To unlock the multi-gigabit data rates promised by millimeter-wave (mmWave) and Terahertz (THz) frequencies, wireless networks must overcome a fundamental physical hurdle: severe path loss. Unlike lower-frequency signals that propagate omnidirectionally, high-frequency signals are easily attenuated by atmospheric absorption and physical obstacles. To bridge this gap, transmitters and receivers use large antenna arrays to focus radio energy into highly directional, narrow "pencil" beams.
However, this directional focus introduces a coordination problem known as beam alignment. Before any data can be transmitted, the base station and the mobile device must point their narrow beams directly at each other. As noted by Zhou et al. in their analysis of IEEE 802.11ay-Based mmWave WLANs: Design Challenges and Solutions, establishing these links requires an initial training phase to discover the optimal antenna configuration, followed by continuous tracking to adapt to environmental changes.
The Latency Bottleneck of Exhaustive Sweeping
The traditional approach to beam alignment is exhaustive beam sweeping. The transmitter sequentially steps through every possible beam direction while the receiver listens across its own set of angles. While reliable, this brute-force search creates a massive temporal and computational bottleneck.
The core issue is scalability. As antenna arrays grow to achieve higher gain, beams become narrower, and the number of potential beam combinations scales quadratically. Xu et al. point out in Hashing Beam Training for Near-Field Communications that this linear-to-quadratic scaling with the number of antenna elements leads to excessive training times, consuming critical signaling resources that would otherwise be used for data transmission.
This overhead becomes catastrophic in high-mobility scenarios, such as vehicle-to-everything (V2X) networks or unmanned aerial vehicle (UAV) communications. In a highway environment, as modeled by Tassi et al. in Modeling and Design of Millimeter-Wave Networks for Highway Vehicular Communication, the relative positions of nodes change rapidly. If the time required to complete an exhaustive sweep exceeds the channel coherence time—the window during which the spatial path remains stable—the alignment process becomes obsolete before it even finishes. As Chafaa et al. observe in Exploiting Channel Sparsity for Beam Alignment in mmWave Systems via Exponential Learning, this mismatch leads to frequent connection drops and severe throughput degradation.
Out-of-Band Sensing and Context Awareness
To bypass the slow process of sweeping the radio frequency (RF) spectrum, a prominent direction of research leverages out-of-band sensing. The intuition is simple: if a base station can see or track a vehicle using non-RF sensors, it can calculate the geometric angle of arrival (AoA) or departure (AoD) instantly, narrowing the beam search space from a massive grid to a few highly probable candidates.
Several modalities have been successfully integrated into this framework:
- Computer Vision: By mounting cameras on base stations, deep learning models can detect objects and predict beam configurations. For instance, Salehi et al. demonstrate in Machine Learning on Camera Images for Fast mmWave Beamforming that a convolutional neural network (CNN) can process standard camera images to locate transmitters and receivers, slashing alignment latency by over 90%. To support this, Kim et al. introduced the VOMTC: Vision Objects for Millimeter and Terahertz Communications dataset to train vision models specifically for identifying wireless devices in real-world environments.
- LIDAR and GPS: In vehicular networks, combining structural point clouds with positioning data provides a rich geometric map. Zecchin et al. show in LIDAR and Position-Aided mmWave Beam Selection With Non-Local CNNs and Curriculum Training that non-local attention mechanisms can extract spatial features from LIDAR data to achieve near-optimal throughput with zero communication overhead. Similarly, Wang et al. utilize real-time GPS coordinates in Position Aware 60 GHz mmWave Beamforming for V2V Communications Utilizing Deep Learning to proactively predict and steer beams before a link degrades.
- Integrated Sensing and Communication (ISAC): Instead of external sensors, the communication signal itself can act as a radar. Cui et al. demonstrate in Sensing-assisted Accurate and Fast Beam Management for Cellular-connected mmWave UAV Network that integrating visual data with radar-like echo signals allows high-speed mobile units like UAVs to maintain precise beam tracking without heavy feedback loops. Saad in Joint Sensing and Communication for Situational Awareness in Wireless THz Systems further shows how tensor decomposition of uplink signals can construct a real-time environmental map, transforming reactive tracking into proactive alignment.
Algorithmic Subspace Reduction and Digital Twins
When external sensors are unavailable, researchers reduce the search space algorithmically by exploiting the underlying mathematical structure of the wireless channel or the environment.
One approach is hierarchical codebook design. Instead of testing narrow beams individually, the system first transmits wide, low-gain beams to identify a general sector, and then recursively narrows the search space. Xia et al. proposed a hierarchical codebook framework in Hierarchical Codebook Design for Beamforming Training in Millimeter-Wave Communication that creates flat, wide beams without sacrificing transmit power, accelerating the training phase under strict hardware constraints.
Alternatively, the search can be framed as an online learning problem. By treating beam directions as arms in a multi-armed bandit problem, algorithms can balance exploring new directions with exploiting known good paths. Zhang et al. developed the Hierarchical Beam Alignment (HBA) algorithm in Fast mmwave Beam Alignment via Correlated Bandit Learning, which uses correlated bandit learning to intelligently "zoom in" on optimal directions, reducing alignment time from seconds to milliseconds.
To scale this to multi-user or multi-surface environments, Li developed a "Hashing Multi-arm Beam" training scheme in Low-Complexity Beam Training for Multi-RIS-Assisted Multi-User Communications that uses pseudo-random hash functions to test multiple directions simultaneously, dropping the overhead complexity from exponential to logarithmic scales.
Finally, the concept of Digital Twins bridges the gap between physical simulation and real-time execution. Rather than performing physical sweeps, the edge network maintains a virtual 3D replica of the environment. Massoli et al. built a pipeline in Vision-Assisted Digital Twin Creation for mmWave Beam Management that uses a single camera to construct a 3D digital twin, running physics-based ray-tracing simulations to predict optimal beams.
To handle real-world uncertainties, Salehi et al. proposed the "Multiverse" paradigm in Multiverse at the Edge: Interacting Real World and Digital Twins for Wireless Beamforming, which hosts a library of digital twins with varying fidelities at the network edge. This allows autonomous vehicles to quickly select the best-fit simulation to predict beams without real-world sweeping latency.
By shifting from blind, exhaustive RF sweeping to context-aware prediction and structured mathematical search, these techniques eliminate the alignment overhead, making high-mobility, high-frequency networks viable for real-time applications.