robot body sense 1 - department of computer sciencekuipers/handouts/s07/lauzon.pdf · agostino...
TRANSCRIPT
1
1
Robot Body Sense
Self-Calibration and Health Monitoring
Thomas Lauzon University of Texas at Austin
March 1 20073/1/2007 2
Plan
1. Problem Presentation1. Self-calibration2. Health Monitoring
2. Methods and Tools for Self-calibration3. Existing work on Health Monitoring4. Conclusion
1. Research Objectives2. Evaluation3. Value
3/1/2007 3
Self-Calibration
The ability to infer system parameters without humanassistance
Ex: On-line use of Laser Range Finder data to determine
the drift error from odometry. Problems:
Sensors and actuators are noisy Natural Wear and Tear over time
Ex: Wheel diameter, structural deformations Weight of passenger varies Damage
3/1/2007 4
Health Monitoring
The ability to assess the behavioral mode of adevice or the system
Problems: How to distinguish natural wear and tear from
damage? Where is the defect located? How severe is the defect? What is the impact on the operability of the system?
Knowing the health condition can help self-calibration and maintenance
3/1/2007 5
Further Reflexions
Change in the environment may favor the use of atype of sensor over another. Terrain change: impact on odometry and belief of
position based on given action command Ex: dirt, oil, water, cement, wood, carpet, tile…
In case of fire: Smoke prevents use of laser and cameras, navigation
must be based on odometry and sonars. Lens can become dirty. The operator can clean it if he
is aware of it. More sensors give better calibration when these are non-defective
3/1/2007 6
Plan
1. Problem Presentation1. Self-calibration2. Health Monitoring
2. Methods and Tools for Self-calibration3. Existing work on Health Monitoring4. Conclusion
1. Research Objectives2. Evaluation3. Value
2
3/1/2007 7
Methods and tools: Calibration
Two types of parameters: Intrinsic: Focal Length… Extrinsic: 3D Position, roll, pitch, yaw
Two types of methods Manual Calibration Automatic Calibration (Self-Calibration)
Two types of calibration Independent (one device) Based on other sensor observations
3/1/2007 8
Methods and tools: Calibration
Important Paper:“Online Self-Calibration For Mobile Robots”Nicholas Roy and Sebastian Thrun, ICRA 1999 Presents a way to calibrate odometry based
on incremental maximum likelihoodestimation.
Unknown parameters are the real rotationaland translational drift δrot and δtrans
3/1/2007 9
Methods and tools: CalibrationTrue values Measured
valuesSystematicerror (drift)
Randomerror withzero mean
3/1/2007 10
Methods and tools: Calibration Process:
1. Acquire Scan s(i)2. Update occupancy grid with s(i)3. Move to a new location, and record o(i)4. Acquire s(i+1)5. For each position error, compute P(s(i+1)|s(i),o(i), δrot ,
δtrans)6. Compute the most likely δrot , δtrans: δrot(i)* , δtrans(i)*
(argmax of previous calculations)7. Compute δ rot* , δtrans* based on new estimated drift and
global estimated drift8. Update position using computed drift
Note: this doesn’t require recording all past data.
3/1/2007 11
Methods and tools: Calibration
Results
3/1/2007 12
Methods and tools: Calibration
Other methods: “Extrinsic Calibration of a Camera and Laser-
Range finder” Robert Pless, Qilong Zhang 2003 Method used: Registration of laser scan line on a
planar pattern
3
3/1/2007 13
Methods and tools: Calibration
Learning to improve models: “Simultaneous Calibration of Action and
Sensor Models on a Mobile Robot” Daniel Stronger and Peter Stone (UT), ICRA
2005
3/1/2007 14
Methods and tools: Calibration
Extended Kalman Filter “Automatic Self-Calibration of a Vision-
System during Robot Motion” ICRA 2006 Agostino Martelli, Davide Scaramuzza and
Roland Siegwart High Accuracy Requires a light source
3/1/2007 15
Methods and tools: Calibration
Hartley’s self-calibration (based on essentialmatrix properties): “A simple Technique for Self-Calibration” 1999
Paulo R. S. Mendonça and Roberto Cipolla Calculates Camera intrinsic parameters (focal
length) of a stereo pair
3/1/2007 16
Calibration: General Remarks
All these papers offer good modern techniques forself-calibration.
Some consider the calibration of one device alone. Some consider the calibration of one device with
respect to another A lot of calibration techniques cover camera
calibration They don’t cover Calibration based on more sensors They don’t suggest anything about defective sensors
or actuators
3/1/2007 17
Plan
1. Problem Presentation1. Self-calibration2. Health Monitoring
2. Methods and Tools for Self-calibration3. Existing work on Health monitoring4. Conclusion
1. Research Objectives2. Evaluation3. Value
3/1/2007 18
Health Monitoring
Most health monitoring systems areprobabilistic or statistical:
These include: Run charts Bayesian methods Hidden semi-Markov models
Other systems use heuristics One recent paper illustrates the use of
stochastic optimization to generate modelsand then executing an action that will rule outmany of these models.
4
3/1/2007 19
HM: Existing Work
“Diagnosis with Behavioral Modes” Johan de Kleer, Brian C. Williams, IJCAI-89
Introduces Sherlock Sherlock hunts for possible defective candidates It rules out which systems cannot be defective by
logic It determines the best measurement to take based
on calculated probabilities
3/1/2007 20
HM: Existing Work
“Hidden semi-Markov model-basedmethodology for multi-sensor equipmenthealth diagnosis and prognosis” Ming Dong, David He, European Journal of
Operational Research 2006
Looks very complicated!!!
3/1/2007 21
HM: Existing Work
“Probabilistic Models for Monitoring and FaultDiagnosis: Application and Evaluation in a MobileRobot” Joaquin Fernandez, Rafael Sanz and Amador
Dieguez Applied AI[2004]
Uses a set of monitors to detect significant differencesbetween perceived and expected states
Presents hybrid decision systems which make use ofheuristics
Some of these are simple to implement (such as awatchdog)
3/1/2007 22
HM: Existing Work
Work in civil/aerospace engineering onstructural Damage: “Introduction to structural health monitoring”
Charles Farrar and Keith Worden [Dec 2006] Gives a historical overview States Pattern Recognition Paradigm Discusses challenges for SHM
“Bayesian Probabilistic Approach to StructuralHealth Monitoring” M.W Vanik, J. L. Beck and S. K. Au [Journal of
Engineering Mechanics 2000]
3/1/2007 23
Plan
1. Problem Presentation1. Self-calibration2. Health Monitoring
2. Methods and Tools for Self-calibration3. Existing work on Health Monitoring4. Conclusion
1. Research Objectives2. Evaluation3. Value
3/1/2007 24
Research objectives
Obtaining Vulcan’s parameters automatically Monitoring the different devices to update the
model parameter values and detect defects
5
3/1/2007 25
Wheelchair Architecture
Wheel chairbase
Stereo pairLaserRangeFinder
Wheels
x
yz
3/1/2007 26
Parameters to consider Geometric Parameters:
Length, width, height of wheelchair Wheel positions and orientations.
Extrinsic Parameters of Sensors 6 dof laser range finder pose 6 dof stereo cameras pose
Dynamics: Acceleration/deceleration Moments of inertia
Odometry: δrot and δtrans
Other parameters?
3/1/2007 27
Evaluation criteria Models of Vulcan’s sensors and actuators shall be
developed They shall incorporate the defined parameters Simulations shall be done under Gazebo (or
USARSIM) Estimation of parameters shall be compared to
ground truth in Gazebo experiments Tests shall include navigating in circle and
comparing estimated 6dof poses of cameras andlaser range finder to ground truth
Breakdown scenarios shall be designed and reactionto these shall be evaluated in terms of accuracy andspeed.
3/1/2007 28
Value of this project
Self-Calibration Determines system capabilities Removes non random errors Allows coping with changing/non-stable configurations:
Wear and tear Different terrain types…
Health monitoring Allows ignoring data from a defective/inefficient sensor
i.e. when navigating in smoke, dirty lenses… Helps maintenance (i.e. motor still working, but
degraded and needs to be replaced) Helps evaluating if wheelchair is still capable of
control/autonomy.
3/1/2007 29
Value of this project (2)
Precise control requires accurate calibration(going through a door…)
Operator is not capable of detecting if thewheelchair behaves spuriously unless: wrongroom, saccades, slow motion, halt oraccident!!
3/1/2007 30
Further Research
If time permits: Relationship between internal representation
of lengths and meters (or inches) Visual Odometry Calibration using optical flow when robot is
moving straight forward (feasibility TBD)
Even further: Prognosis