BeamSeek: Deep Learning-based DOA Estimation

for Low-Complexity mmWave Phased Arrays

Arav Sharma1 , Lei Chi1 , Ari Gebhardt1 , Alon S. Levin2 , Timothy R. Hoerning1 and Sam Keene1
1
The Cooper Union for the Advancement of Science and Art, New York City, USA
2
Department of Electrical Engineering, Columbia University, New York City, USA
{arav.sharma, lei.chi, ari.gebhardt, timothy.hoerning, keene}@cooper.edu, alon.s.levin@columbia.edu

Abstract—A novel approach combining agile beam switch-

ing with deep learning to enhance the speed and accuracy
arXiv:2508.13075v1 [eess.SP] 18 Aug 2025

of Direction of Arrival (DOA) estimation for millimeter-wave

(mmWave) phased array systems with low-complexity hardware
implementations is proposed and evaluated. Traditional DOA
methods requiring direct access to individual antenna elements
are impractical for analog or hybrid beamforming systems preva-
lent in modern mmWave implementations. Recent agile beam
switching techniques have demonstrated rapid DOA estimation,
but their accuracy and robustness can be further improved
via deep learning. BeamSeek addresses these limitations by
employing a Multi-Layer Perceptron (MLP) and specialized data
augmentation that emulates real-world propagation conditions.
The proposed approach was experimentally validated at 60 Fig. 1. Digital beamforming with multiple RF chains (a) vs. analog
GHz using the NSF PAWR COSMOS testbed, demonstrating beamforming with a single RF chain (b).
significant improvements over a correlation-based method across
various Signal-to-Noise Ratio (SNR) levels. Results show that
BeamSeek achieves up to an 8◦ reduction in average estimation
error compared to this baseline, with particular advantages in Under these hardware constraints, only the combined output
noisy channels. This makes it especially suitable for practical from the antenna array is accessible, presenting significant
mmWave deployments in environments characterized by multi- challenges for conventional DOA estimation methods. The
path interference and hardware constraints. simplest approach is beam sweeping, where the array se-
Index Terms—Direction of arrival (DOA), mmWave, phased
quentially scans all angular directions and chooses the beam
array, machine learning, neural networks, beamforming
corresponding to the maximum received power [7]. While
I. I NTRODUCTION straightforward to implement, this approach introduces latency
Millimeter-wave (mmWave) technology is a crucial com- proportional to the angular resolution, limiting its applicability
ponent of next-generation wireless communications systems, in dynamic environments.
offering greater bandwidth and data capacity to meet growing An innovative solution to this challenge was introduced by
demands [1]. A significant challenge in mmWave implemen- Gallyas-Sanhueza et al. [6], who developed an agile beam
tations is their susceptibility to substantial path loss and signal switching technique capable of rapidly estimating DOA within
blockage, which demands highly directional beamforming a single OFDM symbol. While effective in ideal conditions,
to ensure reliable communication links [2]. Enabling this this approach may face challenges in high noise and multipath
requires accurate Direction of Arrival (DOA) estimation to scenarios. Additionally, this method relies on a known pilot
optimize beam steering. signal, which potentially limits its applications. The primary
Traditional DOA estimation approaches for phased array contributions of this work include:
systems rely on methods such as MUSIC (MUltiple SIgnal • A novel approach that combines agile beam switching
Classification) [3] and ESPRIT (Estimation of Signal Pa- with Deep Learning (DL) to enhance DOA estimation
rameters via Rotational Invariance Techniques) [4]. These accuracy for hardware-constrained phased arrays, exper-
techniques require direct access to individual antenna ele- imentally validated at 60 GHz on the National Science
ment data, which is only available in arrays with dedicated Foundation (NSF) Platforms for Advanced Wireless Re-
RF chains for each element. However, for many practical search (PAWR) Cloud Enhanced Open Software Defined
mmWave systems, analog beamforming architectures with a Mobile Wireless Testbed for City-Scale Deployment.
single RF chain are often preferred due to their reduced • A data augmentation method informed by real-world
hardware complexity, power consumption, and cost [5]. For propagation behavior to increase the robustness of Beam-
example, a majority of 5G gNodeB installments currently rely Seek and expand the limited training set.
on a limited number of RF chains [6]. The difference in these • A pilot-free correlation-based DOA estimation method to
beamforming architectures is shown in Figure 1. use a baseline for comparison.

This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.
 Fig. 2. System model used to derive the input data, demonstrating spatial modulation of the incident signal through agile beam switching for B = 3.

II. R ELATED W ORK III. M ETHODOLOGY

A. Hierarchical Beam Scanning A. Input Data

Hierarchical beam scanning techniques effectively mitigate We consider a signal incident on a uniform planar array
the latency associated with exhaustive beam scanning by (UPA), originating from a transmitter located at azimuthal
employing progressively narrower beams to systematically angle α. The colatitude of the transmitter (ϕ) is not considered
reduce the search space [8]. However, these approaches rely due to restrictions of the testbed. Let ⃗s = [s1 , s2 , . . . , sN ] rep-
on a multi-stage search process, which introduces latency resent the transmitted signal of length N , where each element
of its own. Additionally, these techniques suffer in rapidly sn ∈ C represents a sampled data point. The magnitude of the
changing environments, such as high-mobility scenarios or transmitted signal is scaled by a constant γ ∈ R, determined
settings with dynamic obstacles, due to the inherent delay in by the power of the transmitter. To model the incident signal,
iterative refinement of the beam code book. we apply a channel coefficient h ∈ C, which represents the
phase shift and attenuation due to propagation through the
medium. The incident signal, ⃗r can then be expressed as
B. Agile Beam Switching ⃗r = γh⃗s = γh[s1 , s2 , ..., sN ]. While receiving the signal, we
Recent advances in phased array technology enable beam rapidly switch the steering angle of the UPA to B equally
switching within microseconds. Gallyas-Sanhueza et al. [6] spaced beams. Due to the short period of observation, we can
utilized agile switching to modulate a received pilot signal assume that γ and h remain constant. The incident signal is
with a spatially varying pattern and correlated it with theo- segmented and rewritten as
retical patterns to estimate the DOA. This approach reduces
latency and requires only one RF chain, though its real- ⃗r = γh[⃗s1 , ⃗s2 , ..., ⃗sB ], (1)
world performance remains understudied. An implementation
where each component s⃗n ∈ CN/B is an N/B length segment
derived from this method that does not rely on a known
of the transmitted signal and is received by a corresponding
pilot serves as a baseline to evaluate the performance of the
beam with a steering angle βn . As a consequence of beam-
approach introduced in this work.
forming, the gain of each s⃗n varies strongly based on values of
α and βn , and can denoted by a coefficient g(α, βn ) ∈ C. By
C. Deep Learning for DOA Estimation rapidly switching β, we modulate ⃗r by a spatially dependent
Deep learning (DL) approaches have shown promising pattern, as show in Figure 2. This modulated signal can be
results in addressing the limitations of traditional DOA es- expressed as
timation methods. Hassan et al. [9] showed that convolutional ⃗y = γh[g(α, β1 )s⃗1 , g(α, β2 )s⃗2 , . . . , g(α, βB )s⃗B ], (2)
neural networks (CNN) for DOA estimation outperform MU-
SIC in low SNR and multipath scenarios. Similarly, Huang et which can be written as ⃗y = [y⃗1 , y⃗2 , . . . , y⃗B ], where each
al. [10] proposed a deep neural network for DOA estimation y⃗n = γhg(α, βn )s⃗n . The power of y⃗n can be calculated as
with improved robustness to calibration errors. However, most
DL approaches have focused on processing data from digital 1 2
P (⃗yn ) = ∥⃗yn ∥ , (3)
architectures with access to individual antenna elements. The N/B
application of DL techniques to hardware-constrained analog
which is equivalent to
beamforming systems with a limited number of RF chains
presents a significant research opportunity. In addition, many P (⃗yn ) = |γhg(α, βn )|2 P (⃗sn ). (4)
proposed DL methods have only been tested in simulation.
Experimentally verifying the DL method in this work on a It is important to note that γ, h, and P (⃗sn ) are constant
real-world testbed provides additional context regarding its during the period of observation. Therefore, P (⃗yn ) is directly
deployability. proportional to |g(α, βn )|2 , which removes the dependence
 C. BeamSeek
1) Neural Network Architecture: Our approach employs
a Multi-Layer Perceptron (MLP) as shown in Figure 3 to
predict angles from power profiles. The network receives
power profiles of dimension B as input and outputs a single
angle prediction. We use mean squared error as our loss
function, which penalizes larger errors more severely than
smaller ones, encouraging high precision in angle predictions.
The MLP architecture consists of a single SwiGLU gated
linear unit (GLU) feed-forward network (FFN) with a hidden
dimension of 384. SwiGLU, or Swish-Gated Linear Unit,
combines the Swish activation function with a gating mech-
anism similar to GLUs [11]. Mathematically, the SwiGLU is
defined as
FFNSwiGLU (x, W, V, W2 ) = (Swish1 (xW ) ⊗ xV )W2 , (8)
where x is the input vector, W , V , and W2 are weight
Fig. 3. Architecture of the BeamSeek SwiGLU neural network for DOA
matrices, ⊗ represents element-wise multiplication, and the
estimation, featuring parallel gated and linear paths that process beam gain Swish activation function is defined as
profiles through SiLU activation and layer normalization before element-wise
multiplication, dropout regularization, and final angle prediction. Swish1 (x) = SiLU(x) = x · sigmoid(x). (9)
This activation approach enables more expressive feature
transformations by enabling the network to modulate informa-
of our DOA estimation methods on phase and by extension a
tion flow through the multiplication of two separate branches
known pilot. Finally, we can define the power profile vector
of computation. One branch applies the Swish activation while
P⃗α = [P (⃗y1 ), P (⃗y2 ), . . . , P (⃗yB )], (5) the other learns a gating function that controls which infor-
mation passes through. This structure enables more complex
which will serve as the input to our DOA estimation methods. functional mappings while maintaining efficient gradient flow
during training.
B. Correlation-Based DOA Estimation 2) Data Augmentation and Iterative Training: To prevent
overfitting and improve generalization, we implement several
The correlation-based method generates theoretical power regularization techniques, including layer normalization, L2
profiles for each candidate DOA α′ and selects the weight regularization with a coefficient of 10−5 , and a dropout
estimated DOA α̂ that exhibits the highest correlation rate of 0.7 between layers. For training, we process data in
with the input power profile vector. To calculate our batches of 256 examples using the Adam optimizer [12],
theoretical power profiles, we must calculate P⃗α′ = which adapts learning rates for each parameter based on
[|g(α′ , β1 )|2 , |g(α′ , β2 )|2 , . . . , |g(α′ , βB )|2 ] for each α′ and gradient estimates. A learning rate scheduler gradually reduces
βn . The candidate angles are constrained to the range of ±45◦ , the rate from 0.1 to 0.01 over 3,750 iterations of unique
due to the scan range of the device. For a given signal ⃗s and augmented versions of the dataset.
steering vectors α ⃗ ′ and β⃗n , we can model the signal incident BeamSeek relies on an exhaustive calibration scan of all
on the UPA as outer-product ⃗sT α ⃗ ′ and the signal received beams and all angles, differentiating it from the baseline
T ′ ⃗H
by steering angle βn as ⃗s α ⃗ β . This results in the signal correlation-based approach. To compensate for the limited
being scaled by the inner-product of α ⃗ ′ and β⃗n , yielding the data in a single scan, we use data augmentation techniques to
derivation expand the scan into a much larger training set. Our data aug-
g(α′ , βn ) = α ⃗ ′ β⃗ H ,
n (6) mentations emulate perturbations in real-world environments,
including additive white noise and multipath interference,
which can be used to generate the theoretical power profiles. training the neural network to be more robust and accurate.
To select α̂ we maximize the cosine similarity between both Every training iteration, new augmented data is passed to the
power profile vectors: FFN. The foundational data for the iterative training method
was a high-SNR (39.9 dB) beam scan, allowing a more
P⃗α · P⃗αT′ accurate estimate of the SNR after augmentation, as well as
α̂ = arg max . (7) forming a strong base for augmentation and training due to
α′ ∈[−45◦ , 45◦ ] ∥P⃗α ∥∥P⃗α′ ∥
a greater density of features pertinent to the phased array’s
Cosine similarity was chosen as the metric because it does not response to a given incident signal.
consider the magnitude of the vectors, which is significant be- MmWave phased arrays are specifically designed for Line-
cause P⃗α is scaled by unknown constant |γhg(α, βn )|2 P (sn ). Of-Sight (LOS) wireless communications, so the multipath
Fig. 4. The augmentation process of base training data including the LOS power profile P ⃗LOS (a), the data after the addition of spatially correlated multipath
components P ⃗T OT (b), and the final training data after the addition of white Gaussian noise (c) used for one iteration of training.

effects are modeled as Rician. Normally, the Non-Line-of- where P⃗N LOS is the power profile of the NLOS components
Sight (NLOS) components are independently sampled from as a function of α and β, M is the number of NLOS
a Rayleigh distribution as it has been shown that multipath components, Pi is the power contribution of the ith NLOS
contributions, particularly in an urban environment, follow this component, αi and βi are the angle and beam index of the
distribution [13]. In this case, independently adding noise from NLOS component, and σα and σβ are the spread parameters
a Rayleigh distribution across the power profiles would not that shape the the NLOS peak. These parameters are randomly
be suitable because the distortion caused by NLOS signals chosen during augmentation. For a large M , P⃗N LOS will
should be spatially correlated [14]; a strong peak outside of follow a joint Gaussian distribution because it is the sum of
LOS should also be seen by the adjacent beams. The phase many small Gaussian-distributed components. The envelope of
shift caused by multipath interference does not need to be a joint Gaussian distribution is a Rayleigh distribution, so our
considered because BeamSeek only processes power profiles, augmentation model can approximate conventional Rayleigh
which are unaffected by phase. Therefore, the sum of Gaussian fading models in cases with a large M .
spatial distributions is an appropriate model to use here. The The generated NLOS power profile is combined with the
scattered NLOS peaks are generated as LOS power profile according to the Rician model:
K ⃗ 1 ⃗
P⃗T OT = PLOS + PN LOS , (11)
M
X 
(α − αi ) 2
(β − βi )
 2 K +1 K +1
P⃗N LOS = Pi exp − − , (10)
2σα2 2σβ2 where P⃗T OT is the total received power, K is the Rician
i=1
factor, P⃗LOS is the LOS power profile, and P⃗N LOS is the
power profile of the NLOS components. Finally, additive
white gaussian noise, determined by a given SNR value, is
added to P⃗T OT to create the final augmented training data.
This augmentation process, shown in Figure 4, is repeated
at every training iteration to create new, diverse data for the
BeamSeek MLP.
IV. E XPERIMENTAL E VALUATION
A. Hardware Platform
The experiment utilizes the Sandbox 1 (SB1) environment
in the COSMOS (Cloud Enhanced Open Software Defined
Mobile Wireless Testbed for City-Scale Deployment) platform
supported by the National Science Foundation’s (NSF) Plat-
forms for Advanced Wireless Research (PAWR) program [15].
This testbed features a pair of Sivers IMA 57-66 GHz WiGig
transceivers coupled with Xilinx UltraScale+ ZCU111 RFSoC
devices that provide baseband signal processing capabili-
ties, allowing real-time experimentation with programmable
mmWave systems.
Most importantly, the Sivers IMA phased arrays at SB1
Fig. 5. Sivers IMA transceiver mounted on moving XY table. are mounted on movable XY tables, which allows the relative
position and the physical angle (α) of the RX array to be
adjusted. Figure 5 includes a close-up view of the Sivers
IMA WiGig transceiver [15], demonstrating the hardware
constraint. It only provides the combined In-phase/Quadrature
(I/Q) data across the entire array, as opposed to the I/Q
data of each element in the array. Figure 5 also shows the
array mounted on a moving XY table and the surrounding
environment. In contrast to the evaluation of previous meth-
ods, which were only tested in simulation or RF anechoic
chambers, the conditions of this testbed better reflect real-
world deployments. Although there is an unobstructed LOS
between the arrays, there are multiple structures and radios
which may generate interference and multipath effects.
B. Data Acquisition
In the experiment, the arrays are first moved to face each
other head-on. The transmit array remains stationary while
continuously sending QAM-modulated random data. During
the transmission, the receive array is rotated over the full az-
imuthal range (−45◦ to 45◦ ). At each α, an exhaustive sweep
Fig. 6. Average DOA estimation error in degrees for the correlation-based
of all 64 beams in the Sivers IMA transceiver’s integrated method and BeamSeek with eight scan beams (B = 8)
beambook is performed, receiving 512 samples of raw I/Q
data per beam. This set of raw I/Q data, obtained at a very
high SNR level (40 dB), forms the basis of the training set. BeamSeek also scales down to a smaller number beams
The dataset is repeatedly sliced based on the number of scan better than the correlation method; as shown in Figure 8, the
beams, returning equally spaced beam indices across the total difference in the estimation error between three scan beams
scan range, as well as segments of the raw I/Q samples for and sixteen scan beams is smaller. It is important to note that
each beam. The gain profiles of each DOA can be computed the Sivers IMA transceiver’s integrated beambook contains
based on the variance of the I/Q data, which is later augmented only 64 beams to cover a 91◦ range, so errors as small as
during training. 2◦ would have a negligible effect on the performance of the
The test dataset is obtained in a similar fashion, but the system, especially considering the spread of the beam over
beam index is rapidly switched to receive a total of 512 distance. BeamSeek’s accuracy increases for a greater number
samples across the desired number of beams B to emulate of scan beams, but the estimation error for the DOA still
a deployed version of this method. Twenty test gain profiles increases as the divergence from broadside (|α|) increases.
were collected from the testbed at each of the SNR ranges This trend suggests that the model is biased to estimate the
outlined in Table I. The DOA estimation error, defined as DOA closer to 0◦ because it minimizes the error across all
ϵ = α − α̂, is averaged to produce the results in Section V. incorrect predictions. Training the model using a different
optimization criterion may decrease this bias.
TABLE I
SNR VALUES AND C ORRESPONDING L ABELS TABLE II
A BSOLUTE VALUE OF AVERAGE ERROR |ϵ̄| FOR NUMBER OF SCAN BEAMS
Peak SNR (dB) Label B WITH THE CORRELATION METHOD (C ORR ) AND B EAM S EEK (MLP)
0 Low SNR
5 Mid-Low SNR Low SNR Mid-Low SNR Mid-High SNR
10 Mid SNR B Corr MLP Corr MLP Corr MLP
15 Mid-High SNR 3 21.45◦ 14.40◦ 19.55◦ 11.54◦ 16.06◦ 9.54◦
4 21.15◦ 12.92◦ 18.73◦ 10.55◦ 17.68◦ 11.04◦
6 14.84◦ 9.99◦ 12.00◦ 6.97◦ 10.83◦ 9.24◦
V. R ESULTS AND D ISCUSSION 8 11.18◦ 7.73◦ 7.66◦ 4.30◦ 5.96◦ 3.86◦
12 8.99◦ 6.10◦ 5.13◦ 3.02◦ 4.23◦ 2.45◦
The average error in DOA estimation using the baseline 16 7.26◦ 5.47◦ 4.28◦ 2.50◦ 4.14◦ 1.96◦
correlation method and BeamSeek for a varying number
of beams is summarized in Table II. Figure 6 depicts the
average angle estimation error for both methods using 8 VI. C ONCLUSION AND F UTURE W ORK
scan beams. BeamSeek shows a higher accuracy than the In this paper, we introduced BeamSeek, an MLP-based
agile correlation method, which is expected given that it was approach for swift DOA estimation that combines agile beam-
calibrated and has a higher computational complexity. The switching with DL. This approach addresses the limitations of
iterative training method was effective, because BeamSeek high-complexity hardware techniques that require multiple RF
demonstrated stronger noise resilience, as shown in Figure 7. chains, exhaustive beam-sweeping methods with high latency,
 estimation for next-generation wireless systems with hardware
constraints, contributing to the development of more efficient
and reliable mmWave communication links.
ACKNOWLEDGMENT
The authors would like to thank the COSMOS team at
Rutgers University and Columbia University for providing
access to their testbed resources. Specifically, we are grateful
to Ivan Seskar for maintaining the mmWave devices used in
this experiment. We would also like to thank Dr. Alexandra
Gallyas-Sanhueza and her team at IBM T.J. Watson Research
Center for pioneering the agile beamforming method. Finally,
we would like to acknowledge the National Science Founda-
tion’s Platforms for Advanced Wireless Research program for
supporting the COSMOS testbed.
Fig. 7. Average DOA estimation error in degrees at every SNR range for
correlation-based method and BeamSeek with B = 8. R EFERENCES
[1] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N.
Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave
mobile communications for 5G cellular: It will work!” IEEE Access,
vol. 1, pp. 335–349, 2013.
[2] X. Wang, L. Kong, F. Kong, F. Qiu, M. Xia, and S. Arnon, “Millimeter
wave communication: A comprehensive survey,” IEEE Communications
Surveys & Tutorials, vol. 20, no. 3, pp. 1616–1653, 2018.
[3] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
tion,” IEEE Transactions on Antennas and Propagation, vol. 34, no. 3,
pp. 276–280, 1986.
[4] R. Roy and T. Kailath, “ESPRIT-estimation of signal parameters via
rotational invariance techniques,” IEEE Transactions on Acoustics,
Speech, and Signal Processing, vol. 37, no. 7, pp. 984–995, 1989.
[5] J. Choi, “Analog beamforming for low-complexity multiuser detection
in mm-wave systems,” IEEE Transactions on Vehicular Technology,
vol. 65, no. 8, pp. 6747–6752, Aug. 2016.
[6] A. Gallyas-Sanhueza, S. G. Sanchez, A. Paidimarri, A. Tzadok,
B. Sadhu, and A. Valdes-Garcia, “Angle of arrival estimation within
one ofdm symbol leveraging agile beamforming,” in 2024 IEEE Interna-
Fig. 8. Average DOA estimation error in degrees with a varying number of
tional Symposium on Phased Array Systems and Technology (ARRAY),
beams B for correlation-based method and BeamSeek with Mid-SNR.
Oct. 2024, pp. 1–8.
[7] M. Giordani, M. Polese, A. Roy, D. Castor, and M. Zorzi, “A tutorial
on beam management for 3GPP NR at mmWave frequencies,” IEEE
and correlation-based methods that struggle in noisy environ- Communications Surveys & Tutorials, vol. 21, no. 1, pp. 173–196, 2019.
ments. Our experimental results demonstrate that BeamSeek [8] Z. Xiao, T. He, P. Xia, and X.-G. Xia, “Hierarchical codebook design
for beamforming training in millimeter-wave communication,” IEEE
achieves a superior performance to the baseline correlation- Transactions on Wireless Communications, vol. 15, no. 5, pp. 3380–
based method across various SNRs, achieving up to an 8◦ 3392, 2016.
reduction in average angle estimation error. These advantages [9] H. Hassan, A. R. Maud, and M. Amin, “Deep learning based DOA esti-
mation in low SNR and multipath scenarios,” Department of Electrical
make BeamSeek particularly suitable for practical mmWave Engineering, Institute of Space Technology, Islamabad, Pakistan, Tech.
deployments in noisy environments and hardware-constrained Rep., 2024.
systems, such as drones and vehicles. BeamSeek does not [10] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learning for
super-resolution channel estimation and DOA estimation based massive
depend on a known pilot, increasing its scope for Signals MIMO system,” IEEE Transactions on Vehicular Technology, vol. 67,
Intelligence (SIGINT) applications, as well as decreasing its no. 9, pp. 8549–8560, 2018.
overall latency in systems where the pilot is only transmitted [11] N. Shazeer, “Glu variants improve transformer,” 2020. [Online].
Available: https://arxiv.org/pdf/2002.05202
at the start of a frame or symbol. [12] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
This work can be further expanded in many ways. Im- arXiv preprint arXiv:1412.6980, 2014.
[13] T. S. Rappaport, Wireless Communications: Principles and Practice,
plementing a blended approach where a neural network is 2nd ed. Upper Saddle River, NJ: Prentice Hall, 2002.
used to tune the output of the correlation-based method may [14] G. D. Durgin and T. S. Rappaport, “Theory of multipath shape fac-
decrease the computational complexity. Integrating BeamSeek tors for small-scale fading wireless channels,” IEEE Transactions on
Antennas and Propagation, vol. 48, no. 5, pp. 682–693, May 2000.
to perform additional physical layer functions, such as channel [15] T. Chen, P. Maddala, P. Skrimponis, J. Kolodziejski, X. Gu, A. Paidi-
estimation and equalization, will further increase its practi- marri, S. Rangan, G. Zussman, and I. Seskar, “Programmable and
cality. Lastly, adapting BeamSeek for hybrid MIMO systems open-access millimeter-wave radios in the pawr cosmos testbed,” ACM
MobiCom ’21: The 27th Annual International Conference on Mobile
composed of several analog subarrays will enable it to be Computing and Networking, Sep. 2021.
implemented at larger scales.
BeamSeek is a new path toward practical, robust DOA