{"id":"79206107-9116-4087-9e1d-9540225d8334","arxiv_id":"2506.16216","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Coupling a control JEPA with a wireless JEPA lets a remote controller imagine future states and channel quality in latent space, reducing transmit power by about 50% in a synthetic car-racing setup.","lead":"A new machine learning method predicts both a robot's motion and its wireless channel quality in compressed latent spaces, letting a remote controller schedule transmissions only when channel conditions are good. In simulations, the method cut transmit power by roughly half without hurting control performance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported power savings depend on a simulator in which CSI is a deterministic function of the car position, so the cross-modal predictor has never been tested under stochastic or partially observed channel conditions.","rationale":"The reader identified the determinism of the Sionna channel as the weakest assumption, and my reading agrees. The central mechanism of the paper is that a latent control state learned from images contains enough information to predict future CSI, which is then used to schedule low-power transmissions. In the current simulator this is true by construction because the channel is a deterministic function of the car position, which is itself the control state rendered into pixels. The empirical savings therefore cannot discriminate between 'the model learns a useful cross-modal mapping' and 'the problem is easy because the conditioning variable fully determines the target.' This concern is load-bearing because the claimed contribution is precisely about predicting wireless dynamics from control dynamics under unknown and high-dimensional observations. Adding a stochastic channel component is a concrete, feasible check: if the method still saves power under fading, the contribution is robust; if not, the paper must be reframed as a deterministic-channel proof of concept. This does not invalidate the synthetic experiment as reported, so the appropriate verdict remains a conditional accept: the authors should either provide the stochastic-channel evaluation or explicitly restrict the claim. The reader's verdict is therefore unchanged. Additional weaknesses, such as no released code or error bars, compound the uncertainty but are secondary to the determinism issue for the scientific claim.","tokens_in":9942,"tokens_out":7258,"duration_ms":93960,"concrete_test":"Modify the Sionna pipeline so that the channel is gt = g_deterministic(position_t) + g_stochastic, with independent fast-fading or Rician components drawn per slot and not observable from position alone; retrain the wireless JEPA and power predictor under the same protocol and evaluate for H=10 whether the transmit-power saving relative to the power-agnostic baseline is preserved (e.g., still at least 2 dB) and whether the chosen slots are better than random. Also report the 10-step CSI embedding prediction error. If the saving collapses or slot selection becomes no better than random, the cross-modal conditioning claim is only valid for deterministic channels.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim, 'over 50% transmit power reduction with comparable control performance,' is established only in a synthetic pipeline where the wireless channel is a deterministic function of the device position. Section IV-A states that 'the device's position is replicated from gym to render its wireless channel,' and Appendix A describes a fixed Sionna scene with no fading or other stochastic channel component. In this setting, the latent control state learned from pixels contains, in principle, all information needed to predict future CSI, so the wireless JEPA's cross-modal conditioning is guaranteed to be sufficient. The reported 3 dB / 2.2 dB savings therefore validate a supervised interpolation task, not prediction under genuine channel uncertainty. The paper does not provide quantitative multi-step CSI prediction error, and it does not demonstrate that the mechanism survives when channel realizations depend on unobserved fast fading, interference, or shadowing. This is not an internal inconsistency, but it means the headline claim is narrower than the general resource-management contribution implied by the title and framing. The paper should either add a stochastic fading component to the simulation or explicitly scope the claim to deterministic line-of-sight-like channels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a coupled JEPA framework for remote control of a device whose state is observed as images, with the goal of minimizing uplink transmit power while maintaining control performance. A control JEPA learns latent control dynamics from pixels, and a wireless JEPA learns latent CSI dynamics conditioned on the control JEPA's latent state via cross-modal conditioning. On top of these, the authors train an RL policy (actor-critic) from imagined latent trajectories and a power predictor that estimates the transmit power required for future slots, enabling the controller to schedule transmissions at low-power time slots. The method is evaluated on a synthetic pipeline combining the Car Racing gym environment with Sionna ray tracing, reporting roughly 50% transmit power savings (3 dB) at short horizons and 40% (2.2 dB) at longer horizons relative to a power-agnostic baseline, with comparable normalized control return.","tokens_in":10129,"tokens_out":3433,"duration_ms":41470,"significance":"If the results are robust, the paper makes a useful contribution by showing how self-supervised latent dynamics models (JEPAs) can be extended across modalities (pixels and CSI) for joint control and communication optimization. The idea of distilling control-latent information into a wireless dynamics model is interesting and goes beyond prior work that assumes access to raw velocity or known control dynamics. The paper also provides a fairly complete system architecture and reproducible hyperparameters. However, the significance is tempered by the fact that the central quantitative claim is established only in a deterministic, fully observed synthetic channel scenario, and by the absence of quantitative prediction-error metrics and statistical confidence in the reported gains. The contribution is a proof of concept rather than a demonstrated general resource-management solution.","major_comments":[{"comment":"The wireless channel is rendered deterministically from the car position ('the device's position is replicated from gym to render its wireless channel'), so the latent control state contains, in principle, all information needed to predict CSI. Consequently, the headline claim of 'over 50% transmit power reduction' is demonstrated only in a noiseless, fully observed channel setting; real channels with fast fading, interference, or other unobserved factors are not exercised. Please add a stochastic channel component (e.g., Rician fading or shadowing) or explicitly scope the claim to deterministic line-of-sight-like channels and discuss how the cross-modal predictor would cope with unobserved channel factors.","section":"Section IV-A, Appendix A"},{"comment":"The wireless JEPA's prediction quality is evaluated only qualitatively in the latent space (Fig. 5b), with no quantitative metric such as normalized MSE between predicted and true CSI embeddings, or between predicted and actual required transmit power, as a function of horizon. Since the power savings rely entirely on the accuracy of these predictions, the mechanism is not directly validated. Please report a quantitative CSI-prediction error and, if possible, the resulting scheduling accuracy.","section":"Section IV-B, Fig. 5b"},{"comment":"All results are shown as single curves without error bars, confidence intervals, or multiple random seeds. Deep RL and self-supervised JEPA training are high-variance, and the reported 3 dB/2.2 dB savings and return differences could be within run-to-run variation. Please repeat the experiments with several seeds and report mean and variance (e.g., shaded intervals) for the key comparisons.","section":"Figures 4-6"},{"comment":"The transmit-power savings are measured against a 'power agnostic' baseline that schedules at the end of the prediction horizon without considering power, and against no-prediction schemes. There is no oracle or lower-bound baseline that schedules at the true lowest-power slot (knowing the channel perfectly), so the absolute quality of the learned scheduler is unknown. Adding an oracle scheduler or an offline optimal schedule would contextualize the 50% claim and indicate how much of the gain is due to the learned latent dynamics versus simply exploiting the deterministic geometry.","section":"Section IV-B, baseline comparison"}],"minor_comments":[{"comment":"The symbol SNR is used both as the instantaneous signal-to-noise ratio (SNRt) and as a threshold in constraint (1b); the threshold is never defined. Please introduce a distinct symbol, e.g., SNR_thr, and use it consistently in Eq. (8).","section":"Eq. (1b), Eq. (8)"},{"comment":"The phrase 'saving 3 times its transmit power' is ambiguous; it should read 'reducing transmit power by a factor of three' or 'using one-third of the transmit power'.","section":"Section IV-B, page 6"},{"comment":"'Our model is variation of the recently proposed wireless JEPA' should be 'Our model is a variation of the recently proposed wireless JEPA'.","section":"Section III-B, first paragraph"},{"comment":"The prediction horizon H used in experiments varies from 2 to 15, while Table I lists H = 50 as a general hyperparameter. Please clarify whether 50 is the training horizon used during JEPA training and how the evaluation horizons relate to it.","section":"Appendix A, Table I"},{"comment":"The phrase 'We use gradient coloring to identify local neighborhoods' is vague; please specify the color mapping and what exactly the predicted points represent in the latent space.","section":"Fig. 5 caption"},{"comment":"In the text, reference [11] is referred to as 'Time-series jepa'; consider capitalizing 'JEPA' consistently. Also, in the fourth paragraph of Section I, 'In this work, we propose a novel data-driven framework' is repeated nearly verbatim in the conclusion; consider tightening the introduction.","section":"References and text"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the authors are transparent about their synthetic evaluation. My main concern, in addition to the comments above, is that the abstract and title claim a general resource-management contribution, while the experiments only cover a deterministic channel that is a function of the device position. The authors should either add a stochastic fading component or carefully re-scope the claims in the title/abstract and conclusions. I would also encourage the authors to make their code public, as the hyperparameters are already detailed and re-implementation seems feasible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on wireless control co-design. The new thing here is coupling two JEPA models—one for control dynamics from pixels, one for CSI dynamics—and using the latent control state as conditioning for CSI prediction, then scheduling transmissions at predicted low-power slots. That combination is absent from the authors' prior work [11, 15] and it's a sensible next step. The paper is clearly written and the training objectives are straightforward. The synthetic setup actually closes the loop: Car Racing in gym drives a Sionna ray-tracing scene, the RL policy learns in latent space, and the power predictor schedules transmissions. The convergence comparison against DQN is fine, and the method holds control performance out to 10–12 prediction steps while cutting transmit power versus the power-agnostic baseline. The cross-modal distillation idea—offloading velocity-like information from the control JEPA rather than assuming raw velocity—is the strongest conceptual contribution.\n\nThe soft spots are real but not fatal. The biggest one: in the simulator, CSI is a deterministic function of the car's position. Section IV-A says the position is replicated from gym to render the channel, and Appendix A has no fading or other stochastic component. Under those conditions the latent control state contains, in principle, all information needed to predict CSI, so the wireless JEPA is doing supervised interpolation rather than prediction under genuine channel uncertainty. The reported 3 dB / 2.2 dB savings are therefore narrower than the title and abstract imply. On top of that, Figures 4–6 show single curves with no error bars or seeds; the abstract says 'over 50%' but the actual figure shows 50% only at short horizons and 40% at longer ones; the closest prior work [11] is not compared empirically; and the CSI prediction quality is shown only qualitatively in latent space. No code or data are released.\n\nNone of this invalidates the core idea. The paper is honest about being a synthetic proof of concept, and the mechanism is coherent. But the headline claim is not yet robust. I'd want to see a stochastic fading component, multi-seed results, and a comparison to [11] before trusting the power-saving numbers.\n\nThis is a paper for people working on semantic communication, world models for control, and resource management in closed-loop systems. It deserves a serious referee, not a desk reject. If I were handling it, I'd send it out and expect major revision; the central idea can survive the fixes.","headline":"Coupled control/wireless JEPA proof of concept with a sensible cross-modal distillation idea, but the power-saving headline rests on a deterministic synthetic channel and needs stronger evidence.","tokens_in":10691,"tokens_out":2939,"would_cite":true,"duration_ms":33192,"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":"A remote controller that learns latent dynamics of both the visual task and the wireless channel schedules transmissions at favorable moments, cutting transmit power by about half without losing control performance.","keywords":["self-supervised learning","joint-embedding predictive architecture","cross-modal prediction","latent dynamics","wireless resource management","channel state information","deep reinforcement learning","remote control"],"falsifier":"Render the same car racing scene with a randomized fading component added on top of the position-based channel and rerun the evaluation; if the gap to the power-agnostic baseline drops from the reported 3 dB toward zero, or the normalized return falls below the no-prediction baseline, the central claim that latent control states suffice to predict favorable CSI is refuted.","tokens_in":9719,"feed_emoji":"📡","tokens_out":7055,"duration_ms":72490,"temperature":0.7,"pith_summary":"The paper claims that a remote controller can learn, directly from image frames and channel estimates, two coupled latent models: one for the control dynamics of a device and one for the dynamics of the device's wireless channel. The control model's latent states are used as conditioning for the wireless model, so the controller can imagine the device's trajectory and predict which future time slots will need the least transmit power. A deep reinforcement learning policy and a power predictor are then trained on those imagined latent trajectories, allowing the controller to skip transmissions in poor channel slots while still steering the device. In the paper's simulation, this scheme cuts transmit power by about 50 percent relative to a power-agnostic predictive baseline and by roughly three times relative to always-transmit baselines, while the control return stays comparable. The upshot is that radio resource management for pixel-based remote control can be derived from learned latent dynamics rather than from raw state estimates or a known channel model.","feed_headline":"Latent world model halves transmit power for remote control","feed_subtitle":"Imagining future states and channels in latent space saves about half the transmit power with no drop in control performance.","key_machinery":"The load-bearing mechanism is a pair of coupled Joint Embedding Predictive Architectures (JEPAs), where one model's latent state conditions the other. The control JEPA, built from an image encoder, a recurrent state-space model (RSSM) with a deterministic GRU state and a discrete stochastic latent variable, plus reward and termination predictors, learns control dynamics from pixels without reconstructing the images. The wireless JEPA, built from a channel encoder and a recurrent prediction network, embeds CSI into low-dimensional vectors and predicts future embeddings, with its hidden state conditioned on the control JEPA's latent state; a stop-gradient EMA target prevents representation collapse. This cross-modal conditioning is what lets the controller convert an imagined image trajectory into an imagined channel trajectory, and it is what the power predictor uses to choose cheap transmit slots.","core_discovery":"On the paper's own terms, the central discovery is that future channel state information can be predicted from predicted control states, and that this cross-modal link can be learned in latent space without reconstructing pixels. The authors show that the latent representation produced by a control JEPA, trained with a discrete recurrent state-space model, carries enough information about the device's motion to drive a wireless JEPA's forecasts of CSI embeddings several steps ahead. The controller uses those forecasts to schedule the next transmission at the slot with the lowest estimated required power and to compute the power level that meets the SNR constraint, while an actor-critic policy on the latent states keeps the control task running during unscheduled slots. Results show that this architecture achieves a normalized return close to that of no-prediction baselines up to roughly 10 prediction steps, while using about 50 percent less transmit power than a power-agnostic scheduler at short horizons and 40 percent less at longer horizons.","pith_inferences":["The paper leaves implicit that its power savings are an upper bound for realistic deployments: the simulation renders the channel deterministically from the car's position, so if fast fading, interference, or other factors are added, the wireless JEPA would need extra conditioning inputs or a stochastic channel model to keep the same savings.","A natural extension is to treat the coupled JEPAs as a lightweight digital twin of both the physical and radio environments, using the same latent space to drive control, scheduling, and power allocation in networked autonomy.","A testable next step is to run the same pipeline on real channel traces or in a multi-user setting with shared resources, where the claimed 50 percent power reduction would have to be re-measured against contention and interference rather than free-space ray tracing.","One could add a confidence estimate to the wireless predictions and trigger a fallback transmission when the imagined channel is uncertain, trading a little power to avoid the errors that would otherwise grow at long prediction horizons."],"forward_implications":["The controller can operate for horizons of about 10 time steps without receiving fresh images, using imagined latent states for both action selection and scheduling, while keeping control return comparable to always-transmitting baselines.","Transmit power can be allocated per slot by evaluating predicted CSI embeddings against the SNR constraint, so scheduling and power control emerge from the same latent model instead of separate channel estimators.","Sending a small block of consecutive samples when communicating, here κ=4, improves the control return at long horizons with only about 0.4 dB extra average power compared with κ=3.","The learned CSI embeddings preserve the spatial structure of the environment, meaning the latent wireless model generalizes along the car's trajectory rather than memorizing discrete channel states.","Because the JEPA representations omit redundant pixel detail, policy learning converges faster than model-free DQN, reaching high normalized returns in under half a million training steps."],"supporting_citations":[{"why":"Defines the joint-embedding predictive architecture idea on which both the control and wireless JEPAs are built.","marker":"[7]"},{"why":"Supplies the latent world model and actor-critic training recipe used to learn the control policy from imagined trajectories.","marker":"[9]"},{"why":"Provides the recurrent state-space model (RSSM) that forms the core of the control JEPA.","marker":"[12]"},{"why":"The closest prior work on vision-based remote control with JEPA state prediction, which this paper extends from dropped packets to power-aware scheduling.","marker":"[11]"},{"why":"Introduces the wireless JEPA that the paper adapts by conditioning on latent control states instead of raw velocity.","marker":"[15]"},{"why":"Ray-tracing channel renderer used to generate the CSI data from the car's position in the simulation pipeline.","marker":"[18]"},{"why":"The pixel-based car racing environment provides the observations, actions, and rewards used to train and evaluate the control agent.","marker":"[17]"},{"why":"Model-free deep Q-learning baseline used to compare convergence speed and final control performance.","marker":"[19]"}],"fun_headline_variants":["Latent cross-modal prediction cuts wireless transmit power by half","Imagining future channels in latent space halves transmit power","Control-aware latent CSI forecasts save 50% transmit power","Coupled latent models slash transmit power for remote control","Latent world model schedules transmissions to halve power"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the car's latent control state contains all information needed to predict future channel state, because in the simulation the channel is rendered deterministically from position; if real channels also depend on fast fading, interference, or other unobserved factors, the wireless JEPA forecasts and the claimed power savings would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Latent cross-modal prediction cuts wireless transmit power by half","Imagining future channels in latent space halves transmit power","Control-aware latent CSI forecasts save 50% transmit power","Coupled latent models slash transmit power for remote control","Latent world model schedules transmissions to halve power"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000188,"raw_usage":{"total_tokens":1328,"prompt_tokens":936,"completion_tokens":392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":552,"tokens_out":392,"duration_ms":3915,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:43:42.132628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same car racing scene with a randomized fading component added on top of the position-based channel and rerun the evaluation; if the gap to the power-agnostic baseline drops from the reported 3 dB toward zero, or the normalized return falls below the no-prediction baseline, the central claim that latent control states suffice to predict favorable CSI is refuted.","supporting_citations":[{"cited_title":"Learning latent dynamics for planning from pixels,","cited_arxiv_id":null,"evidence_quote":"Provides the recurrent state-space model (RSSM) that forms the core of the control JEPA."},{"cited_title":"Learning latent wireless dynamics from channel state information,","cited_arxiv_id":null,"evidence_quote":"Introduces the wireless JEPA that the paper adapts by conditioning on latent control states instead of raw velocity."},{"cited_title":"Human-level control through deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Model-free deep Q-learning baseline used to compare convergence speed and final control performance."}],"review_version":1}