{"id":"d887b51c-163e-4c89-8791-f807776237d1","arxiv_id":"2507.20399","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An adaptive wake-up time controller on the MUSIC angle-of-arrival algorithm enables opportunistic 5G SSB-based localization for factory delivery vehicles, cutting onboard energy by 43.09% in simulation and hardware tests while keeping worst-case simulated error below 0.3 m.","lead":"ACCESS-AV is a localization system for factory delivery vehicles that uses the 5G synchronization signals already broadcast by factory base stations instead of dedicated roadside units or extra sensors. It adds a control loop that decides when to run the power-hungry MUSIC angle-estimation algorithm, reporting 43.09% lower onboard energy while keeping simulated errors under 30 centimeters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 43.09% energy saving and <0.3 m worst-case accuracy are not demonstrated against a fixed 240 ms schedule or under any low-SNR/NLOS condition; Section 3.2.2 fixes LoS/Rician at ~40 dB, so the adaptive PID's robustness claim is untested.","rationale":"The energy study on the Jetson platform is genuine evidence, and the comparison against fixed tbase baselines is sensible. The concern is not that the hardware measurements are wrong; it is that the headline 'adaptive' contribution is not isolated. First, no low-SNR or NLOS condition is ever run, so the deferral mechanism central to the robustness claim is untested. Second, no fixed 240 ms schedule is used as a baseline, so the 43% energy saving may be attributable simply to choosing a larger wake-up interval rather than to PID-based adaptation. These are structural experimental omissions rather than internal contradictions. The reader's conditional verdict is appropriate: the paper should add these two controls before the robustness and adaptivity claims are accepted. My concern overlaps with the reader's LoS/high-SNR point but shifts the focus to the missing comparison that would isolate the adaptive benefit.","tokens_in":18655,"tokens_out":7831,"duration_ms":92164,"concrete_test":"Re-run the Section 7.1 simulator with a SNR trace that includes a 5 s NLOS/shadowing segment (e.g., SNR 10 dB, Rician K=0 dB), and in the same sweep replace the PID with fixed WT=240 ms. Compare peak position error and total onboard energy for adaptive, fixed-240 ms, and fixed-tbase variants. If adaptive error exceeds 0.3 m after the outage, or if fixed-240 ms matches its accuracy and energy, then the adaptive robustness and energy claims are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4's 'Robustness via Adaptation' claims that deferring MUSIC at low SNR avoids unreliable estimates, but the accuracy experiments in Section 7.1 never present a low-SNR condition: Section 3.2.2 fixes a LoS/Rician channel, and the SNR trace in Figure 5 stays within 39.2-40.6 dB over the entire trajectory. The PID's SNR-error term is therefore effectively inert, and the behavior that is supposed to provide robustness is not exercised. In addition, the energy comparison in Table 1 and Figure 8 is against fixed hardware-constrained baselines of 65-150 ms, not against a fixed schedule at the adaptive upper bound tmax=240 ms. Since Algorithm 1 clamps WT to [tbase, 240 ms], a non-adaptive 'run MUSIC every 240 ms' policy would by itself cut the number of computations by roughly the observed 43%, making the adaptive mechanism's added value unproven. Finally, the paper does not state what position the vehicle uses during a deferral; if it holds the last fix, up to 240 ms of motion (about 1 m at 15 km/h) plus odometry error must be absorbed, and this error source is absent from the model. Section 9 lists NLOS as future work, confirming the limitation is acknowledged but not analyzed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ACCESS-AV, a localization framework for autonomous delivery vehicles (ADVs) in smart factories that uses 5G Synchronization Signal Blocks (SSBs) from existing remote radio heads (RRHs) instead of dedicated roadside units. An onboard MUSIC algorithm estimates angle-of-arrival (AoA), and a PID controller adaptively schedules wake-up times (WT) based on SNR and velocity to reduce energy consumption. The authors evaluate accuracy in a Python simulation of a 300 m track and energy on an NVIDIA Jetson AGX Xavier across eight power modes. They report worst-case localization error below 0.3 m and average energy savings of 43.09% relative to fixed-interval baselines of vanilla MUSIC, ESPRIT, and Root-MUSIC, plus roughly a 130x reduction in onboard sensor cost.","tokens_in":19026,"tokens_out":4411,"duration_ms":42015,"significance":"The main strength is the energy study, which uses real hardware and compares the adaptive controller against fixed baselines; this part is not circular. The system concept—opportunistic use of SSBs and adaptive computation-scheduling—is timely and relevant for edge-autonomous systems. If the accuracy and energy claims held under realistic factory radio conditions, the contribution would be useful. However, the accuracy evidence currently relies on a nearly constant, high-SNR, LoS-only simulated channel, and the energy comparison does not isolate the benefit of adaptation over simply using the maximum allowable wake-up time.","major_comments":[{"comment":"The simulation always assumes LoS with Rician fading and the SNR trace stays within 39.2–40.6 dB. Consequently, the SNR-error term in Algorithm 1 is effectively constant, and the robustness behavior claimed in Section 4 (deferring MUSIC at low SNR) is never exercised. Add experiments that vary SNR across a realistic range (e.g., 0–30 dB) and include NLOS or shadowing; report accuracy and energy under those conditions.","section":"Section 3.2.2 and Section 7.1, Figure 5"},{"comment":"The energy savings of 43.09% are computed relative to the hardware-constrained baselines (65–150 ms), not a fixed schedule at the adaptive upper bound tmax=240 ms. Since Algorithm 1 clamps WT to [tbase, 240 ms], a non-adaptive MUSIC executed every 240 ms would already cut the computation count by a similar factor. Compare the adaptive policy against fixed schedules at both tbase and tmax to demonstrate that adaptivity itself provides energy gain beyond simply increasing the wake-up interval.","section":"Section 7.2 and Table 1"},{"comment":"The paper does not state how the vehicle position is propagated during a deferred wake-up (i.e., when WT > tbase). If the last MUSIC fix is held until the next update, the vehicle can move approximately 1 m at 15 km/h during 240 ms, and that uncompensated motion is not in the error model. Specify the propagation method (e.g., odometry/kinematic model) and include its noise when computing the reported errors.","section":"Section 6 and Algorithm 1"},{"comment":"Accuracy results are reported as averages over 100 runs without error bars or worst-case statistics. In a Rician fading channel, MUSIC errors vary run-to-run; provide distributions (e.g., min/max or 95th percentile) to support the repeated 'worst-case below 0.3 m' claim.","section":"Section 7.1"},{"comment":"The PID gains (0.75, 0.08, 0.195), weights (0.6, 0.4), and tmax=240 ms are hand-chosen for this track. No sensitivity analysis is provided, so it is unclear whether the claimed accuracy/energy trade-off is robust to these choices. A sensitivity study (e.g., varying Kp, w_snr, and tmax by ±50%) would support generalizability.","section":"Section 6"}],"minor_comments":[{"comment":"The statement 'fading is multiplied with a standard deviation of 3 dB' is ambiguous for Rician fading; Rician fading is parameterized by the K-factor and total power, not a dB standard deviation. Clarify the K-factor and how shadowing is (not) modeled.","section":"Section 3.2.2"},{"comment":"The sentence 'the worst-case error is 0.8 m on average across vanilla MUSIC, ESPRIT, and Root-MUSIC' is unclear; 'worst-case on average' is contradictory. State the metric precisely.","section":"Section 7.1"},{"comment":"The raw data and smoothed SNR are nearly flat; state explicitly that this is the best-case scenario or add a lower-SNR scenario to represent realistic factory conditions.","section":"Figure 5"},{"comment":"The cost comparison of a $30.71 receiver versus a $4,000 LiDAR is presented as a 130x reduction, but the replacement of LiDAR by a single 5G receiver assumes the localization task remains equally safe; the comparison would be stronger if it included the full sensor suite cost of a typical ADV.","section":"Section 7.4"},{"comment":"The PID controller updates the integral term with a fixed Δt, but the wake-up time varies; clarify whether Δt is the SSB period (20 ms) or the current WT.","section":"Algorithm 1"},{"comment":"Modes 1, 2, and 3 have identical CPU frequencies but the table lists the same base wake-up time for each; explain the differences between these modes (e.g., other DVFS domains or memory settings).","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The adaptive system concept and the real-hardware energy measurements are worth publishing, but the accuracy experiments need to be substantially expanded before the paper is acceptable. The current evidence does not support the headline robustness claim under low-SNR or NLOS conditions. I recommend major revision with the required additions above; the paper may also benefit from an explicit limitations section beyond the future-work paragraph in Section 9."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes ACCESS-AV: use 5G SSBs already transmitted in a private factory network as opportunistic AoA beacons, run MUSIC on an onboard Jetson, and gate how often MUSIC runs with a PID controller that reacts to SNR and velocity. That is a legitimate engineering idea. The energy measurements are real hardware measurements on the Jetson, the cost argument for dropping dedicated RSUs is sensible, and the design-space exploration over power modes is a nice touch. The integration is new as a combination, even though every building block is known.\n\nThe problems are in the evaluation. The SNR trace in Fig. 5 is essentially flat at 39-40.6 dB over the whole run, so the SNR term in the PID controller is never exercised. That makes the 'robustness via adaptation' claim hard to accept. If the system is meant to defer MUSIC under low SNR, this paper never shows what happens when SNR actually drops. Second, and more important, the energy comparison is against fixed baselines of 65-150 ms, which are the hardware-constrained lower bounds. There is no fixed 240 ms baseline. Since the adaptive WT clamps to [tbase, 240 ms], a non-adaptive policy that just runs MUSIC every 240 ms would cut the number of computations roughly as much as the reported 43% savings. Without that comparison, the added value of the PID controller is unproven. The accuracy simulation is also done only in idealized LoS/Rician with no low-SNR condition and no error bars; the sub-30 cm claim is a simulation average, not a measured result. Finally, the paper never states what happens to position estimation during the deferral interval; if the vehicle holds the last fix or propagates with odometry, the error from up to 240 ms of motion and potential drift is not analyzed. The paper does list NLOS as future work, which confirms the limitation is acknowledged but not addressed.\n\nNone of this makes the paper worthless. The system concept is practical, and the hardware energy numbers are honest. But the headline claims outrun the evidence. The paper needs a fixed tmax baseline, a low-SNR or NLOS case, and an explicit error-accumulation model for the deferral period. With those, it could be a solid systems paper.\n\nWorth a serious referee: yes, because the idea and the hardware work deserve scrutiny, but the referee should push hard on the missing baseline and the untested adaptivity.\n\nRegards.","headline":"A sensible integration of 5G SSB-based localization with PID-gated MUSIC scheduling, but the headline energy and accuracy claims are not actually demonstrated against the right baseline or under the conditions the adaptivity is meant for.","tokens_in":19519,"tokens_out":4077,"would_cite":false,"duration_ms":43089,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive wake-up timing for the MUSIC localization algorithm can hold delivery-vehicle error under 0.3 m while cutting onboard energy by 43.09%.","keywords":["5G","Synchronization Signal Block","MUSIC","Angle of Arrival","Vehicle-to-Infrastructure","Autonomous delivery vehicle","Localization energy efficiency","Adaptive PID wake-up time"],"falsifier":"Repeat the same adaptive MUSIC experiments in a channel simulator that adds realistic factory shadowing, non-line-of-sight intervals, and intermittent SSB reception, then check worst-case localization error and energy per trajectory; if worst-case error exceeds 0.3 m or energy saving falls far below 43.09% at the same power modes, the central claim is not supported.","tokens_in":18454,"feed_emoji":"🚚","tokens_out":8415,"duration_ms":81417,"temperature":0.7,"pith_summary":"The paper sets out to show that autonomous delivery vehicles in a 5G-equipped factory can locate themselves using only the synchronization signals the network already broadcasts, with no dedicated roadside units and no extra sensors. Its system, ACCESS-AV, still runs the angle-of-arrival estimator MUSIC, but replaces fixed periodic execution with an adaptive wake-up time: a PID controller decides when to run MUSIC based on current signal-to-noise ratio and vehicle speed. The paper's central result is that this gating cuts average onboard energy by 43.09% compared with fixed-schedule MUSIC, ESPRIT, and Root-MUSIC baselines, while simulated worst-case localization error stays below 0.3 m. If true, the result matters because it would extend delivery-vehicle battery life and make factory automation cheaper and more sustainable without new infrastructure.","feed_headline":"Delivery robots trim positioning energy 43%, keep error under 30 cm","feed_subtitle":"A control loop runs the heavy angle-estimation step only when signal quality and speed make it worthwhile.","key_machinery":"The mechanism is the adaptive wake-up time (WT): a PID controller with proportional, integral, and derivative gains takes a weighted combination of normalized SNR error and velocity error, updates an integral and a derivative term, and outputs a correction to a base interval, clamped to [$t_{\\text{base}}$, $t_{\\text{max}}$]. $t_{\\text{base}}$ is the platform's hardware-constrained minimum MUSIC interval; $t_{\\text{max}}$ is set by a user-chosen position-uncertainty bound (240 ms for a 1 m margin at 15 km/h). This single timing decision carries the whole argument: it decides when the expensive eigenvalue-decomposition step runs, skipping estimates under poor signal or motion conditions that would produce unreliable angles and wasted energy.","core_discovery":"The central claim is that an adaptive communication-computation code-sign can make MUSIC-based V2I localization both accurate and energy-efficient: instead of computing at every 20 ms SSB arrival or at a fixed hardware-limited rate, the vehicle computes only when the PID controller's weighted SNR and velocity error says the estimate is worth computing. In software simulation of a 300 m factory track under a line-of-sight Rician channel, this scheme kept worst-case position error within 0.3 m in both X and Y directions across all eight power modes of the embedded platform, while hardware energy measurements across the entire trajectory showed 43.09% average savings against the fixed baselines. The same design also removes the need for roadside units and yields a 130-fold onboard sensor cost reduction.","pith_inferences":["A natural extension is to gate other computation-heavy perception steps with the same PID-style SNR and velocity wake-up logic, since the energy-saving lever is deciding when to compute, not the MUSIC algorithm itself.","If the channel model were changed from always-line-of-sight Rician fading to include shadowing and intermittent SSB reception, the controller would likely defer more computations at low SNR, and the 0.3 m worst-case bound would need to be re-validated; the paper does not make that claim.","The 130x sensor cost comparison counts only the radio receiver versus a LiDAR and ignores the embedded compute board, so the full cost picture for a factory operator is less favorable than the headline ratio.","The design-space view of accuracy versus energy per power mode suggests the same framework could be used to auto-select a power mode at runtime based on current energy budget and accuracy requirement, which the paper does not implement."],"forward_implications":["Factories with private 5G can get vehicle localization without buying roadside units, since the SSBs already transmitted by remote radio heads serve as the signal source.","Adaptive scheduling keeps worst-case error under 0.3 m across all tested power modes, so operators can pick a low-power mode and still meet accuracy needs.","Energy savings scale: a single vehicle in the highest-power mode saves about 398.84 kJ per 24-hour day, and a 100-vehicle fleet about 39.884 MJ per day.","Fixed-schedule algorithms accumulate position error over time, while the adaptive controller defers computation under poor conditions and thereby avoids those cumulative errors."],"supporting_citations":[{"why":"Defines the MUSIC algorithm and its eigenvalue-decomposition cost, the computation ACCESS-AV schedules adaptively.","marker":"[12]"},{"why":"Establishes V2I AoA localization as more accurate than GPS, motivating the localization approach.","marker":"[16]"},{"why":"Describes SS/PBCH (SSB) design in 5G NR, supporting the opportunistic use of periodic SSBs.","marker":"[20]"},{"why":"Supplies the Rician fading plus line-of-sight channel model used to generate the simulated SNR profile.","marker":"[35]"},{"why":"Provides the PID control structure used for adaptive wake-up time calculation.","marker":"[42]"},{"why":"Lists the eight power modes and CPU frequencies used for hardware energy measurements.","marker":"[44]"},{"why":"Provides roadside-unit capital and operating costs used in the infrastructure cost-saving analysis.","marker":"[46]"}],"fun_headline_variants":["Smart factory ADVs cut localization energy 43% without losing accuracy","Adaptive SSB-based V2I localization: 43% energy saving, <30cm error","Compute only when needed: ADVs save 43% power in 5G localization","Energy-aware MUSIC scheduling cuts ADV localization power 43%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The stated accuracy and energy numbers depend on the simulation's assumption that the vehicle always keeps a clear, strong line-of-sight signal from the 5G tower with only mild fading; on a real factory floor with shadowing, stronger multipath, and interrupted synchronization broadcasts, the controller's decisions would differ and the 0.3 m claim would not follow.","fun_headline_variants_meta":{"raw":{"variants":["Smart factory ADVs cut localization energy 43% without losing accuracy","Adaptive SSB-based V2I localization: 43% energy saving, <30cm error","Compute only when needed: ADVs save 43% power in 5G localization","Energy-aware MUSIC scheduling cuts ADV localization power 43%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000727,"raw_usage":{"total_tokens":3253,"prompt_tokens":940,"completion_tokens":2313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2228}},"tokens_in":556,"tokens_out":2313,"duration_ms":14956,"temperature":1.0,"reasoning_tokens":2228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:44:18.413346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the same adaptive MUSIC experiments in a channel simulator that adds realistic factory shadowing, non-line-of-sight intervals, and intermittent SSB reception, then check worst-case localization error and energy per trajectory; if worst-case error exceeds 0.3 m or energy saving falls far below 43.09% at the same power modes, the central claim is not supported.","supporting_citations":[{"cited_title":"Music-lite: Efficient music using approximate computing: An ofdm radar case study","cited_arxiv_id":null,"evidence_quote":"Defines the MUSIC algorithm and its eigenvalue-decomposition cost, the computation ACCESS-AV schedules adaptively."},{"cited_title":"A localization algorithm based on v2i communications and aoa estimation","cited_arxiv_id":null,"evidence_quote":"Establishes V2I AoA localization as more accurate than GPS, motivating the localization approach."},{"cited_title":"Irukulapati, Hai Wang, and Henrik Sahlin","cited_arxiv_id":null,"evidence_quote":"Describes SS/PBCH (SSB) design in 5G NR, supporting the opportunistic use of periodic SSBs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Rician fading plus line-of-sight channel model used to generate the simulated SNR profile."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PID control structure used for adaptive wake-up time calculation."},{"cited_title":"https://docs.nvidia.com/jetson/archives/r35.4.1/DeveloperGuide/text/SD/ PlatformPowerAndPerformance/JetsonXavierNxSeriesAndJetsonAgxXavierSeries.htmll","cited_arxiv_id":null,"evidence_quote":"Lists the eight power modes and CPU frequencies used for hardware energy measurements."},{"cited_title":"Tonguz and Wantanee Viriyasitavat","cited_arxiv_id":null,"evidence_quote":"Provides roadside-unit capital and operating costs used in the infrastructure cost-saving analysis."}],"review_version":2}