RFID Technology-based Mapping and Exploration by Humans and Mobile Robots Dr. Alexander Kleiner Institut für Informatik Arbeitsgruppe “Grundlagen der Künstlichen.

RFID Technology-based Mapping and Exploration by Humans and

Mobile Robots

Dr. Alexander KleinerInstitut für Informatik

Arbeitsgruppe “Grundlagen der Künstlichen Intelligenz”, Prof. Dr. Bernhard Nebel

Albert-Ludwigs-Universität Freiburg

kleiner@informatik.uni-freiburg.dewww.informatik.uni-freiburg.de/~kleiner

A. Kleiner RFID Technology-based Mapping and Exploration by Humans and Mobile Robots 2

Related robotics activities of our groupCompetitions we recently participated

Rescue robots (RoboCup Rescue)

Multi-robot exploration (RoboCup Rescue Sim.)

RFID-based SLAM

This talk

Fast robots (Sick Robot Day)

All-terrain navigating robots (TechX challenge)

MotivationThe “golden” 72 hours

Courtesy S. Tadokoro

Courtesy R. Murphy

Tom Haus (firemen at 9/11): “We need a tracking system that tells us where we are, where we have been, and where we have to go to”

Outline

• Introduction• RFID SLAM

– Centralized – Decentralized (DRFID SLAM)– Ongoing work: SLAM with active RFID

• Multi-robot exploration– Local exploration– Global exploration

• Conclusions

IntroductionMapping and Exploration within US&R environments

• Mapping: Computing globally consistent maps from pose tracking and data association by one or multiple agents

• Exploration: Efficient coverage of an unknown environment by one or multiple agents

• Requirements within harsh real-world domains (e.g. US&R):– Real-time computation– Decentralized with limited radio communication– Mixed-initiative teams: Integration of robotic solutions into human

organizations (e.g. first responders)• Limitations of existing solutions:

– Data association problem:• Dynamic illumination conditions• Unstructured environment in 3D

– “Loop closure” requirement • Cannot be guaranteed during time critical missions (e.g. victim search)

IntroductionSolutions presented in this talk

• Decentralized team coordination and Simultaneous Localization And Mapping (SLAM) via local information nodes, e.g. RFIDs– World-wide unique labeling of places

• No data association problem– Exchange of map pieces between teams of robots and

humans• “Loop-closure” on joint maps

– Information sharing via local node memories• Facilitates decentralized mapping & exploration with indirect

communication– Topological node graph structure

• Efficient multi-robot task assignment and path planning

Outline

• Related activities• Introduction• RFID SLAM

• Conclusions

RFID-SLAMHardware Setting Robot & Human

Robot:- 4WD (four shaft encoders)- Inertial Measurement Unit (IMU)- RFID antenna for detecting tags lying beneath the robot- RFID deploy device

Human:-IMU for step counting and angle estimation- RFID glove

Glove for sensing RFIDs (TZI Bremen)

13.56 MHz RFIDs

RFID antenna and deploy device

IMUZerg robot

RFID:- 13.56 MHz (short range below a meter)- 2048 Bit RAM, programmable by the user

Test person

RFID-SLAMAutonomous deployment of RFIDs

Robot pose trackingKalman filter for updating pose estimates

• Typically applied with constant error covariance• However, particularly outdoors, wheel odometry errors

are situation dependent, e.g. the specific type of ground• Solution: Adaptive odometry model

Robot pose trackingSlip detection from redundant wheel odometry

• Wheels on the same side turn at different speeds under slip! Measuring with 4 shaft encoders

• Decision tree model for slip detection– with the wheel-velocity

differences Δvleft, Δvright, Δvrear, Δvfront, as classifier input

• Determination of error estimate σslip by computing the RMS error with scan matching GT

Robot Pose Tracking Slippage-sensitive odometry

Odometry distance estimate with 3σ bound compared to ground truth computed from laser-

based scan matching

Conventional odometry: covariance bound does

not hold

Slippage-sensitive

odometry: Reduced

distance error within valid

bounds

Human pose tracking“Odometry” from accelerometer, gyros, and magnetometer

• Human foot steps generate vertical accelerations

• We use a method from Ladetto et al. that extracts acceleration maxima for counting foot steps

• Individual step length is automatically calibrated from GPS readings (if available), yielding distance estimates from steps

• Orientation changes are measured by gyroscope and magnetometer

• Kalman filter-based pose tracking from increments yields estimate d = (x,y,θ) with 3x3 covariance matrix Σ

Acceleration patterns during walking

Tracked pose (green)

• Analogy to spring-mass: Find low energy arrangement of springs (estimates ) and connected masses (nodes):

• Estimation of inter-node displacements by pose tracking methods• Construction of joint graph G from all observations , consisting of measured

distances and 3x3 covariance matrix • Loops are detected if nodes have been observed twice

– Modeled by an observation edge with , where Δθ denotes the angle difference, and covariance reflecting max. detection range of the antenna

• From G a globally consistent map is calculated by minimization of the Mahalanobis distance (Lu & Milios 97) :

Centralized RFID-SLAMBuilding globally consistent maps

ijij Σ,d̂

,0,0d̂ ii

ijΣijij Σ,d̂

ijijij yx ̂,ˆ,ˆd̂ ij

Outline

• Conclusions

Decentralized RFID-SLAM (DRFID-SLAM)Indirect communication via node memories

• The basic idea: – To utilize the memory of nodes for learning the topology of the

graph– Mobile agents are accumulating graphs Aj G from observations

on their path – Nodes are learning local graphs Ri G representing the topology

of their vicinity– Agents propagate information through the network, i.e. synchronize

nodes if they are in range• Information update:

– When traveling from node i to node k:• Add new estimate to Rk

• Update Aj from Rk by graph merging, and vice versa

• Double edges: – Locally fused by adjacent nodes – ID value for each fusion for preventing doubly counting

ijij Σ,d̂

Decentralized optimization of A2

Second agent trajectory

Graph A2

DRFID-SLAM cont.Example

RFID node

ijij Σ,d̂

Decentralized optimization of A3

First agent trajectory

Graph A1

Third agent trajectory

Graph A3

RFID-SLAM jointly by humans and robots Corrected map

Robot (orange) and pedestrian (red) odometry Corrected track (green) compared to GPS ground truth (blue)

-Robot driving at 1.58 m/s for 2.5 km

-10 RFIDs

- Optimization time below a second

RFID-SLAM jointly by humans and robotsCovariance bounds

3σ bounds from slippage-sensitive odometry

3σ bounds after the optimization

RFID-SLAM by a team of humansCentralized graph optimization in a large-scale environment

Pedestrian tracks recorded in the City of Freiburg

Corrected RFID graph

DRFID-SLAM (from data) Decentralized graph optimization during the outdoor exp.

• Simulation of all 720 possible sequences of 6 agents exploring the environment

• The more agents visited the area, the better the individual map improvements

DRFID-SLAM ExperimentDecentralized graph optimization during the outdoor exp.

Outline

• Conclusions

Ongoing workSLAM with active RFID

Robot team with 9 sensor nodes

Zigbee Sensor node developed in Freiburg (Dept. of Microsystems

Engineering)

Sensor nodes placed outdoors for experiments

USARSim environment for simulated large scale experiments

SLAM with active RFIDDistance estimation from RSSI (signal strength)

Relation between Transceiver-Receiver (TR) separation d and signal strength P (Seidel & Rapport 1992):

Path loss at reference distance d0

Noise with variance σ

However, many outliers in the data!

We use the RANSAC (Random Sample Consensus) method to identify outliers.

Resulting model for the utilized Zigbee modules

SLAM with active RFIDBearing estimation by voting grids

• Omni-directional antennas provide no bearing information!• Bearing can be determined from intersections of range

measurements at different robot locations (by odometry)• Each cell on the grid votes for the RFID location

Observations during navigation Step-wise integration yields location estimate

SLAM with active RFIDResults from USARSim experiments

Results from experiments on different USARSim maps

Odometry on “Pywood” map

RSLAM on “Plywood” map Groundtruth of “Plywood” map

SLAM with active RFIDResults from real world experiments

Robot odometry RSLAM

Experiment 1:

Experiment 2:

Outline

• Conclusions

RFID-based explorationHybrid: local exploration and global planning

• Local exploration (LE):– Indirect communication– Scales-up with # of robots and environment size– Inefficient exploration due to local minima

• Global task assignment and path planning– Based on node graph abstraction of the environment– Monitors LE and computes new agent-node

assignment If exploration overlap is high– Requires communication

Local explorationNavigation and target selection

• Navigation based on (limited) robot-centric grid map generated from laser ranges

• Exploration targets taken from grid frontier cells [Yamauchi, 1997]• Coordination:

– Automatic node deployment w. r. t. a pre-defined density– Discretization of node vicinity into equally sized patches– Node memory for counting visits of each patch [Svennebring and

Koenig, 2004])

– Frontier selection by minimizing the following cost function:

lfi : frontier cell location, LRS: set of nodes within range, Pr: set of patches around node r, d(.): the Euclidean distance

Discretized visited areas counted in memory

Local exploration cont.Discretization of the node’s vicinity π

RFID node

Robot trajectories

Relative addressing!

Results Local Team CoordinationVirtual rescue scenarios from NIST (RoboCup’06)

Each color denotes the path of a single robot

Largest explored area (by 8 robots)

Final 1 (indoor, 1276m2) Final 2 (outdoor. 1203m2)

Global explorationTask assignment and planning

• Task assignment:– Sequential robot planning to best targets [Burgard et al., 2005]– Genetic algorithm (GA) for finding optimal planning sequence

• Score computed from multi-robot plan cost

• Initialized by greedy sequence

• Computation of multi-robot plan:– A* time space planning to multiple goals [Bennewitz et al., 2001]– Plan costs: joint plan length + conflict penalties (infinite if

deadlock)– Heuristic: based on pre-computed shortest Dijkstra tree ignoring

conflicts

Results Global Team Coordination Task assignment and planning on node graph (USARSim outdoor map)

Robot start nodes

Goal nodes

Multi-robot plan

Conflicts vs. # of robots: Greedy (red), GA assignment (blue), GA sequence (green)

Rescue Virtual CompetitionVideos from RoboCup’06

Semi-Final`06 Final`06

Conclusions and Future Work

• Robust and efficient methods for DSLAM and exploration – Limited radio communication– No requirement for direct loop closure– Local information exchange– Joint human and robot exploration– Coordination scalable in terms of communication and

computation

• Future work: – Experiments with far-range RFID technology, such as ZigBee

transponders– Spontaneous node-to-node communication for synchronizing

position estimates and explored areas

Thanks for your attention, any questions?

RFID Technology-based Mapping and Exploration by Humans and Mobile Robots Dr. Alexander Kleiner Institut für Informatik Arbeitsgruppe “Grundlagen der Künstlichen.

fre technologybased

mapping slam

robot rfid

mobile robotsdr

mobile robotsrfidslamauto

usr environmentsmapping

teams of robots

unknown environment

Documents

Arbeitsgruppe Meteorologische Umweltforschung/Klimatologie

Akustische Kenntlichmachung von künstlichen...

ETH Arbeitsgruppe Radiochemie

Kleiner Münsterländer

VDI ZRE Studie: Potenziale der schwachen künstlichen ...

Direktblendung der künstlichen - LiTG

GK Info 13 (Hö)PROLOG – Eine Sprache der Künstlichen...

Arbeitsgruppe Visionen Inklusive Projekte

Spielplatz Jever, Kleiner Moorweg -...

Inhaltsverzeichnis 1700Pilze Buch - 3600 Pilze … · Web.....

Lichtplanung im künstlichen Himmel...Lichtplanung im...

kleiner Zentralbankkurs

Arbeitsgruppe Fachreferat Naturwissenschaften

EXPO HYPER– REALISMCeci n’est pas un corps...

Bewerbung kleiner

Kleiner Nazareno