City College of New York
1
Dr. Jizhong XiaoDepartment of Electrical Engineering
CUNY City [email protected]
Advanced Mobile Robotics
City College of New York
Probabilistic Robotics
Probabilistic Sensor Models
Beam-based ModelLikelihood Fields ModelFeature-based Model
City College of New York
• Prediction (Action)
• Correction (Measurement)
Bayes Filter Reminder
111 )(),|()( tttttt dxxbelxuxpxbel
)()|()( tttt xbelxzpxbel
City College of New York
4
Sensors for Mobile Robots
• Contact sensors: Bumpers
• Internal sensors– Accelerometers (spring-mounted masses)– Gyroscopes (spinning mass, laser light)– Compasses, inclinometers (earth magnetic field, gravity)
• Proximity sensors– Sonar (time of flight)– Radar (phase and frequency)– Laser range-finders (triangulation, tof, phase)– Infrared (intensity)
• Visual sensors: Cameras
• Satellite-based sensors: GPS
City College of New York
5
Proximity Sensors
• The central task is to determine P(z|x), i.e., the probability of a measurement z given that the robot is at position x.
• Question: Where do the probabilities come from?• Approach: Let’s try to explain a measurement.
City College of New York
6
Noise IssuesSensors are imperfect!
Specular Reflection
City College of New York
7
Beam-based Sensor Model• Scan z consists of K measurements.
• Individual measurements are independent given the robot position and the map of environment.
},...,,{ 21 Kzzzz
K
kk mxzPmxzP
1
),|(),|(
A cyclic array of k ultrasound sensors
City College of New York
8
Beam-based Sensor Model
K
kk mxzPmxzP
1
),|(),|(
City College of New York
9
Typical Measurement Errors of an Range Measurements
1. Beams reflected by obstacles
2. Beams reflected by persons / caused by crosstalk
3. Random measurements4. Maximum range
measurements
City College of New York
10
Proximity Measurement
• Measurement can be caused by …– a known obstacle.– cross-talk.– an unexpected obstacle (people, furniture, …).– missing all obstacles (total reflection, glass, …).
• Noise is due to uncertainty …– in measuring distance to known obstacle.– in position of known obstacles.– in position of additional obstacles.– whether obstacle is missed.
City College of New York
11
Beam-based Proximity ModelMeasurement noise
zexp zmax0
2
2exp )(
21
22
1),|(
zz
hithit emxzP
otherwisezz
mxzPz
0e
),|( expunexp
Unexpected obstacles
zexp zmax0
Gaussian Distribution
Exponential Distribution
Likelihood of sensing unexpected objects decreased with range
City College of New York
12
Beam-based Proximity ModelRandom measurement Max range/Failures
max
1),|(z
mxzPrand
otherwise
zzifmxzP
01
),|( maxmax
zexp zmax0zexp zmax0
Uniform distribution
Point-mass distribution
If fail to detect obstacle, report maximum range
City College of New York
13
Resulting Mixture Density
),|(),|(),|(
),|(
),|(
rand
max
unexp
hit
rand
max
unexp
hit
mxzPmxzPmxzPmxzP
mxzP
T
How can we determine the model parameters?
Weighted average, and 1maxexp randunhit
System identification method: maximum likelihood estimator (ML estimator)
},,,,,{ maxexp randunhit
City College of New York
14
Raw Sensor DataMeasured distances for expected distance of 300 cm.
Sonar Laser
City College of New York
15
Approximation
• The likelihood of the data Z is given by
• Our goal is to identify intrinsic parameters that maximize this log likelihood– Mathematic derivation can be found pp163~167– Psudo-code can be found in pp160
• An iterative procedure for estimating parameters
• Deterministically compute the n-th parameter to satisfy normalization constraint.
},,,,,{ maxexp randunhit
City College of New York
16
Approximation Results
Sonar
300cm 400cm
The smaller range, the more accurate the measurement
Laser is more accurate than sonar, narrower Gaussians
Laser
City College of New York
17
Example
z P(z|x,m)
Laser scan, projected into a previously acquired map
The likelihood, evaluated for all positions and projected into the map. The darker a position, the larger
P(z|x,m)
City College of New York
18
Summary Beam-based Model
• Assumes independence between beams.– Justification?– Overconfident! Suboptimal result
• Models physical causes for measurements.– Mixture of densities for these causes.– Assumes independence between causes. Problem?
• Implementation– Learn parameters based on real data.– Different models should be learned for different angles at which
the sensor beam hits the obstacle.– Determine expected distances by ray-tracing.– Expected distances can be pre-processed.
City College of New York
19
Likelihood Fields Model
• Beam-based model is …– not smooth for small obstacles and at edges
(small changes of a robot’s pose can have a tremendous impact on the correct range of sensor beam, heading direction is particularly affected).
– not very efficient (computationally expensive in ray casting)
• Idea: Instead of following along the beam, just check the end point.
City College of New York
20
Likelihood Fields Model• Project the end points of a sensor scan Zt into the global
coordinate space of the map
• Probability is a mixture of …– a Gaussian distribution with mean at distance to closest
obstacle,– a uniform distribution for random measurements, and – a small uniform distribution for max range measurements.
• Again, independence between different components is assumed.
randrandhithittkt pppmxzp maxmax),(
)sin()cos(
cossinsincos
ksens
ksenskt
ksens
ksens
z
z zyx
yx
yx
kt
kt
City College of New York
21
Likelihood Fields Model
)sin()cos(
cossinsincos
ksens
ksenskt
ksens
ksens
z
z zyx
yx
yx
kt
kt
Distance to the nearest obstacles. Max range reading ignored
City College of New York
22
Example
P(z|x,m)
Example environment Likelihood fieldThe darker a location, the less likely it is to perceive an obstacleSensor
probability
O1
O2
O3
Oi : Nearest point to obstaclesZmax
City College of New York
23
San Jose Tech Museum
Occupancy grid map Likelihood field
City College of New York
24
Scan Matching
• Extract likelihood field from scan and use it to match different scan.
Sensor scan, 180 dots
The darker a region, the smaller likelihood for sensing an object there
City College of New York
25
Properties of Likelihood Fields Model
• Highly efficient, uses 2D tables only.• Smooth w.r.t. to small changes in robot position.
• Allows gradient descent, scan matching.
• Ignores physical properties of beams.
• Will it work for ultrasound sensors?
City College of New York
26
Additional Models of Proximity Sensors
• Map matching (sonar,laser): generate small, local maps from sensor data and match local maps against global model. (e.g., transform scans into occupancy maps)
• Scan matching (laser): map is represented by scan endpoints, match scan into this map. (e.g. ICP algorithm to stitch scans together)
• Features (sonar, laser, vision): Extract features such as doors, hallways from sensor data.
City College of New York
27
Features/Landmarks
• Active beacons (e.g., radio, GPS)• Passive (e.g., visual, retro-reflective)• Standard approach is triangulation
• Sensor provides– distance, or– bearing, or– distance and bearing.
City College of New York
28
Distance and Bearing
City College of New York
29
Feature-Based Measurement Model• Feature vector is abstracted from the measurement:
• Sensors that measure range, bearing, & a signature (a numerical value, e.g., an average color)
• Conditional independence between features
• Feature-Based map: withi.e., a location coordinate in global coordinates & a signature • Robot pose:
• Measurement model:
Zero-mean Gaussian error variables with standard deviations
},,{}{)(2
2
2
1
1
1
21
t
t
t
t
t
t
ttt
s
r
s
rffzf
),,,(),)(( mxsrpmxzfp tit
i
it
ittt
},,{ 21 mmm
2
2
2
)),(2tan)()(
,,
2,
2,
s
r
j
xjyj
yjxj
it
it
it
sxmymaymxm
s
r
Tt yxx }{
Tjyjxjj smmm },,{ ,,
2r
r
City College of New York
30
Sensor Model with Known Correspondence• A key problem for range/bearing sensors: data association, (arise
when the landmarks cannot be uniquely identified)• Introduce a correspondence variable , between feature
and landmark . If , then the i-th feature observed at time t corresponds to the j-th landmark in the map
Algorithm for calculating the probability of a landmark measurement: inputs: feature , with know correspondence , the robot pose and the map, Its output is the numerical probability
itf jm
}1,,1{ Nc it Njc it
itf Nc it
),,( mxcfp tit
it
City College of New York
31
Summary of Sensor Models• Explicitly modeling uncertainty in sensing is key to robustness.• In many cases, good models can be found by the following approach:
1. Determine parametric model of noise free measurement.2. Analyze sources of noise.3. Add adequate noise to parameters (eventually mix in densities for
noise).4. Learn (and verify) parameters by fitting model to data.5. Likelihood of measurement is given by “probabilistically comparing” the
actual with the expected measurement.• This holds for motion models as well.• It is extremely important to be aware of the underlying assumptions!
City College of New York
32
Thanks
Homework 4, 5 posted