Sensing-Assisted Predictive Beamforming

• Sensing-assisted predictive beamforming is a method that uses various sensing modalities to predict optimal beam directions before misalignment occurs.
• It integrates RF, vision, LiDAR, and digital twin data to enhance beam management under high mobility and reduce overhead from exhaustive beam sweeping.
• The technique employs model-based filtering, optimization, and learning mechanisms, making it effective in mmWave, massive MIMO, UAV, vehicular, and cell-free systems.

Sensing-assisted predictive beamforming (SAPB) denotes a class of beam management techniques that uses sensed motion, geometry, environment, or scene dynamics to predict future beam directions or focusing points before misalignment occurs. Across recent literature, SAPB spans integrated sensing and communication (ISAC), non-RF sensing, channel-fingerprint inference, digital-twin-aided path prediction, and cooperative multi-station tracking. Its common purpose is to reduce or replace exhaustive beam sweeping, fast-timescale CSI acquisition, and feedback latency under mobility, particularly in mmWave, massive MIMO, near-field, UAV, vehicular, and cell-free settings (Jiang et al., 10 Jun 2025).

1. Physical basis and problem formulation

The central problem in SAPB is that reactive beam tracking becomes inefficient when coherence time is short, mobility is high, or beamwidth is very narrow. In vehicular and UAV regimes, repeated sweeping or CSI feedback can consume a nontrivial fraction of available resources, while stale beam decisions lead to misalignment and rate loss. Several works therefore reformulate beam management as a prediction problem: estimate a latent mobility or scene state from sensing, propagate that state forward over one or more future slots, and synthesize the corresponding beamformer without waiting for explicit channel reacquisition (Liu et al., 2022).

In codebook-based mmWave systems, this predictive task is often written as beam-index selection. A representative objective is

$$b_t^* = \arg\max_{b \in \mathcal{B}} \frac{|w_r^H H_t w_t|^2}{\sigma_n^2},$$

with multi-slot extensions that predict beam indices for the current and several future slots directly as a classification problem over a finite codebook (Ma et al., 14 Sep 2025). This formulation is especially natural when the communication stack already operates with oversampled receive or transmit codebooks.

A major bifurcation in the literature is between far-field and near-field SAPB. Far-field models mainly encode angle and uniform Doppler, so predictive accuracy often depends on explicit trajectory priors or state evolution models. Near-field models, by contrast, exploit spherical wavefront curvature and element-dependent Doppler, which can expose both range and transverse velocity. This is why near-field work describes “prior-knowledge-free prediction” as a defining advantage: full-dimensional sensing closes the sensing–prediction loop without relying on assumed trajectories (Jiang et al., 10 Jun 2025).

2. Sensing modalities and observation spaces

SAPB is not tied to a single sensing source. The observation space may be purely RF, purely non-RF, or explicitly fused across modalities, depending on deployment architecture and latency constraints.

Sensing source	Typical observable	Representative papers
Near-field or ISAC echoes	Delay, Doppler, angle, range, velocity	(Jiang et al., 10 Jun 2025, Jiang et al., 2024, Wang et al., 2023)
Infrastructure vision	Motion masks or image sequences	(Ma et al., 14 Sep 2025)
Uplink channel fingerprints	Beam-domain CIR amplitudes from SRS	(Pjanić et al., 10 May 2025)
Multimodal sensors	LiDAR, camera, and UE coordinates	(Patel et al., 2024)
Environment priors	LoS/NLoS path directions and partial gains from a digital twin	(Jiang et al., 2024)

Vision-assisted work at 60 GHz uses a BS-mounted RGB camera to produce motion masks from frame differencing and thresholding, then predicts codebook beams for the current and six future time slots from compact visual sequences (Ma et al., 14 Sep 2025). Channel-fingerprint-driven work in 5G NR instead uses uplink SRS, Hann windowing, IDFT-based beam-domain CIR extraction, and transformer inference on amplitude-only features, without GNSS, inertial, radar, LiDAR, or vision inputs (Pjanić et al., 10 May 2025). Multimodal beamforming generalizes this by inferring an AoD-only beamspace from UE-side LiDAR and coordinates together with BS-side images (Patel et al., 2024).

Environment-aware SAPB introduces another sensing layer: a digital twin comprising an EM 3D model and ray tracing can predict LoS and NLoS path directions and rank candidate paths by partial pre-target gain, enabling predictive joint sensing/communication beam design without exhaustive search (Jiang et al., 2024). In a different direction, SSB-based ISAC for low-altitude UAV communications reuses periodic NR synchronization signal blocks as active probing signals and combines 2D range–velocity profiling with augmented beamspace MUSIC under hybrid UPA constraints (Tang et al., 21 Apr 2026).

RF-only state tracking also appears in distributed and cell-free architectures. Distributed MIMO V2X uses multi-RSU bi-static delay and Doppler sensing fused at a CPU (Akçalı et al., 29 Jan 2025), while cell-free massive MIMO tracks idle users via multi-static OFDM sensing and then switches to pilot-free predictive beamforming when service is requested (Kama et al., 8 Oct 2025). A related multi-user cell-free framework partitions users into communication and sensing groups and performs sensing only when state uncertainty exceeds a threshold tied to half-power beamwidth (Kama et al., 21 Apr 2026). At the other end of the spectrum, UAV-assisted massive MIMO work has already proposed explicit fusion of wireless and sensor data, an EKF state-space model, and predictive beamforming, with significantly improved position/orientation estimation accuracy and higher spectral efficiency (Lee et al., 2023).

3. Estimation, tracking, and learning mechanisms

Model-based Bayesian filtering remains one of the most common SAPB backbones. The canonical pattern is a motion model, a nonlinear measurement model, and EKF prediction/update recursions that maintain both a state estimate and its covariance. This structure appears in near-field tracking across multiple CPIs (Jiang et al., 2024), distributed MIMO predictive beamforming (Akçalı et al., 29 Jan 2025), SSB-based UAV tracking with maneuver-aware covariance correction (Tang et al., 21 Apr 2026), and cell-free sensing management (Kama et al., 8 Oct 2025). In the LAWN setting, the state is

$\mathbf{s}_k^{(i)} = [x, y, z, v_x, v_y, v_z]^T,$

with nearly constant velocity dynamics and measurement functions in range, radial velocity, azimuth, and elevation (Tang et al., 21 Apr 2026).

Near-field SAPB also supports optimization-based single-CPI estimation. In mono-static near-field velocity sensing, the received echo depends on per-element ranges and Dopplers that encode both radial and transverse motion. The resulting maximum-likelihood problem jointly estimates motion components from echo data and then updates the beamformer without explicit CSI estimation (Wang et al., 2023). A later near-field formulation combines AGD-AO within a single CPI and EKF across multiple CPIs, arguing that AGD-AO stabilizes descent under small gradients while EKF is more robust in low-SNR regions because it fuses information over time (Jiang et al., 2024).

Learning-based SAPB covers several distinct paradigms. One line performs sequence modeling directly from sensed inputs. A CNN+GRU+attention teacher predicts beam indices for current and future slots from motion-mask sequences; a lightweight student then preserves long-horizon prediction through knowledge distillation despite shorter input histories (Ma et al., 14 Sep 2025). Another line uses ConvLSTM to map historical sensing-informed channels to next-slot beamforming matrices under CRLB-based sensing constraints (Liu et al., 2022). A third uses transformers on SRS-derived beam-domain channel fingerprints to predict future downlink beam energy maps in commercial 5G NR (Pjanić et al., 10 May 2025). In multi-user mmWave ISAC, deep reinforcement learning allocates single beams to users whose AoD precision is already sufficient and multiple sensing beams to users that require updates, using only sensing echoes and no user feedback or prior state evolution information (Wang et al., 9 May 2025).

These threads suggest that SAPB is not a single algorithmic family. It is better understood as a systems concept whose inference engine may be Bayesian, optimization-based, supervised, distillation-based, or reinforcement-driven, provided the output is a proactive beam decision.

4. Beam design paradigms

The most common beam design output is either a predicted codebook index or a predicted analog/digital beamformer. In codebook systems, predictive beam tracking is naturally cast as multi-step classification. The output may be a Top-1 beam for direct use or a Top-$K$ subset for robust follow-up refinement and proactive scheduling (Ma et al., 14 Sep 2025).

Near-field SAPB replaces angle-only steering with spatial focusing. A representative near-field steering component is

$$a_n(r(t)) = \frac{1}{\|r(t)-r_n\|}\exp\!\left(-j\frac{2\pi}{\lambda}\|r(t)-r_n\|\right),$$

and the element-specific Doppler is

$$f_D^{(n)}(t) = -\frac{1}{\lambda}\frac{d}{dt}\|r(t)-r_n\|.$$

Given a predicted position $\tilde r$, beam synthesis takes the form

$w(t+\Delta)=\frac{a(\tilde r)}{\|a(\tilde r)\|},$

optionally with per-element or per-subcarrier Doppler compensation in RF or baseband (Jiang et al., 10 Jun 2025). This same logic underlies pilot-free predictive beamforming based on near-field velocity sensing, where focusing weights and Doppler correction are derived from sensed state rather than from estimated CSI (Wang et al., 2023).

Joint sensing–communication formulations embed SAPB into constrained optimization. Digital-twin-assisted ISAC maximizes sensing SNR over a dominant predicted LoS or NLoS path while satisfying a communication SINR requirement, using SDR and, optionally, a null-space heuristic (Jiang et al., 2024). Symbiotic sensing and communication in bistatic vehicular systems maximizes achievable sum rate under a CRLB constraint on weak-target AoA estimation, with PDD-based alternating algorithms for fully digital and hybrid analog-digital arrays (Xia et al., 2024). ConvLSTM-based predictive beamforming in vehicular ISAC also uses CRLB constraints, but shifts the optimization burden into self-supervised network training (Liu et al., 2022).

A more explicitly robust version appears in SSB-based UAV SAPB, where instantaneous CSI is replaced by predictive correlation matrices derived from EKF state distributions. The expected SINR is then

$$\overline{\mathrm{SINR}}_u = \frac{\mathbf{w}_u^{\mathrm{H}}\mathbf{R}_{\theta,\phi,u}\mathbf{w}_u} {\sum_{v\ne u}\mathbf{w}_v^{\mathrm{H}}\mathbf{R}_{\theta,\phi,u}\mathbf{w}_v+\sigma_{\mathrm{eff},u}^2},$$

and hybrid beamforming is optimized via SDR/SCA under uncertainty-aware constraints (Tang et al., 21 Apr 2026).

In distributed and cell-free settings, the beamformer can be simpler because prediction supplies geometry directly. Distributed MIMO uses per-RSU steering vectors based on one-step or two-step EKF angle prediction (Akçalı et al., 29 Jan 2025). Cell-free pilot-free beamforming maps tracked positions to predicted LoS channels and then applies per-AP MRT or predictive MMSE precoding, eliminating uplink pilot acquisition when users transition from idle to active service (Kama et al., 8 Oct 2025, Kama et al., 21 Apr 2026).

5. Empirical regimes and reported gains

The reported gains of SAPB vary by sensing modality and system architecture, but several representative results recur. In vision-assisted long-term beam tracking, the attention-enhanced teacher achieves average Top-5 accuracy exceeding 93% across the current and six future time slots, with ATop-5 approximately 94.6% and ADBA approximately 95.0%. The student closely matches this behavior with about 90% fewer parameters than the teacher and 60% shorter input sequences, while the teacher already reduces complexity by about 90% versus state-of-the-art sensing-aided baselines (Ma et al., 14 Sep 2025).

In multi-user multimodal beamforming, using LiDAR, camera, and coordinates to estimate AoD-only beamspace yields more than $4\times$ improvement in median sum spectral efficiency at 42 dBm EIRP with 4 active users relative to a single-user beam-prediction baseline extended to the multi-user case (Patel et al., 2024). In digital-twin-assisted ISAC, the proposed design achieves near-optimal target sensing SNR in both LoS- and NLoS-dominant areas while ensuring the required communication SINR (Jiang et al., 2024).

Learning-based predictive beamforming for vehicular ISAC reports achievable sum-rate close to the genie-aided perfect-ICSI upper bound while satisfying sensing constraints, showing that next-slot beamforming can be learned directly from historical sensing-informed channels (Liu et al., 2022). In distributed MIMO V2X, distributed antennas provide more uniform and robust sensing performance, coverage, and data rates than a co-located massive MIMO array while the vehicular user is in motion (Akçalı et al., 29 Jan 2025).

Resource-managed cell-free SAPB pushes the overhead argument further. In the multi-user state-based framework, sensing is needed only occasionally after an initial convergence period, approximately 1% of epochs, while downlink spectral efficiency remains close to the perfect-CSI case (Kama et al., 21 Apr 2026). In low-altitude UAV communications, SSB-based robust SAPB yields the highest gain at short SSB periods, for example more than 32% at $T=10$ ms, and remains advantageous at longer SSB periods because it avoids per-user feedback overhead and widens beams when covariance inflation indicates maneuver uncertainty (Tang et al., 21 Apr 2026).

Maritime UAV–buoy SAPB provides a contrasting environment. There, the pronounced RMSE degradation of the communication-only benchmark confirms that sensing-assisted state refinement is essential when motion is wave-driven and clutter-contaminated; the proposed PCRB-driven design remains robust under denser buoy deployments and harsher sea conditions (Li et al., 15 Jun 2026).

6. Limits, misconceptions, and research directions

A common misconception is that SAPB is synonymous with vision-based beam prediction. The literature is much broader: it includes near-field echo processing, channel-fingerprint inference from uplink pilots, digital-twin path prediction, SSB-based monostatic sensing, multi-static OFDM tracking, and multimodal LiDAR–camera–position fusion (Jiang et al., 10 Jun 2025). A related misconception is that SAPB is inherently a neural-network technique. In practice, some of the strongest formulations are still EKF-, ML-, SDP-, CRLB-, or Schur-complement-based (Jiang et al., 2024).

Another important distinction concerns “prior-knowledge-free” prediction. Near-field work argues that spherical waves and non-uniform Doppler enable direct sensing of full motion state, which can remove the need for trajectory priors under arbitrary motion (Jiang et al., 10 Jun 2025). That claim does not automatically transfer to far-field or codebook-only systems, which often still depend on motion smoothness, synchronized sensing, or architectural priors such as a calibrated digital twin or fixed sensing geometry (Tang et al., 21 Apr 2026).

The dominant limitations are also consistent across papers. Generalization across environments and sensing modalities remains difficult; multimodal fusion can improve accuracy but raises complexity and calibration demands (Patel et al., 2024). Many formulations assume LoS dominance, sparse multipath, accurate synchronization, or static infrastructure models; dynamic objects and material mismatches can degrade digital-twin predictions (Jiang et al., 2024). Near-field schemes face transition-region sensitivity, calibration demands, oscillator phase-noise issues, and degraded transverse-velocity observability as range moves toward the far field (Jiang et al., 2024). Cooperative and cell-free methods improve robustness, but fronthaul latency, subset selection, and centralized fusion overhead become critical (Yang et al., 18 Aug 2025).

Current research directions therefore concentrate on broader ISAC co-design and stronger uncertainty handling. These include learning-enhanced predictors such as deep Kalman filters and Transformers, RIS and holographic MIMO for improved LoS and aperture, cooperative BS arrays for geometric diversity, new waveforms adapted to near-field mobility such as OTFS, AFDM, and ODDM, flexible antennas, multimodal knowledge distillation, and tighter integration of sensing accuracy with communication objectives (Jiang et al., 10 Jun 2025). This suggests that SAPB is evolving from a beam-tracking heuristic into a general architecture for state-aware, uncertainty-aware, and overhead-aware beam management.