ISAC-Assisted Proactive Handover in mmWave Networks: Value of Foresight and Localization Requirements

This work has been submitted to the IEEE Conference for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

ISAC-Assisted Proactive Handover in mmWave Networks: Value of Foresight and Localization Requirements Seyed Bagher Hashemi Natanzi∗ , Bo Tang† ∗† Department of Electrical and Computer Engineering, Worcester Polytechnic Institute (WPI), USA

arXiv:2611.03771v1 [eess.SY] 2 Nov 2026

Email: {snatanzi, btang1}@wpi.edu Abstract—Millimeter-wave (mmWave) links can lose tens of decibels within milliseconds when a person or vehicle crosses the line of sight, and reactive handover recovers only after the outage has begun. Integrated sensing and communication (ISAC) promises to let the network see blockers coming and move users to another cell in advance. This paper asks how much such foresight is worth relative to the best reactive handover, and what sensing accuracy is needed to harvest it. We formulate cost-aware cell selection with knowledge of the future signal-tonoise ratio as a sequential decision problem, define the value of foresight through a policy-independent bound, and develop an ISAC-assisted handover framework in which a radio unit senses blockers with its own communication waveform and a simulated Open RAN xApp plans handovers. Ray-traced experiments in a 28 GHz street canyon, reported on held-out scenarios, show that, with a short look-ahead, perfect foresight closes most of the gap between service-aware (Layer-3) reactive handover and an instantaneous oracle. Harvesting this value, however, requires the user and nearby blockers to be localized to about a decimeter, well beyond what our radar tracker and meterlevel user positioning provide; with realistic errors, even an uncertainty-aware planner largely defers to reactive handover. Index Terms—ISAC, mmWave blockage, proactive handover, O-RAN, ray tracing, digital twin.

I. I NTRODUCTION The wide bandwidths of the millimeter-wave (mmWave) bands are central to the capacity of 5G and 6G networks, but their short wavelength makes links fragile: when a pedestrian or a vehicle crosses the line of sight (LoS), the received power drops by tens of decibels within milliseconds and stays low for up to seconds [1], [2]. Today’s mobility procedures are reactive. Beam failure recovery and handover are triggered by degraded measurements, filtered over a time-to-trigger, so the user experiences at least part of every outage. Outages dominate the experience of latency- or rate-critical services. Proactive mitigation requires knowing the blocker before it arrives. Prior work has predicted blockage and triggered proactive handover from cameras or dedicated radars [3]–[5], which adds sensors, cost, and privacy concerns. Integrated sensing and communication (ISAC) offers an alternative: the radio unit senses its surroundings with its own OFDM waveform [6], [7], 3GPP has studied ISAC use cases and standardized sensingtarget channel models [8], [9], and Open RAN (O-RAN) lets *Corresponding author: snatanzi@wpi.edu

an xApp act on such information via the E2 interface [10], [11]. Whether this pays off depends on two quantities that are rarely separated. The first is the value of foresight: what a network with perfect knowledge of the future could gain over the best reactive handover. The second is the realizable part: what a practical predictor and policy actually deliver. Existing studies typically report gains of a specific predictor over link-failure recovery or untuned triggers, so neither quantity is known, and it remains unclear which sensing accuracy a network would need. Measuring the value requires a strong reactive baseline and a policy-independent bound; measuring the realizable part requires a model in which the blocker that the radar observes is the same object that attenuates the link. This paper addresses both questions. Our contributions are: • A formulation of cost-aware cell selection with knowledge of the future signal-to-noise ratio (SNR), solved exactly by dynamic programming, which bounds any proactive policy and defines the value of foresight relative to service-aware reactive handover (Layer-3 event A5). • An ISAC-assisted handover framework in which a radio unit senses blockers with its own waveform, tracks them with digital-twin-aware processing, predicts each cell’s SNR, and plans handovers in an O-RAN xApp (modeled in simulation), including an uncertainty-aware planner. • A spatially consistent simulation built on Sionna RT 2.2 [12], [13] in which every moving body is both a 3GPP radar target and a blocker of communication paths, released as open source.1 • Held-out findings: which blockers matter, that foresight closes most of the improvable gap with a short horizon and cost-aware switching, and, by calibrated error analysis, the user and blocker localization accuracy needed. II. R ELATED W ORK mmWave blockage. Measurements report deep and fast fading caused by pedestrians at mmWave frequencies [2], and 3GPP models blockage with a stochastic model A and a geometric knife-edge model B [9]. System-level studies mostly treat blockers statistically, which decouples the blocker from anything the network could observe. 1 https://github.com/CLIS-WPI/sense2connect

Sensing-aided mobility. Vision-aided methods predict blockage and trigger proactive handover from camera images [3], and radar-aided methods predict beams and blockages from a co-located automotive radar [4], [14], supported by multimodal datasets [5]. RaDaR [15] equips each small cell with a separate FMCW radar, uses object, base-station and user heights to decide whether an object obstructs the LoS, and predicts the time to blockage, with gains over handover after link failure. Look Before Switch [16] uses gNB-side ISAC to track the vehicular user itself and triggers handover from its estimated position and maneuver, cutting the handover interruption by more than half relative to conventional handover. We differ in three respects: we sense third-party blockers with the communication waveform rather than the user or an auxiliary radar; we compare against a tuned service-aware reactive policy (A5) and a cost-aware oracle rather than linkfailure recovery or fixed A3 settings; and we separate the value of foresight from the part that sensing realizes. III. S YSTEM M ODEL AND P ROBLEM F ORMULATION A. Network, Channel, and Blockage We consider a set C of cells, each served by an O-RAN radio unit (O-RU) with an N -element array, and a single-antenna user equipment (UE) u that can be served by any cell in C. The channel from O-RU c to the UE at frequency f and time t is X hc (f, t) = 10−Lℓ (t)/20 αℓ (t) ac (θℓ ) e−j2πf τℓ (t) , (1) ℓ

where αℓ , τℓ , and θℓ are the gain, delay, and departure direction of path P ℓ without blockers, ac is the array response, and Lℓ (t) = k Lk,ℓ (t) is the loss that the moving bodies k cause on path ℓ, a function of their geometry relative to the path segments. We use 3GPP blockage model B [9], in which each body acts as a screen on every path segment it intersects, with loss
Lk = −20 log10 1 − (Fh1 + Fh2 )(Fw1 + Fw2 ) , (2) where each p knife-edge term
is F = π −1 arctan ± π2 πλ−1 (D1 + D2 − r) and D1 , D2 , r are the edge and direct distances of the segment; losses add in decibels over bodies and segments. We denote by γc (t) the SNR that cell c provides to the UE under the receiver model at hand and by γ̄c (t) its value without blockers. The UE requires a rate R, i.e., an SNR γreq , and is in outage whenever γct (t) < γreq for its serving cell ct ; every handover interrupts the user plane for τHO , which also counts as outage. B. Sensing and Prediction Model The O-RUs transmit OFDM sensing symbols as part of their frames. With Nsc subcarriers of spacing ∆f (bandwidth B = Nsc ∆f ) and a coherent processing interval of M symbols spaced Ts , a monostatic radar resolves range, radial velocity, and the maximum unambiguous velocity as ∆R =

c0 , 2B

∆v =

λ , 2M Ts

vmax =

λ , 4Ts

(3)

where c0 is the speed of light. Each moving body is a sensing target with state xk (t) (position, velocity, class). The network obtains estimates x̂k from its tracker and a UE position estimate p̂u with errors, and predicts each cell’s SNR in dB as γ̂c (t) = γ̄c (t) − L̂c (t), (4) where γ̄c comes from the digital twin and the predicted LoSonly loss L̂c applies (2) to the predicted blocker trajectories on the line between O-RU c and p̂u . C. Handover Problem and the Value of Foresight A handover policy selects the serving cell ct ∈ C at every step t of duration ∆t and is judged by its outage time X  J({ct }) = ∆t ⊮ γct (t) < γreq ∨ t ∈ I({ct }) , (5) t

where I contains the τHO /∆t steps after each switch, so that an interrupted step counts once. Reactive policies use only past measurements: A3 hands over when the neighbor’s filtered reference signal received power (RSRP) exceeds the serving one by an offset plus hysteresis for a time-to-trigger [17]; the Layer-3 event A5 hands over when the serving cell falls below one threshold and the neighbor exceeds another, which makes it aware of the service requirement. The instantaneous oracle selects the better cell at every step without delay or interruption. The cost-aware oracle knows γc (t) for all cells and minimizes (5) exactly by the Viterbi algorithm over the cell and the remaining interruption; no policy with any predictor can do better. We call the outage difference between the best reactive policy and the instantaneous oracle the improvable gap, and the share of it closed by the cost-aware oracle the value of foresight. A gridaligned variant restricted to the decision times of a practical controller isolates the cost of that restriction. IV. ISAC-A SSISTED H ANDOVER F RAMEWORK Sensing and tracking. Per coherent processing interval, the sensing O-RU forms range–Doppler–angle maps from the received echoes, detects targets with two-dimensional cellaveraging constant false alarm rate (CFAR) detection, rejects ghost detections that mirror real targets on building walls known from the digital twin (image method), clusters detections with DBSCAN, and tracks them with an extended Kalman filter on range, angles, and radial velocity. Optionally, tracks are constrained to lane center lines (vehicles) and sidewalks (pedestrians, with a free model for crossings). Prediction and planning. An xApp in the near-real-time RAN intelligent controller, modeled in simulation, receives tracks over E2 with period Tr and acts after a loop delay τE2 . At each report it predicts γ̂c by (4); the planner minimizes (5) on γ̂c over a horizon H and its decision grid and applies the first decision after the loop delay. The genie-planner uses the true future SNR and isolates the effect of prediction errors. For comparison, trigger policies hand over when the serving cell’s LoS is predicted blocked and the other cell’s clear.

O-RU (facade, 8 m) LoS over near lane: 2.7 m (lamppost), 3.3 m (facade)

lamppost 5 m

Placeholder: figs/fig˙onset2.pdf. Onset panels.

truck

UE

car near lane

far lane

(a) Street cross-section

Fig. 1. (a) Street cross-section: over the near lane the LoS passes above cars but through tall vehicles. (b), (c) Model-B loss of an example bus/truck and pedestrian outage (with its onset) at 1 ms resolution, with the handover instants of A3 and of the trigger genie on top of A3 (Section IV) at 10 dB margin.

Placeholder: figs/fig˙leadtime.pdf. Share of 10 dB events with a confirmed track before onset.

Fig. 2. Share of outage events whose blocker is tracked a given time before onset, map-constrained vs. unconstrained tracker (95% Wilson intervals; n: events in the cell of O-RU 0, pooled over mounts and densities).

Uncertainty-aware planning. The uncertainty-aware planner keeps A5 and intervenes only if the expected outage reduction, averaged over K samples from the tracks’ (inflated) posterior and the UE position error, exceeds a threshold θ. V. E XPERIMENTS AND R ESULTS A. Setup Scenario. We simulate an 80 m street canyon (periodic, constant traffic density) at fc = 28 GHz with two cells (Fig. 1(a)). O-RU 0 is on the north side, on a lamppost (5 m) or a facade (8 m); O-RU 1 is a 5 m lamppost on the south sidewalk, half a block away. Each O-RU has an 8 × 8 halfwavelength array; UEs carry one antenna at 1.5 m on the south sidewalk. Traffic consists of cars (5 × 2 × 1.6 m), buses (12 × 2.5 × 3.2 m), trucks (8 × 2.5 × 3.5 m), and pedestrians who walk the sidewalks or cross the street; every moving body is a TR 38.901 sensing target [9] (five scattering points for vehicles, one for pedestrians). Each 60-s run has two UEs (1.0 and 1.2 m/s); low/high density places 4/12 cars, 1/2 buses, 1/2 trucks, and 6/12 sidewalk and 2/4 crossing pedestrians (8–12, 6–10, and 1.0–1.5 m/s for cars, buses/trucks, and pedestrians). Communication paths are traced without bodies and attenuated by (2) every ∆t = 10 ms; only the onset analysis uses 1 ms. Sensing. O-RU 0 operates as a monostatic radar under an ideal full-duplex assumption, transmitting on one element

and receiving on its array; Sionna’s RCS solver traces paths through the scattering points. One sensing symbol per slot (numerology 3, 1024 subcarriers) and a 32-ms processing interval give ∆R ≈ 1.2 m, ∆v ≈ 0.17 m/s, and vmax ≈ 21 m/s in (3); one interval every Tr = 0.1 s occupies 2.3% of the resources. Sensing-based policies (the sensing, true-loss, and uncertainty-aware planners and all conditions of Table II) pay it as a rate loss, about 0.26 dB on γreq ; A3, A5, the oracles, and the genie-planner do not. The E2 loop delay is τE2 = 20 ms (modeled, without an implemented E2 service model), and the UE position, an exogenous input from UE-based or network positioning (the radar senses only third-party bodies), carries a Gaussian error of 1 m per horizontal axis. Link budget. Transmit power (33 dBm) and UE noise figure (13 dB) follow TR 38.802 [18]; unspecified losses [19] are set to 0 dB; the rate is Shannon’s capped at the highest NR spectral efficiency [20]. Here, γc adds the path powers with the full array gain after their model-B loss, an optimistic approximation of the received multipath power, since no single beam collects all paths with full gain. A rate of 400 Mbit/s requires γreq = 9.3 dB, and τHO = 20 ms. We define the link margin as the median unblocked best-cell SNR minus γreq . The 3GPP budget gives 35.9 dB on these short links; since longer links, unspecified losses, or higher rates lower it, a common extra path loss sweeps the margin from 0 to 30 dB. Protocol. All policies are tuned per margin on five tuning seeds (A3 over offsets and hysteresis up to 15 dB; the tuned A5 time-to-trigger, 40 ms, is the smallest value of its grid); methods were fixed using ten development seeds, which also calibrate the tracker error model below. All performance results come from a single final evaluation on ten held-out seeds, each simulated for both mounts and both traffic densities. Since the four runs of a seed share the traffic and random draws, the seed, with the mean of its four runs, is the statistical unit: we compare schemes pairwise with the Wilcoxon signedrank test over the ten seeds and give 95% bootstrap CIs over seeds for paired differences (Fig. 4); Fig. 3(a) shows t-intervals

Placeholder: figs/fig˙outage˙value.pdf. Outage and value of foresight.

Fig. 3. (a) Outage time at 400 Mbit/s versus link margin for A3, A5, the sensing-driven planner, the genie-planner (H = 0.5 s), and the two oracles (95% CI over the 40 runs). (b) Share of the A5-to-oracle gap closed by the cost-aware oracle, split by the class of the blocker on the user’s cell, with the genie-planner (triangles) and the sensing-driven planner (below zero: worse than A5, in multiples of the gap); the numbers above the bars: absolute gap [s/UE-min].

over the 40 runs, Fig. 2 Wilson intervals over events.

in 83–90% of heavy-vehicle and 53–65% of pedestrian outages (Fig. 2). Of three uses of the digital twin, static-clutter subtraction is equivalent to blind zero-Doppler removal, because static clutter has exactly zero Doppler in simulation; imagemethod ghost rejection removes 14–21% of the unmatched clusters and, at a fixed false-alarm budget, raises heavy-vehicle detection from 0.83 to 0.88; and, without ghost rejection, mapconstrained tracking raises detection from 0.83 to 0.91 for heavy vehicles and from 0.28 to 0.45 for pedestrians. Sensing thus sees heavy vehicles reliably and pedestrians only about half of the time: on the development seeds, a pedestrian track is confirmed after a median of 1.2 s and then lost for 0.9 s at a time, and along-lane position errors range from 0.53–1.84 m.

B. Which Blockers Matter

D. The Value of Foresight

Users see about 8.5 LoS outage events (10 dB model-B loss) per minute; Table I summarizes them by class. Cars never cause them: for a UE at height hu and an O-RU at height ho and horizontal distance d, the LoS height at distance x from the UE is z(x) = hu +(ho −hu )x/d, i.e., 2.7 m–3.3 m over the near lane: above a car, but through a bus or truck (Fig. 1(a)). Pedestrians block near the UE end of the ray. A blocker shadows the whole angular sector of the serving O-RU, including the ground reflection: the best surviving path lies a median 26–33 dB below the unblocked LoS, so the same cell keeps serving the UE only with a large link margin, and otherwise only a second cell can. That cell is clear in most heavy-vehicle outages but in few pedestrian outages, because a pedestrian close to the UE often shadows both cells. The outages also differ sharply in dynamics (Fig. 1(b), (c)): the loss caused by a bus or truck rises over 0.47 s, whereas a pedestrian next to the UE cuts the link almost at once: 62% of pedestrian onsets are shorter than 10 ms.

Fig. 3(a) shows outage versus margin. A5, the best reactive policy on the tuning seeds, is our reference; on the held-out seeds it is on par with the wide-grid A3 at low margins, and its advantage grows with the margin (0.09 vs. 0.16 s/UE-min at the reference), since A3 switches on RSRP differences even if the serving cell meets the service rate. The cost-aware oracle closes 73–88% of the improvable gap at every margin from 5 dB up (Fig. 3(b)); at 10 dB it reaches 3.13 vs. 3.51 for A5 and 3.04 for the instantaneous oracle. It does so largely by making fewer, better-timed switches, so it avoids interruption as well as blockage. In absolute terms, perfect foresight lowers the outage of A5 by 11% (0.38 s/UEmin) at 10 dB and by 62% (0.057 s/UE-min) at the reference; most of the remaining outage occurs when no cell is usable. High traffic density raises the outage about threefold (A5: 1.76 vs. 5.26 s/UE-min at 10 dB). The genie-planner closes 62–79% of the gap (73% at 10 dB); against the grid-aligned oracle, which exceeds the bound by at most 1.9%, a horizon of 0.5 s suffices. Trigger-based policies do not: even with perfect prediction, the trigger genie stays at the level of the A3 it builds on (3.64 vs. 3.58 at 10 dB), because blockage triggers ignore sub-threshold losses and the cost of the switches they add. Foresight thus needs cost-aware planning, but little lookahead.

TABLE I B LOCKERS OF 10 dB L O S OUTAGES ( RANGES : MIN – MAX OVER MOUNTS ; ONSET: MEDIAN 10–90% RISE TIME ; TRACKED : FINAL MAP - CONSTRAINED TRACKER , O-RU 0 EVENTS ; VALUE : SHARE OF IMPROVABLE GAP AT 10 dB CLOSABLE PER CLASS , REST IS NOT, F IG . 3( B )). Class

Share of 10–90% Other cell Tracked Share of outages onset [s] clear 0.5 s ahead value Bus/truck 41–42% 0.47 58–61% 83–90% 29% Pedestrian 58–59% ≤0.001 29–30% 53–65% 53% Car 0% – – – ≈0%

C. Sensing Performance With the detector tuned to about 3.8 false alarms per processing interval, the track detection probability is 0.94 for buses and trucks and 0.48 for pedestrians (held-out seeds; per target sample), and a confirmed track exists 0.5 s before onset

TABLE II E RROR BUDGET: PAIRED OUTAGE DIFFERENCE TO A5 [ S /UE- MIN ] WITH PERFECT TRACKS AND WITH CALIBRATED TRACKER ERRORS INJECTED ; EACH “+” ROW ADDS ONE ERROR TYPE TO PERFECT TRACKS , AND ALL VALUES ARE TOTAL PAIRED DIFFERENCES TO A5; THE SIZE RULE GIVES CARS BUS SIZE , SO THEY ARE PREDICTED TO BLOCK THE L O S ( † : NOT SIGNIFICANT, SEED - LEVEL W ILCOXON p ≥ 0.05). Condition 10 dB Reference Perfect tracks +0.33 -0.042 + whole-track misses +2.71 +0.108† + persistent false tracks +0.70 +0.006† + correlated pos./vel. error +1.30 +0.049 + class-agnostic size rule +0.89 -0.018† All tracker errors +4.92 +0.099 All + 1 m UE position error +6.91 +0.101 Real sensing-driven planner +6.83 +0.099

Pedestrian blockage accounts for 48–65% of the improvable gap and heavy vehicles for 10–29%; cars contribute only at margins of 0–5 dB, where their smaller losses already cause outage. The blockers that sensing observes well, heavy vehicles, cause slow outages that A5 follows reactively, while most of the value lies in abrupt pedestrian blockage near the UE, which the radar tracks in only about half of the cases. E. What ISAC Delivers With tracker predictions, the planner falls behind A5 at every margin (Fig. 3; 10.34 vs. 3.51 at 10 dB; 0.19 vs. 0.09 at the reference), with 25.3 handovers per UE-minute against 6.6 for A5 and a ping-pong rate of 0.56: missed and false blockage predictions translate into wrong switches. Fed with the true LoS loss, while still paying the sensing overhead, the planner has a lower mean outage than A5 from 15 dB up (0.05 vs. 0.09 at the reference); the advantage is significant at 20– 30 dB and at the reference (p ≤ 0.049 each, without correction for the number of margins) but not at 15 dB (p = 0.084). Error budget. To attribute the shortfall, we inject our tracker’s errors into perfect tracks (Table II) with a model calibrated on the development seeds: whole-track misses with the measured acquisition delays and outage durations, persistent false tracks with the measured rate, lifetime, and share of fragments next to real targets, position and velocity errors along the lane that are correlated over time, and the classagnostic size rule of our predictor, which gives every vehicle track the size of a bus. Missed tracks are the largest single cause at every margin from 5 dB up, and their persistence matters: dropping detections independently per frame costs only +0.87 s/UE-min at 10 dB, because the tracker bridges single misses. All errors plus the 1 m UE position error reproduce the real planner’s shortfall (+6.91 vs. +6.83 s/UEmin at 10 dB). Accuracy requirements. Fig. 4 adds one error source at a time to perfect tracks; σ is the standard deviation per horizontal axis (along the lane for blockers). Blocker velocity errors up to 1 m/s keep the mean outage below that of A5. For margins of 15–30 dB, linear interpolation of the seedmean difference places the break-even points at a blocker

Placeholder: figs/fig˙error˙sweep.pdf. Planner minus A5 vs. error standard deviation.

Fig. 4. Accuracy requirements: outage of the planner minus that of A5 when one error source is added to perfect tracks, (a) at 20 dB margin and (b) at the reference (95% CI). Below zero, foresight pays off.

position error of 0.12–0.17 m and a UE position error of 0.08– 0.13 m (0.64 and 0.19 m at the reference). These are singleerror thresholds, not guarantees under simultaneous errors; as exploratory evidence (pointwise tests without correction), the advantage remains significant only up to 0.125 m and 0.075 m, respectively, at 30 dB (less at lower margins). With a UE position error of 1 m, no blocker accuracy restores the advantage. A pedestrian next to the UE is about 0.5 m wide, so errors of this size move it in or out of the predicted line of sight; in this deployment and error model, the user’s own position is the binding constraint. Uncertainty-aware decisions. Accounting for the uncertainty does not recover the gain either. Tuned on the tuning seeds (K = 8, no covariance inflation, θ = 0.3 s at 10 dB and 0.1 s at the reference), the uncertainty-aware planner never intervenes at 10, 15, 20, and 30 dB and in at most 6.6% of the epochs otherwise. It is significantly behind A5 (+0.37 and +0.008 s/UE-min at 10 dB and the reference). At 10 dB this deficit is entirely the sensing overhead (+0.37), since it takes A5’s decisions; at the reference, the overhead accounts for +0.002 and the residual difference from overhead-matched A5 (+0.006† ) is not significant. In our experiments, the tuned policy thus rarely acts on the predictions. VI. D ISCUSSION AND L IMITATIONS In the studied deployment, most of the value of foresight lies in pedestrian blockage close to the user. There, both the blocker and the user must be localized to about a decimeter, more than a lamppost radar and meter-level user positioning provide. Making ISAC pay off thus requires precise user positioning and sensing geometries that see the user’s surroundings from close by, such as UE-side or bistatic sensing. The study has limitations: one street canyon with two cells; an O-RAN loop modeled by its report period and delay rather than implemented, e.g., on OpenAirInterface and FlexRIC [21], [22]; monostatic sensing at one radio unit with ideal full duplex and isotropic elements without a back plane, which add mirror ghosts; model B with opaque bodies, without fullwave diffraction, window transmission, or user self-blockage; blockers at constant velocity, which favors prediction, and ideal signaling without link or handover-command failures,

which favors reactive handover, and without conditional handover or Rel-18 L1/L2-triggered mobility (LTM), which would strengthen reactive handover; Shannon rates; a white Gaussian user position error; and an error model calibrated to our tracker. VII. C ONCLUSION Perfect foresight can improve on service-aware reactive handover: a planner that weighs the cost of switching closes most of the remaining outage gap with less than a second of look-ahead. Turning this into practice is difficult, because most of the gain comes from sudden pedestrian blockage right next to the user. To capture it, the network must locate the user and nearby blockers to about a decimeter. With realistic radar tracking and meter-level user positioning, even an uncertainty-aware planner falls back to reactive handover in our deployment and error model. Knowing precisely where the user and its surroundings are is therefore a key requirement for ISAC to help mmWave mobility. ACKNOWLEDGMENT This material is based upon work supported in part by NSF under Awards CNS-2120442 and IIS-2325863, and NTIA under Award No. 51-60-IF007. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the views of the NSF and NTIA. AI assisted with coding and enhancing the text; the authors verified all content. 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] G. R. MacCartney, T. S. Rappaport, and S. Rangan, “Rapid fading due to human blockage in pedestrian crowds at 5G millimeter-wave frequencies,” in Proc. IEEE GLOBECOM, 2017. [3] G. Charan, M. Alrabeiah, and A. Alkhateeb, “Vision-aided 6G wireless communications: Blockage prediction and proactive handoff,” IEEE Trans. Veh. Technol., vol. 70, no. 10, pp. 10 193–10 208, 2021. [4] U. Demirhan and A. Alkhateeb, “Radar aided 6G beam prediction: Deep learning algorithms and real-world demonstration,” in Proc. IEEE WCNC, 2022, pp. 2655–2660. [5] A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais, U. Demirhan, and N. Srinivas, “DeepSense 6G: A large-scale real-world multi-modal sensing and communication dataset,” IEEE Commun. Mag., vol. 61, no. 9, pp. 122–128, 2023. [6] F. Liu, Y. Cui, C. Masouros, J. Xu, T. X. Han, Y. C. Eldar, and S. Buzzi, “Integrated sensing and communications: Toward dual-functional wireless networks for 6G and beyond,” IEEE J. Sel. Areas Commun., vol. 40, no. 6, pp. 1728–1767, 2022. [7] C. Sturm and W. Wiesbeck, “Waveform design and signal processing aspects for fusion of wireless communications and radar sensing,” Proc. IEEE, vol. 99, no. 7, pp. 1236–1259, 2011. [8] 3GPP, “Feasibility study on integrated sensing and communication,” 3GPP TR 22.837, v19.4.0, 2024. [9] ——, “Study on channel model for frequencies from 0.5 to 100 GHz,” 3GPP TR 38.901, Rel. 19, 2025. [10] M. Polese, L. Bonati, S. D’Oro, S. Basagni, and T. Melodia, “Understanding O-RAN: Architecture, interfaces, algorithms, security, and research challenges,” IEEE Commun. Surveys Tuts., vol. 25, no. 2, pp. 1376–1411, 2023. [11] L. Bonati, S. D’Oro, M. Polese, S. Basagni, and T. Melodia, “Intelligence and learning in O-RAN for data-driven NextG cellular networks,” IEEE Commun. Mag., vol. 59, no. 10, pp. 21–27, 2021.

[12] J. Hoydis, F. Ait Aoudia, S. Cammerer, M. Nimier-David, N. Binder, G. Marcus, and A. Keller, “Sionna RT: Differentiable ray tracing for radio propagation modeling,” in Proc. IEEE GLOBECOM Workshops, 2023. [13] J. Hoydis, S. Cammerer, F. Ait Aoudia, M. Nimier-David, L. Maggi, G. Marcus, A. Vem, and A. Keller, “Sionna,” 2022, https://nvlabs.github.io/sionna/. [14] U. Demirhan and A. Alkhateeb, “Radar aided proactive blockage prediction in real-world millimeter wave systems,” in Proc. IEEE ICC, 2022, pp. 4547–4552, doi: 10.1109/ICC45855.2022.9838438. [15] M. Al-Quraan, A. Zoha, A. Centeno, H. Bany Salameh, S. Muhaidat, M. A. Imran, and L. Mohjazi, “Enhancing reliability in federated mmWave networks: A practical and scalable solution using radar-aided dynamic blockage recognition,” IEEE Trans. Mobile Comput., 2024, doi: 10.1109/TMC.2024.3373529. [16] Y. Li, F. Liu, H. Xiong, Z. Wang, Narengerile, and C. Masouros, “Look before switch: Sensing-assisted handover in 5G NR V2I networks,” IEEE Trans. Mobile Comput., vol. 25, no. 6, pp. 8063–8076, Jun. 2026, doi: 10.1109/TMC.2025.3647295. [17] 3GPP, “NR; radio resource control (RRC) protocol specification,” 3GPP TS 38.331, v17.4.0, 2023. [18] ——, “Study on new radio access technology: Physical layer aspects,” 3GPP TR 38.802, v14.2.0, 2017. [19] ——, “Study on coverage enhancement for NR,” 3GPP TR 38.830, Rel. 17, 2021. [20] ——, “NR; physical layer procedures for data,” 3GPP TS 38.214, v17.1.0, 2022. [21] N. Nikaein, M. K. Marina, S. Manickam, A. Dawson, R. Knopp, and C. Bonnet, “OpenAirInterface: A flexible platform for 5G research,” ACM SIGCOMM Comput. Commun. Rev., vol. 44, no. 5, pp. 33–38, 2014. [22] R. Schmidt, M. Irazabal, and N. Nikaein, “FlexRIC: An SDK for nextgeneration SD-RANs,” in Proc. ACM CoNEXT, 2021.

Related documents

• Cramér-Rao Bound Optimization for Joint Beamforming and Mode Selection in RDARS-Assisted ISAC Systems#1156746
• Physics-Informed Unsupervised Learning for Reliable Beamforming in Cluttered MIMO-ISAC#1326420
• Confidence-Guided Adaptive Beam Selection for Vision-Aided Beam Tracking in mmWave Systems#1357535
• Conflict-Aware Beam Scheduling for Ultra-Dense Directional mmWave IoT#139698
• A Two-Stage Shape- and Motion-Aware Framework for mmWave-based Human Mesh Recovery#1488364