IEEE Consumer Electronics Magazine - July/August 2022 - 71

Table 2. ML Regression Model Performance With Data
Imputation.
ML
Model
LR
SVR
Configuration
Linear
Polynomial
Kernel
MLP 200, 200, 200
Figure 4. Warehouse aisle for experimental
testbed.
position increases as the agent gains no information.
While during the update step, the agent
gains information through the sensors' measurements
and becomes more confident regarding
its state.
KF considers the position probability of the
robot to be Gaussian probability density function
(PDF), which can be characterized using the mean
and the standard deviation. The process starts by
initializing the belief of the robot at the start time.
For this, we take in our initial prediction output
from theMLmodel. Next, using amotionmodel for
the system, which is a constant velocity model in
our design, we estimate the robot's location in the
next time step. During the update-cycle of KF, after
the time step has elapsed, we obtain the measurement
reading from the ML model and update our
previously predicted belief of the robot's state. It
is to note that the robot's position, i.e., the state
and the measurements, is modeled as Gaussian
PDF. The tuning of R and Q parameters is done
basedon the sensor.
EXPERIMENTAL EVALUATION
In this section, we provide a detailed evaluation
of the proposed KF localization framework.
For the experimental testbed, we select a working
warehouse22 as shown in Figure 4. Within the
two aisles of the warehouse, we have mounted
ten 60 GHz routers on the ceiling configured as
APs. The placement of the routers is done in a
zig-zag manner. On the robot, we place a 60 GHz
router configured as Client. For APs and Client,
we have used TP-Link AD7200 consumer grade
routers. For the data collection, we collect the
July/August 2022
Augmentation
With
(0.51 m, 2.6
m)
(0.71 m,
2.17 m)
(0.15 m,
0.91 m)
Without
(0.58 m,
2.84 m)
(0.48 m,
2.33 m)
(0.27 m,
1.54 m)
SNR data from all the available APs. We collect
the training data within the two aisles by moving
the robot 0.609 m (2-feet) in the y-dimension of
the aisle and keeping the x-dimension fixed. For
both the aisles, we collect two different training
datasets. For test-time evaluation, another separate
held-out dataset consisting of 674 test samples
is collected. We assume the robot moves in
one dimension, which is the y-dimension for
localization and tracking. This is realistic in
many practical warehouses for autonomous
agents, where only one agent is allowed to move
in a single-aisle at any given time.3
Machine Learning Model Analysis
The optimized ML localization model is a
multilayer neural network with three hidden
layers consisting of 200 neurons. The ML prediction
is a regression model, where the output is
the 2-D prediction of the robot's position inside
the warehouse. For model selection, we train
and evaluate different regression models of varying
complexity. The three models that we experiment
with are linear regression (LR), multilayer
perceptron (MLP), and support vector regression
(SVR). These ML models are considered as
they represent varying complexity in terms of
computation and learning capability. We train
and optimize each regression model individually
on the training dataset and use the held-out test
dataset to evaluate the performance.
Table 2 presents the performance comparison
between regression models trained with and
without data augmentation. MLP achieves significant
performance improvement, with RMSE of
0.19 m and 0.92 m in x and y location prediction,
respectively. This shows 29.3% and 40.25%
71
IEEE Consumer Electronics Magazine - July/August 2022

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - July/August 2022
