

    • We may indentify a Master Role and a Supporter Role

    • Roles can be played at different responsibility levels:

      • Can be >>> Assume >>> Acquire >>> Advocate

    • Ball assignments depend on Ball Possesses

      • HaveBall condition allows to discriminate

        which robot is really carrying the ball

        • It is an Environment constraint acting as a kind of Macroparameter, evaluated by different teammates

        • It allows to synchronize the activation of a new cooperation pattern

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Notifying Roles

    • Roles can be switched provided a notification is exchanged between teammates

    • A notification implies a communication between teammates based on a

      first-notified/first-advocated basis

    • A notify(Role) rule is:

      Supporter (mate) -->> reply (role, mate)

      Master (mate) --> request (role, mate)

    • Environment Rules require that a Master role must be advocated, whereas a Supporter role should be acquired.

    • haveBall and notify (Role) are the two allowed asynchronous communication from outside for a single robot

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Constructing Clamping Behaviors

    • A role is switched from acquire to advocate, or from assume

      to acquire, provided a notification is made to its teammate

    • Two complex Clamping Behaviors for Master and Supporter

      can be constructed from notify (x) and haveBall (z)

      • The Master robot shows a chase_ball behavior

        • haveBall (me) & not haveBall (mate) -->> acquire (Master)

        • Acquire (Master) & Notify (Master) -->> advocate (Master)

      • The Supporter robot shows an approach_ball behavior

        • Not acquire (Master) & canBe (Supporter) -->> assume (Supporter)

        • Assume (Supporter) & Notify (Supporter) -->> acquire (Supporter)

    • The robot chasing the ball suggests a teammate to become supporter by advocating a master role, and forcing the other robot to acquire a supporter role by approaching the ball.

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    Single Robot Architectureand effect on coordination

    • Conditions are defined as fuzzy functions. A value is returned depending on how strongly the condition is met

    • Team coordination is obtained by incorporating some conditions depending on messages coming from other robots, when the condition is evaluated

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    The Artisti Veneti Team

    www.dei.unipd.it/~robocup

    E. Pagello, RoboCup: Distributed Planning and Sensoring in MRS


    ad
  • Login