Page 1
Learning from the Sky:Robot-Aided Mapping, Radio Access
and Localization
WiLab-Huawei Workshop Jan. 2021
David Gesbert
EURECOM, Sophia-Antipolis, France
Collaboration with Omid Esrafilian@EURECOM,
Rajeev Gangula@EURECOM, Junting Chen@USC, U. Mitra@USC
Page 2
September 2017: Google “loon” en route to Puerto Rico
UAV-aided Networks
Page 3
UAV-aided networks: Use cases for micro-drones
Hot-spots, sport
events, flashcrowds Range extension
Disaster recovery
IoT data harvesting, smart
city, agriculture, caching
D2D connectivity: internet connectivity,
car2car connectivity for assisted driving, mesh
connectivity, battlefield connectivity,…
Page 4
Reliable robotic UAV placement
-
Initial UAV position
Optimal position
Learning trajectory
Optimal path
Can integrate local/instantaneous features (vs. probabilistic placement)
Relying on radio sensing capabilities Position vs. path planning
On-line vs off-line
Finite user population vs. fluid models
Obstacle avoidance & navigation
Page 5
Probabilistic vs map-based prediction
Probabilistic (LoS) prediction
Ex: LoS probability model
11 January 2021 5
[Hourani14] A. Al-Hourani, S.
Kandeepan, and S. Lardner, “Optimal
LAP Altitude for Maximum
Coverage”, IEEE Comm. Lett., 2014.
Map-based prediction
Page 6
Segmented Channel Path Loss Models
Class s=1,2,3,..
Av. RX power
Path loss exponent
shadowing
Fixed offset
distance
Received power map UAV 100
meter above center of Bristol
Page 7
ML-based Radio Map Reconstruction
APPROACH 1:
Classical ML
(KNN applied to
RSSI-domain image)
APPROACH 2:
Model-based ML
(Segment Classification
Followed by KNN )
[J. Chen & D. Gesbert, 2017]
Page 8
Expert knowledge is important!
K nearest neighbors (KNN) with Kernel:
KNN applied to RSSI image
(no channel model)
KNN applied to
model-classified data
(hard/soft reconstruction)
Question: How to scale with #users?
Page 9
3D Reconstruction
1km
Reconstructed map with 1500 users*UAV trajectory
800 m
*K=32 UAV locations, spatial smoothing applied
Optimized flying altitude in closed form (Globecom 17)
Page 10
Joint 3D and radio map reconstruction
10
An orthoimagery of
an area at center
Washington DC, USA
[Chen, Esrafilian, Gesbert, Mitra, Robotics, Science and Systems, MIT, 2017]
Radio map
reconstruction
Radio map estimate
3D map estimate
Joint approach
Page 11
Trajectory design use cases
Scenario 1: UAV as cellular relay
Scenario 2: “Smart” IoT data harvesting
Scenario 3: UAV-aided mesh connectivity
Map information can be too much information !
Single user-drone radio map
Page 12
UAV as an autonomous cellular relay (live demo @ 2.5GHz)
https://www.youtube.com/watch?v=GI_lOsg_qmQ
UAV (EURECOM)
User (off-the-shelf phone)Base Station (EURECOM)
Page 13
Scenario 2: IoT “data harvesting”
[Esrafilian, Gangula, Gesbert, IEEE Journal IoT, 2019]
Problem: Find path and schedule which harvests the “most
data” from ground nodes, under fixed flying time
Assumes map knowledge
But problem not differentiable
Page 14
“Map compression”: Local probabilistic model
-> to make problem differentiable
-
• User
Global LOS Probability model:
Local map-aided LOS Probability:
Page 15
Map-based Trajectory design & User Scheduling
-
Shadowing
Fixed offset
Path loss exponent
Page 16
Scenario 2: IoT data harvesting
Page 17
Scenario 3: UAV-aided Mesh connectivity
- - p 17
•Previous work: Morgenthaler, Wi-UAV 2012, Yanmaz., et al WCNC 2014, PIMRC 2015, etc.
•However, no optimal UAV placement!
Page 18
Scenario 3: UAV-aided Mesh connectivity
- - p 18
• Challenge: Optimal placement depends on routing algorithm (OLSR..)
• Two Phase approach:
• Clustering of nodes
• UAV placement to optimize inter-cluster connectivity.
• Placement:
Where:Average path-loss Transmit power
Number of nodes in cluster k
• This problem is again solved with map compression -> SCA can be used!
Page 19
Scenario 3: UAV-aided Mesh connectivity
- - p 19
Page 20
Outline
UAV placement and path design
Channel prediction
Learning maps
Communication trajectory design
Trajectories design with Active Learning
Perspectives
Page 21
Active Learning #1: Learning 3D building Maps
NMSE = 0.2488
Arbitrary Paths
Dynamic Programming-Optimized Path
NMSE = 0.343 (averagr across arbitray paths)
Refinement
Graph
[Esrafilian, Gesbert, 2017]
Page 22
Active Learning #2: Learning the channel
Goal: design a flight path to estimate channel
parameters with minimal error variance -> DP
[Esrafilian, Gangula, Gesbert, sub IEEE Journal IoT, 2018]
Page 23
Active Learning #3: User localization
• Flight goal:
• Collect RSSI measurements from K users
• To learn the channel parameters and localize the users
• Leveraging the 3D map
• User localization:
• PSO (particle swarm opt)
• Trajectory design:
• Active Learning based on
Fisher Information matrix.
[O. Esrafilian, R. Gangula, D. Gesbert, "3D Map-based Trajectory Design in UAV-aided Wireless Localization
Systems", submitted to the IEEE Internet of Things Journal, April 2020]
Page 24
Active Learning #3: User localization
Page 25
Perspectives
Coordination across multiple UAVs
Low complexity algorithms (DP is complex!)
Onboard RF/antenna design
Onboard vs. offboard computing
Advanced mix robotics-communications models
Fusion of radio and vision/LIDAR data ([Esra Gang Gesb 2021])
All references under www.eurecom.fr/cm/gesbert
Plenty of hard problems (theoretical/experimental)
UAV-aided networks: Promising technology Makes the network flexible, closer to the end user
Comm-robotics interactions (mapping..)
Side benefits: User localization, reliable flying terminal,..