IMULet: A Cloudlet for Inertial Tracking
Lauri Tuominen and Mo Alloulah
25 Feb 2021
1 © Dania Alloulah
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
person
chair
chair
robot
drone
bike
person
table
Vision
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
Active RF 802.11mc


person
chair
chair
robot
drone
bike
person
table
Vision
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
Active RF 802.11mc


person
chair
chair
robot
drone
bike
person
table
Vision
Passive RF-Pose3D
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
Active RF 802.11mc


person
chair
chair
robot
drone
bike
person
table
Vision
Passive RF-Pose3D
…
Multi-modal:
SurroundSense
Navigation/Localisation is a fundamental technology
Cornerstone of the physical internet
Active RF 802.11mc


person
chair
chair
robot
drone
bike
person
table
Vision
Passive RF-Pose3D
…
Name of game: Robustification, robustification, robustification
The future of computing is spatial


Name of game: Robustification, robustification, robustification
The future of computing is spatial


multi-modal
Name of game: Robustification, robustification, robustification
The future of computing is spatial


multi-modal, multi-layered
Name of game: Robustification, robustification, robustification
The future of computing is spatial


multi-modal, multi-layered, and semantic
Name of game: Robustification, robustification, robustification
The future of computing is spatial


multi-modal, multi-layered, and semantic
Lest …
Name of game: Robustification, robustification, robustification
The future of computing is spatial


multi-modal, multi-layered, and semantic
Lest …
Name of game: Robustification, robustification, robustification
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
Ground 0


• independent of environmental dynamics


• i.e. very useful
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
Ground 0


• independent of environmental dynamics


• i.e. very useful
T
r
i
c
k
s
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
Ground 0


• independent of environmental dynamics


• i.e. very useful
T
r
i
c
k
s
DL
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
Ground 0


• independent of environmental dynamics


• i.e. very useful
Comparison between DL (IONet),


pedestrian dead reckoning (PDR), and


traditional strapdown inertial navigation


systems (SINS)


Why Deep Learning?
T
r
i
c
k
s
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
Ground 0


• independent of environmental dynamics


• i.e. very useful
T
r
i
c
k
s
Comparison between DL (IONet),


pedestrian dead reckoning (PDR), and


traditional strapdown inertial navigation


systems (SINS)


Chen et al. "Ionet: Learning to cure the curse of drift in inertial odometry."
arXiv preprint arXiv:1802.02209 (2018).
Why Deep Learning?
T
r
i
c
k
s
What do we need to do?
Inertial
Radar Lidar
Vision
Acoustic Infrared
…
… …
…
Ground 0


• independent of environmental dynamics


• i.e. very useful
T
r
i
c
k
s
Comparison between DL (IONet),


pedestrian dead reckoning (PDR), and


traditional strapdown inertial navigation


systems (SINS)


Chen et al. "Ionet: Learning to cure the curse of drift in inertial odometry."
arXiv preprint arXiv:1802.02209 (2018).
Why Deep Learning?
T
r
i
c
k
s
Major improvements over classical methods
DL, is that it?
DL, is that it?


System-level innovations needed for the real-world
DL, is that it?


System-level innovations needed for the real-world


• ultra-low power tags


• compute-, energy-, and form factor-constrained
DL, is that it?


System-level innovations needed for the real-world


• ultra-low power tags


• compute-, energy-, and form factor-constrained


• deal with DL model fragility


• signal characteristic variabilities in the field


• i.e. hard to generalise
DL, is that it?


System-level innovations needed for the real-world


• ultra-low power tags


• compute-, energy-, and form factor-constrained


• deal with DL model fragility


• signal characteristic variabilities in the field


• i.e. hard to generalise


• configurability and scalability


• use cases/market segments
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
(1) ultra-low power
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
(1) ultra-low power
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
(2) configurable
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
(2) configurable
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
(3) customisable D1 D2 D3 D4 D5 D6
D1 -91.3 -38.3 -37.3 -33 -34.6 -31
D2 -38.3 -91.6 -38.3 -32.6 -36.3 -30.8
D3 -37.3 -38.3 -91.2 -32.4 -35.8 -30.8
D4 -33 -32.6 -32.4 -91.8 -29.8 -28.9
D5 -34.6 -36.3 -35.8 -29.8 -91.8 -30.6
D6 -31 -30.8 -30.8 -28.9 -30.6 -91.2
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
(3) customisable D1 D2 D3 D4 D5 D6
D1 -91.3 -38.3 -37.3 -33 -34.6 -31
D2 -38.3 -91.6 -38.3 -32.6 -36.3 -30.8
D3 -37.3 -38.3 -91.2 -32.4 -35.8 -30.8
D4 -33 -32.6 -32.4 -91.8 -29.8 -28.9
D5 -34.6 -36.3 -35.8 -29.8 -91.8 -30.6
D6 -31 -30.8 -30.8 -28.9 -30.6 -91.2
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
Promising


early performance
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
Promising


early performance
IMULet
LSTM stack
states
estimates delayline
Edge Cloud
raw 9-DOF IMU signals
acceI
gyro
magneto
Displacement
Estimate
On-device
compressed
representation
FC
FC
Adaptation
z K
layer 1
CNN
layer 0
Latent space tracking
s
Promising


early performance
Check out paper for details
In Summary


IMULet is a cloudlet for inertial tracking


• DL + edge + hooks for efficiency and generalisability
In Summary


IMULet is a cloudlet for inertial tracking


• DL + edge + hooks for efficiency and generalisability


Key component towards infrastructure-less localisation


• as mandated by physical internet/spatial computing
In Summary


IMULet is a cloudlet for inertial tracking


• DL + edge + hooks for efficiency and generalisability


Key component towards infrastructure-less localisation


• as mandated by physical internet/spatial computing


Scaling evaluation in-progress
Cheers
45 © Dania Alloulah

IMULet: A Cloudlet for Inertial Tracking

  • 1.
    IMULet: A Cloudletfor Inertial Tracking Lauri Tuominen and Mo Alloulah 25 Feb 2021 1 © Dania Alloulah
  • 2.
    Navigation/Localisation is afundamental technology Cornerstone of the physical internet
  • 3.
    Navigation/Localisation is afundamental technology Cornerstone of the physical internet
  • 4.
    Navigation/Localisation is afundamental technology Cornerstone of the physical internet person chair chair robot drone bike person table Vision
  • 5.
    Navigation/Localisation is afundamental technology Cornerstone of the physical internet Active RF 802.11mc person chair chair robot drone bike person table Vision
  • 6.
    Navigation/Localisation is afundamental technology Cornerstone of the physical internet Active RF 802.11mc person chair chair robot drone bike person table Vision Passive RF-Pose3D
  • 7.
    Navigation/Localisation is afundamental technology Cornerstone of the physical internet Active RF 802.11mc person chair chair robot drone bike person table Vision Passive RF-Pose3D …
  • 8.
    Multi-modal: SurroundSense Navigation/Localisation is afundamental technology Cornerstone of the physical internet Active RF 802.11mc person chair chair robot drone bike person table Vision Passive RF-Pose3D …
  • 9.
    Name of game:Robustification, robustification, robustification
  • 10.
    The future ofcomputing is spatial Name of game: Robustification, robustification, robustification
  • 11.
    The future ofcomputing is spatial multi-modal Name of game: Robustification, robustification, robustification
  • 12.
    The future ofcomputing is spatial multi-modal, multi-layered Name of game: Robustification, robustification, robustification
  • 13.
    The future ofcomputing is spatial multi-modal, multi-layered, and semantic Name of game: Robustification, robustification, robustification
  • 14.
    The future ofcomputing is spatial multi-modal, multi-layered, and semantic Lest … Name of game: Robustification, robustification, robustification
  • 15.
    The future ofcomputing is spatial multi-modal, multi-layered, and semantic Lest … Name of game: Robustification, robustification, robustification
  • 16.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … …
  • 17.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … …
  • 18.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … … Ground 0 • independent of environmental dynamics • i.e. very useful
  • 19.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … … Ground 0 • independent of environmental dynamics • i.e. very useful T r i c k s
  • 20.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … … Ground 0 • independent of environmental dynamics • i.e. very useful T r i c k s DL
  • 21.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … … Ground 0 • independent of environmental dynamics • i.e. very useful Comparison between DL (IONet), pedestrian dead reckoning (PDR), and traditional strapdown inertial navigation systems (SINS) Why Deep Learning? T r i c k s
  • 22.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … … Ground 0 • independent of environmental dynamics • i.e. very useful T r i c k s Comparison between DL (IONet), pedestrian dead reckoning (PDR), and traditional strapdown inertial navigation systems (SINS) Chen et al. "Ionet: Learning to cure the curse of drift in inertial odometry." arXiv preprint arXiv:1802.02209 (2018). Why Deep Learning? T r i c k s
  • 23.
    What do weneed to do? Inertial Radar Lidar Vision Acoustic Infrared … … … … Ground 0 • independent of environmental dynamics • i.e. very useful T r i c k s Comparison between DL (IONet), pedestrian dead reckoning (PDR), and traditional strapdown inertial navigation systems (SINS) Chen et al. "Ionet: Learning to cure the curse of drift in inertial odometry." arXiv preprint arXiv:1802.02209 (2018). Why Deep Learning? T r i c k s Major improvements over classical methods
  • 24.
  • 25.
    DL, is thatit? System-level innovations needed for the real-world
  • 26.
    DL, is thatit? System-level innovations needed for the real-world • ultra-low power tags • compute-, energy-, and form factor-constrained
  • 27.
    DL, is thatit? System-level innovations needed for the real-world • ultra-low power tags • compute-, energy-, and form factor-constrained • deal with DL model fragility • signal characteristic variabilities in the field • i.e. hard to generalise
  • 28.
    DL, is thatit? System-level innovations needed for the real-world • ultra-low power tags • compute-, energy-, and form factor-constrained • deal with DL model fragility • signal characteristic variabilities in the field • i.e. hard to generalise • configurability and scalability • use cases/market segments
  • 29.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s
  • 30.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s
  • 31.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s
  • 32.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s
  • 33.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s (1) ultra-low power
  • 34.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s (1) ultra-low power
  • 35.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s (2) configurable
  • 36.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s (2) configurable
  • 37.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s (3) customisable D1 D2 D3 D4 D5 D6 D1 -91.3 -38.3 -37.3 -33 -34.6 -31 D2 -38.3 -91.6 -38.3 -32.6 -36.3 -30.8 D3 -37.3 -38.3 -91.2 -32.4 -35.8 -30.8 D4 -33 -32.6 -32.4 -91.8 -29.8 -28.9 D5 -34.6 -36.3 -35.8 -29.8 -91.8 -30.6 D6 -31 -30.8 -30.8 -28.9 -30.6 -91.2
  • 38.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s (3) customisable D1 D2 D3 D4 D5 D6 D1 -91.3 -38.3 -37.3 -33 -34.6 -31 D2 -38.3 -91.6 -38.3 -32.6 -36.3 -30.8 D3 -37.3 -38.3 -91.2 -32.4 -35.8 -30.8 D4 -33 -32.6 -32.4 -91.8 -29.8 -28.9 D5 -34.6 -36.3 -35.8 -29.8 -91.8 -30.6 D6 -31 -30.8 -30.8 -28.9 -30.6 -91.2
  • 39.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s Promising early performance
  • 40.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s Promising early performance
  • 41.
    IMULet LSTM stack states estimates delayline EdgeCloud raw 9-DOF IMU signals acceI gyro magneto Displacement Estimate On-device compressed representation FC FC Adaptation z K layer 1 CNN layer 0 Latent space tracking s Promising early performance Check out paper for details
  • 42.
    In Summary IMULet isa cloudlet for inertial tracking • DL + edge + hooks for efficiency and generalisability
  • 43.
    In Summary IMULet isa cloudlet for inertial tracking • DL + edge + hooks for efficiency and generalisability Key component towards infrastructure-less localisation • as mandated by physical internet/spatial computing
  • 44.
    In Summary IMULet isa cloudlet for inertial tracking • DL + edge + hooks for efficiency and generalisability Key component towards infrastructure-less localisation • as mandated by physical internet/spatial computing Scaling evaluation in-progress
  • 45.