orca – an organic control architecture for fault-tolerant...

UniversitUniversitäättzu Lzu Lüübeckbeck

Institut fInstitut füürrTechnische InformatikTechnische Informatik

Erik MaehleErik MaehleErik MaehleUniversity of Lübeck

Institute of Computer EngineeringRatzeburger Allee 160

D-23538 LübeckE-Mail: [email protected]

University of LUniversity of LüübeckbeckInstitute of Computer EngineeringInstitute of Computer Engineering

Ratzeburger Allee 160Ratzeburger Allee 160DD--23538 L23538 Lüübeckbeck

EE--Mail: Mail: [email protected]@iti.uni--luebeck.deluebeck.de

ORCA – An Organic Control Architecture forFault-Tolerant Mobile Robots *

ORCA ORCA –– An An OrganicOrganic ControlControl ArchitectureArchitecture forforFaultFault--TolerantTolerant Mobile Mobile RobotsRobots **

Lübeck-Tartu Workshop on InformaticsTartu, September 27, 2006

LLüübeckbeck--TartuTartu Workshop on Workshop on InformaticsInformaticsTartuTartu, September 27, 2006, September 27, 2006

* Supported by DFG under MA1412/7-1, associated to DFG SPP 1183 ‚Organic Computing‘



Research at ITI

Parallel & Fault-Tolerant Computing

NetRAID

ReconfigurableComputing

DynaCORE

Mobile Robotics

ORCA

Biomedical Technology

Wireless Patient Monitoring



Motivation

Autonomous mobile robots in human environments

-> unstructured, -> complex controldynamically changing systemsenvironment

-> no explicit model -> no explicit faultof the environment model

fault-tolerance, safety engineering bottleneck



Organic Systems

Organic systems adapt dynamically to environment and malfunctions

- uncertainties - unknown environment - unforeseen situations

Our approach: controlled self-organization

Inspiration: autonomic nervous system and immune system

Aim: - detect, react, and adapt to malfunctions

- avoid critical system states at any time

- low cost implementation and engineering



Dependability and Fault-Tolerance

• Definition of a fault model

• FT-Techniques for

- detection

- localisation

- recovery

of the modelled faults

• Redundancy in hardware, software or time

• Behavior in case of not modeled faults is unpredictable.

• Redundancy is expensive and difficult to implement.

• Reaches its limit for complex distributed systems.

Classical Fault Tolerance

In principle transferable to robots, but:

Duplex-System



Immune System

• Recognition also for unforeseen antigens (stochastic process)

• staged reaction: unspecific/specific (macrophages, antibodies)

• diversified

• self-organising

• self-regulating

• distributed

• able to learn (memory cells)

• emergent behavior

Can biological principles like in the immune system orautonomic nervous system be used for building Organic

Robots (‚Orgabots‘)?

B-Lymphocyte



ORGANIC COMPUTING

An Organic Computing System (OC-System) isdefined as a self-organizing system that adaptsdynamically to its respective environment. OC-Systems have the so-called ‚self-x properties‘, i.e. they are:

- self-configuring- self-optimizing- self-healing- self-explaining- self-protecting

Source: GI Informatics Dictionary



• Introduction

• ORCA Architecture

• Methods

• Walking Robot OSCAR

• Self-Organizing Walking

• Conclusion

Overview



Organic Robot Control Architecture

Goals:Self-organization adapt to malfunctions without a formal model

-> detect irregular, unexpected sensor/behavior patterns -> react in a safe and goal-directed way-> learn how to improve the system behavior

Learning online and in-situ under hard real-time constraints Safety avoid critical system states at any time

-> even when learning counteractions-> avoid potentially chaotic system behavior

Goal-Directedness stable improvement of behavior -> stability against getting worse by further learning

Low Cost overhead should be as low as possible (computation time, implementation, engineering)

Approach: controlled emergence on lower functional control levels



ORCA - Organic Robot Control Architecture

Reflexes

Motor (PWM)

Motor-Controller

Motors (PWM)

Perception Proprioception Motor-Controllers

Reflexes

Behaviours

Planning

Gait-Pattern-Gen.Gait-Pattern-Gen.Gait-Generation /-Selection

Gait-Pattern-Gen.Gait-Pattern-Gen.Sensor Modules

decisionallevel

functionallevel

hardwarelevel

OCU

OCUs

BCU = Basic Control UnitOCU = Organic Control Unit



OCU-BCU-Interaction

OCU

BCU

OCU

BCU

BCU

Integrated OCU Separate OCU



OCU-Architecture

- Monitor: anomaly detection

- Memory: short term history

- Reasoner: hard real-time determination of a counteraction

Monitor Reasoner

Memory



Methodological Approaches

- Signals reflecting the ´health´ of a signal or a BCU

e.g. - noisiness, confidence of output results, - load state; error state

- Adaptive action selection IF movement=blocked

THEN activation of commanded behaviour -=5%,

activation of non-commanded behaviour +=5%

- Learning to treat malfunctions

=> Learning at BCU- and OCU-level



Supervised Learning

- no formal models

- fast adaptation

- manageability

- controlled self-organization (safety)

-> hybrid crisp-fuzzy systems

- high-dimensional problems

-> adaptive filters

Supervised Process(robot, BCU, OCU, ...)

OCU



Hybrid Crisp-Fuzzy Systems (W. Brockmann, Osnabrück)

Key features:

- granular computing (partitioning of the input space)

- mixed crisp and fuzzy processing

- rule-based specification as well as learning

- very fast computation, hard real-time learning

- decomposition to fight complexity

- individual attributes for each partition

-> tag-bit for guided online-learning

-> safety warranty



Guided Online-Learning

- Typical BCU knowledge:IF e=X & ed=X THEN c=0, tag=free

IF e=PB & ed=X THEN c=80, tag=free

IF e=NB & ed=X THEN c=-80, tag=free

IF e=PB & ed=PB THEN c=100, tag=locked

IF e=NB & ed=NB THEN c=-100, tag=locked

- Typical OCU knowledge:IF d=Z THEN m=0

IF d=PS THEN m=-1 IF d=NS THEN m=1

IF d=PM THEN m=-2 IF d=NM THEN m=2

IF d=PB THEN m=-10 IF d=NB THEN m=10

e: control error, ed: its derivative, c: control action (-100…100), d: deviation, m: modification



Example for Guided Learning



Result of Guided Learning

A priori Knowledge After Learning

Control surface of angular controller (motor voltage in per cent)

.. ..



OSCAR - Organic Self-Configuring and Adapting Robot

Walking Robots arealready well studied, e.g.:- Genghis (MIT)- Katharina (Magdeburg)- Lauron III (Karlsruhe)- Scorpion (Bremen)- TUM Walking Robot- Fred (ITI)



Technical Data of OSCAR

Hexapod with 18 DOF (Servos)

Ground contact Sensor:Simple switch per leg

Control Computer:- JControl/Smartdisplay- Servo Driver Modul- I2C-Bus

Programming Language:JAVA

Mechanical parts from Lynxmotion



Hexapod Insect Walking

Stick Insect (Aretaon asperrimus)[Quelle: Bettina Bläsing, Uni Bielefeld]



Basic Approaches for Autonomic Walking

Stick Insect (Carausius morosus) Leg and Joints

AEP: Anterior Extreme PositionPEP: Posterior Exreme Position

Swing Phase

StancePhase

3 DOF



Distributed Coordination Rules for Leg Control [Cruse et. al. 90]

(1) Swing phase inhibits swing phaseof anterior leg.

(2) Start of a stance phase excites start of a swing phase of anterior leg.

(3) Increasingly posterior position excitesstart of swing phase of posterior leg.

(4) Tarsus position determines targetposture of swing phase of posterior leg.

(5) Increased resistance increases force(co-activation) and increased loadprolongs stance phase.

(6) Treading-on-tarsus reflex.

ANN Model WALKNET



Distributed Walking Control in OSCAR

Swing phase has constant length.

Duration of stance phase determines velocity.

Slightly different leg orientationdepending on leg position (F, M, L).

So far only one single rule implementedwith circular neighborhood relation for all legs.

Rule 1:



Self-Organizing Gait Patterns in OSCAR

SlowSlow GaitGait TetrapodTetrapod GaitGait TripodTripod GaitGait

EmergentEmergent GaitsGaits withwithIncreasingIncreasing SpeedSpeed!!

Medizinische Universität zu Lübeck

Institut fürTechnische Informatik

T I I

Monitoring of Leg Movement

Joint Sensors ( α, β, γ Servos):

- Current- Position

Foot Sensor:

- Switch or- Pressure

Medizinische Universität zu Lübeck

Institut fürTechnische Informatik

T I I

Current and Position Signals of Joints

Normal Movement Disturbed Movement

Swing StanceSwing Stance

Anomalies can in principle be detected.

Open Question: Can reasons for anomalies (obstacle, fault) be distinguished?



beta-servo: position and current

Marek Litza

Monitoring –Tolerance Band


Institut fInstitut füürrTechnische InformatikTechnische InformatikMarek Litza

Health Signal

Our Goal: Self-organization such that anomalies can be tolerated!



Insect Walking with Lost Legs

[Quelle: Holk Cruse, Uni Bielefeld]



Additional Rule for Tolerating Lost Legs [Schilling et al. 06]

1HF 1HF

RuleRule (1HF):(1HF):

(1) Swing phase inhibits swing phaseof anterior leg,

but is suppressed if the middle leg isloaded.

Extensions to WALKNET



Simulations with Expanded Walknet

[Quelle: Malte Schilling, Uni Bielefeld]



Leg Coordination After Leg Loss for OSCAR

Same single rule, onlyneighborhood relationchanges.



OSCAR Walking with Lost Middle Legs

Simulated Amputation

See also similar exeriments e.g.with Hannibal (MIT), Scorpion (U Bremen)



OSCAR with Lost Front or Hind leg



Leg-Shifting in OSCAR

In case of lost hind or front leg, defective leg is shifted into middle positionand leg orientation is changed accordingly.

However, no correspondence in biology!

x xx



Conclusion

ORCA – Organic Robot Control Architecture

- BCUs (Basic Control Units) for functional control

- OCUs (Organic Control Units) for monitoring, fault adaption and optimization

- Learning based on hybrid crisp fuzzy methods and adaptive filters

OSCAR - Organic Self-Configuring and Adapting Robot

- Self-organizing gait patterns inspired by WALKNET

- Monitoring of leg health status by sensor signals from joints and feet

- Self-organizing gait patterns also in case of lost legs



Further Work

- Tools for hybrid crisp fuzzy methods (U Osnabrück)

- Experiments with adaptive filters in mobile robot simulations (Fraunhofer AIS)

- Integration of learning methods into ORCA/OSCAR

- New hardware platform for OSCAR (2 * ARM9 + 6 * ATmega32)

- Additional sensors (inertial, ultrasonic, laser, inclinometer, compass, .. )

- Tolerance of less severe leg faults than complete leg losses

- Combination with obstacle avoidance in more challenging environments

- Application of ORCA to other robot platforms (e .g. fish-like underwater robots)

Acknowledgement of contributions of ORCA Group MembersWerner Brockmann (U Osnabrück), Adam El-Sayed-Auf, Karl-Erwin-Grosspietsch(Fraunhofer AIS), Bojan Jakimovski, Marek Litza, Florian Mösch

orca – an organic control architecture for fault-tolerant...

Documents