9. Related my works
• Reliability estimation of vehicle localization result, In IEEE IV, 2018.
• Mobile robot localization considering uncertainty of sensor observations, In
IEEE/RSJ IROS, 2018.
• Misalignment recognition using Markov random fields with fully connected latent
variables for detecting localization failures, In IEEE RA-L with IROS, 2019.
• Hybrid localization of model- and learning-based methods: Fusion of Monte Carlo
and E2E localizations via importance sampling, In IEEE ICRA, 2020.
• Mobile robot localization considering uncertainty of depth regression from camera
images, In IEEE RA-L 2022 with ICRA.
• Reliable Monte Carlo localization for mobile robots, arXiv, 2022 (preprint).
10. Topics and references of this presentation
Topics
• A general localization model and its limitation [Thrun+ ’05]
• A new method to realize “reliable localization” [Akai+ ’22]
References
• An arXiv preprint https://arxiv.org/abs/2205.04769
• A ROS package for 2D LiDAR-based localization
als_ros: https://github.com/NaokiAkai/als_ros
11. als_ros vs. amcl (experiments with wrong initial pose)
als_ros outperformed amcl
12. als_ros vs. amcl (experiments with wrong initial pose)
als_ros outperformed amcl
14. A general localization model
A graphical model for general localization
• Estimate posterior over the robot pose
x : Robot pose (latent)
u : Control input
z : Sensor measurement
m : Map
t : Time
Posterior over the robot
pose is estimated using a
recursive Bayes filter
Measurement model
15. Posterior and confidence
What does posterior tell us?
• Uncertainty of estimate can be determined from distribution range
Estimated pose
Uncertainty
Unconfident estimate Confident estimate
Posterior over the pose
• Uncertainty also tells us confidence of estimate
16. Confidence and Reliability
Confidence and reliability are similar but different
• A confident but unreliable case exists ➡ miss convergence
Ground truth Ground truth
Reliable but unconfident Unreliable but confident
Need a new approach to know reliability of estimate
• In practical cases, reliability wants to be known
17. A new approach to reliable localization
N. Akai “Reliable Monte Carlo localization for mobile robots,” arXiv, 2022.
18. Localization correctness classifier
A classifier to distinguish localization correctness is required
• Threshold- and learning-based methods can be used
➢ Set a threshold to mean absolute error on residual errors
➢ Learn misalignment using machine learning algorithms
Using the classifier is a simple solution but inaccurate
• The classifier of course makes miss classification
• How to estimate reliability from the classifier?
Propose a new probabilistic localization model
19. A new localization model [Akai+ IV’18]
A localization model to estimate reliability
• Decision by classifier d is introduced as an observable variable
➢ Assume the classifier works as a sensor to measure localization correctness
• State s ∈ {success, failure} is introduced as a latent variable
➢ p(s = success) indicates reliability of estimate
x : Robot pose (latent)
u : Control input
z : Sensor measurement
m : Map
t : Time
d : Decision by classifier
s : Localization state (latent)
20. A new localization model [Akai+ IV’18]
Intuitive understanding
• The general localization model
位置
制御入力
センサ観測
地図
Robot pose
Control
input
Map
Sensor
measurements
21. A new localization model [Akai+ IV’18]
Intuitive understanding
• A roll of the decision in the model
学習器の出力
(正誤判断)
位置
制御入力
センサ観測
地図
Robot pose
Control
input
Map
Sensor
measurements
Decision by classifier
Do not use its output
directly to distinguish
localization correctness
22. A new localization model [Akai+ IV’18]
Intuitive understanding
• Relationship between the decision and the localization state
学習器の出力
(正誤判断)
位置
制御入力
センサ観測
地図
自己位置
推定状態
xt
m
xt-1
zt-1
ut-1
zt
ut
dt
dt-1 zt
st-1 st
Localization state
Robot pose
Control
input
Map
Sensor
measurements Estimate the localization
state based on the decision
with Bayes filter
23. A new localization model [Akai+ IV’18]
Estimate joint posterior over the robot pose and localization state
• This distribution is estimate using Rao-Blackwellized particle filter
Distribution over
the robot pose
Distribution over the
localization state
24. A new localization model [Akai+ IV’18]
Comparison of the posteriors over the pose
• A new likelihood distribution is derived in the new model
Distribution over
the decision
➡ Decision model
25. Decision model
Enabling uncertainty handling of the classifier
• The decision model represents statistical performance of the classifier
• A likelihood calculation example with the threshold-based classifier
28. Further extension in als_ros [Akai+ ’22]
Three characteristic functions of als_ros
• Reliability estimation [Akai+ IV’18]
• Robust localization [Akai+ IROS’18]
• Quick recovery from failure [Akai+ ICRA’20]
x : Robot pose (latent)
u : Control input
z : Sensor measurement
m : Map
t : Time
d : Decision by classifier
s : Localization state (latent)
c : Sensor measurement
classes (latent)
A graphical model used in als_ros
29. Robust localization
als_ros can work in highly dynamic environments
• Known and unknown obstacles are estimated while doing localization
31. Reliable localization
Desired requirements
• Robustly work in dynamic environments
• Immediately detect localization failures
• Quickly recover if failed
als_ros achieves them!
Next target
• Extension of the handheld localization device
• Applying localization technologies to besides
robotic applications (e.g., building management)
32. Summary
Probabilistic approach to reliable localization
• Propose the new graphical model that simultaneously estimates a
robot pose, sensor measurement classes, and localization correctness
• The new model simultaneously achieves
➢ Robust localization in dynamic environments
➢ Immediate failure detection based on estimated reliability
➢ Quick re-localization from failure state
arXiv preprint https://arxiv.org/abs/2205.04769
als_ros: https://github.com/NaokiAkai/als_ros