Predicting and Steering Around mmWave Blockage from Recorded Trajectories

A mobile robot that depends on a millimeter-wave (mmWave, roughly 30–300 GHz) radio link has a fragile lifeline. These signals travel almost like light. When a person, pillar or another robot steps into the direct path, received power collapses and the link drops. Two remedies exist, and today they are designed separately. The first is a reconfigurable intelligent surface (RIS), a wall-mounted panel of many small, cheap elements whose reflection phase can be switched in software, so it can bounce the signal around an obstacle. The second is to have the robot itself move or slow down to stay out of shadowed regions. Existing combinations of the two plan offline. They assume a perfectly known channel and a fixed radio map, or they re-run a ray-tracing simulator so slowly that moving obstacles defeat it. This proposal asks whether a robot can instead learn, directly from its own recent history of motion commands and measured signal power, what will happen to the link over the next fraction of a second under each RIS setting. It would then choose its velocity and the RIS setting together, with no channel model, no radio map, and no hand-tuned controller weights.

The tool is data-enabled predictive control (DeePC). DeePC is a receding-horizon controller that, instead of identifying a model, predicts future outputs as linear combinations of past recorded trajectories stacked in a data matrix. Two things are new here. We treat the received power as one more output of the robot's behavior. We also keep a separate small data matrix for each entry of the RIS's discrete reflection codebook, so the RIS choice becomes a discrete scheduling decision inside the controller. The regularization weights that make DeePC work under noise are notoriously sensitive. We tune them by Bayesian optimization on closed-loop episodes rather than by hand or through a surrogate plant model. Our approach succeeds only if it beats a decoupled reactive system on outage-free task completion. That reactive system picks the best RIS beam from measured power and plans motion with a standard controller.

Core Claim

Blockage by moving bodies unfolds over the time it takes a body to cross the beam. That is much slower than fast fading, which the radio's physical layer already handles. On that slower timescale, the joint signal formed by the robot's velocity commands and its received power, conditioned on a fixed RIS configuration, is predictable enough over short horizons for a regularized, piecewise data-matrix predictor to give useful warning before an outage. If this holds, a single predictive controller that enumerates a small RIS codebook can pre-emptively switch reflection paths and adjust speed. It should keep the link alive more often than reacting after the power drop, with lower detour cost than static radio-map planning, and without the pilot overhead of estimating the cascaded base-station-to-RIS-to-robot channel.

The claim has a real risk that we state up front. Received power in decibels is not a linear function of anything, and Willems' lemma guarantees the data-matrix representation only for linear time-invariant systems. Our bet rests on short horizons, scheduling by codebook entry, and the robustness given by regularization. It is the same bet gain-scheduled DeePC makes for nonlinear ship dynamics. If the bet fails, the predictor's error at blockage onset will show it directly. We test that first.

Relation to Prior Work

The closest system work motivates the gap rather than solving it. Intelligent Reflecting Surface Enhanced Indoor Robot Path Planning: A Radio Map-Based Approach jointly optimizes RIS phases and an indoor robot's shortest path over a precomputed channel-gain or rate map. It reports a 2.6 dB gain with continuous phases over no RIS, and near-continuous performance with 2–3 bit phases. It assumes perfect channel knowledge everywhere and a static obstacle set, and it plans entirely offline. We keep its use of the RIS as a path-restoring actuator and its finding that coarse 2–3 bit codebooks suffice, which is what makes our enumeration over codebook entries tractable. We drop the map and the perfect channel knowledge. Network Digital Twin for 5G-Enabled Mobile Robots updates a radio map online from LiDAR mapping plus ray tracing on a real robot and shows radio-aware paths keep higher received power. It only routes around dead zones, cannot reshape the environment, and is limited by ray-tracing latency. We replace simulation with a measured-trajectory predictor and add the RIS as a decision. On the radio side, Channel Estimation and Transmission for Intelligent Reflecting Surface Assisted THz Communications acquires the cascaded channel of a passive surface through hierarchical beam training for static endpoints. It constrains our design: estimating that channel for a moving robot is exactly the overhead we avoid by working at the level of a codebook entry's measured power. The multimodal beam predictors Multi-Modality Sensing in mmWave Beamforming for Connected Vehicles Using Deep Learning and When Vision-Language Model (VLM) Meets Beam Prediction: A Multimodal Contrastive Learning Framework predict base-station beams from camera, LiDAR and GPS. They are open-loop classifiers evaluated largely in line-of-sight, with no RIS and no coupling to motion. They supply a strong learned-predictor baseline, not the method.

The control side supplies the method. DeePC itself (Jeremy Coulson et al.) and its stability-guaranteed variant Data-Driven Model Predictive Control with Stability and Robustness Guarantees give the predictor and show that without regularization the noisy closed loop diverges on the four-tank benchmark. Regularized and Distributionally Robust Data-Enabled Predictive Control proves the regularizers correspond to a Wasserstein robustness radius and shows a U-shaped cost in that radius, but leaves the radius to the engineer. Gain-Scheduled Data-Enabled Predictive Control: A DeePC Approach for Nonlinear Systems splits data into local matrices indexed by a measured scheduling variable, with large speed and memory savings. We reuse this idea with the RIS codebook index as the scheduling variable. The difference is that our schedule is a decision the controller makes, not a passively measured quantity. For tuning, DeePC-Hunt differentiates through DeePC to tune its regularizers but needs an approximate plant model offline. That model is unavailable here, since there is no trustworthy model of the cascaded channel under moving blockers. Automatic LQR tuning based on Gaussian process global optimization tunes controller weights by Bayesian optimization on real closed-loop trials, improving 31.9% over model-derived gains in 20 trials. We adopt its trial-level black-box tuning for DeePC's regularizers. Its seconds-per-step acquisition cost is irrelevant because we tune per deployment, not per channel coherence interval.

Research Program

The controlled system is a differential-drive robot moving through an indoor space toward a goal. It is served by one mmWave access point and one RIS with a $K$-entry discrete codebook. The input $u_t$ is the robot's velocity command. The output $y_t$ is its odometry pose plus received power in dB under the currently active RIS entry. Data collection runs exploratory episodes with persistently exciting velocity inputs while the RIS cycles through its entries and pedestrians or other robots walk through the scene. For each entry $k$ we build a block data matrix $H^{(k)}$ with past and future input and output blocks $U_p^{(k)},Y_p^{(k)},U_f^{(k)},Y_f^{(k)}$. Following DeePC-GS, each local matrix is kept small by slicing data near the current region.

At each control step, for every candidate entry $k$ we solve

$$\min_{g,\sigma,u,y}\; \sum_{i=1}^{N}\Big(\|p_i-p^\star\|^2_Q + \|u_i\|^2_R + \rho\,[\,\bar P - P_i\,]_+^2\Big) + \lambda_g\|g\|_2^2+\lambda_\sigma\|\sigma\|_1$$

subject to $U_p^{(k)}g=u_{past}$, $Y_p^{(k)}g=y_{past}+\sigma$, $U_f^{(k)}g=u$ and $Y_f^{(k)}g=y$, with velocity limits on $u$. Here $N$ is the prediction horizon, $p_i$ and $P_i$ are the predicted pose and power at step $i$, $p^\star$ is the next waypoint, $\bar P$ is the minimum power for the required link rate, $\rho$ weights link-margin violations, $Q,R$ weight tracking and effort, and $\lambda_g,\lambda_\sigma$ are the regularizers that set the robustness radius. The $K$ quadratic programs are independent and run in parallel. The controller applies the first input and the RIS entry with the lowest cost, adding a switching penalty to prevent chattering. The past window is then refilled with the measured power under the newly chosen entry. Power for non-active entries is not observed, so the cost of a switch includes the uncertainty from that entry's stale past window. Quantifying how much this hurts is part of the study.

Tuning removes the manual step. A Gaussian-process Bayesian optimizer searches over $\lambda_g$, $\lambda_\sigma$, $\rho$ and the horizon, scoring each candidate by the episode-level outcome: task completion without link outage, minus a travel-time penalty. All tuning runs on closed-loop episodes in the target environment, with no plant model. We then test one cheap online refinement: raising $\lambda_g$ when the data-consistency residual $\|\sigma\|$ spikes, which signals that the recent window has left the span of the recorded data, as during a novel blockage. This tests whether the U-shaped optimum of the regularization paper moves with blockage conditions enough to justify scheduling it.

The testbed is a reconstruction of the indoor grid scenario of the radio-map path-planning paper: same geometry style, 2–3 bit RIS codebook, access point and obstacle layout. We extend it with moving blockers modeled as absorbing bodies on random walking paths, and with a geometric mmWave channel that includes the RIS reflection path. Received-power sequences from DeepSense 6G (V2I scenarios 31–34) serve a narrower purpose: checking on real measured mmWave power over time whether short-horizon data-matrix prediction of power is competitive with learned predictors. The file reports these scenarios as mostly line-of-sight, so they test prediction of power dynamics, not blockage recovery. Real-world blockage validation would need new measurements, for example on a platform like the SUMMIT-XL testbed fitted with an RIS. We list that as a follow-up, not a claim.

Baselines

The naive reference is a shortest-path planner with a fixed RIS entry chosen once at the start for the best average power. It shows how often blockage breaks an uninformed robot.

The strongest existing method on this question is the offline radio-map planner of Intelligent Reflecting Surface Enhanced Indoor Robot Path Planning: A Radio Map-Based Approach, run with jointly optimized RIS phases and given perfect static channel knowledge, a generous advantage. It is the only published system that co-designs RIS configuration and robot path against dead zones. Its reported figure is a 2.6 dB gain from continuous RIS phases over no RIS. For outage-free completion under moving blockers, there is no reported result found on Lacuna for this or any method. We run it on the identical layouts and blocker seeds and give it the same discrete codebook.

The baseline we consider most dangerous is a decoupled reactive system. A standard model-based MPC tracks the shortest path using the known robot kinematics. A separate rule switches the RIS to whichever entry gives the highest measured power after periodic sweeps, at the same sweep budget our data collection uses. It isolates the value of prediction and coupling. Beating it means a statistically significant reduction in outage episodes at equal or lower travel time, across at least the faster blocker speeds.

A learned-predictor variant uses a recurrent network, trained on the same exploratory data, to predict per-entry power over the horizon, embedded in a sampling-based MPC with the same cost. It gets the same Bayesian-optimization budget for its hyperparameters. Beating it would mean matching outage rates with less training data or lower solve time. Matching it on outages alone would still favor the simpler, regularized predictor only if data or compute is lower.

Three ablations isolate the key ingredients. A single global data matrix without codebook scheduling, with the RIS entry appended as an input, tests whether scheduling matters. A version with hand-picked regularizers from a coarse grid by an expert, versus Bayesian-optimized ones at the same episode budget as the 20-trial regime in the LQR tuning paper, tests the tuning claim. A version that receives the same predictions but only reacts once power has already crossed $\bar P$ tests whether lead time is the source of the gain.

Why not just…?

Why not just switch the RIS reactively and keep a standard motion controller? This is a real threat. If blockers move slowly relative to how fast a sweep can find a new entry, reacting after the drop costs only brief outages, and joint prediction adds complexity for nothing. Nothing in the file settles this. The reactive baseline and the predict-then-react ablation are designed to answer it. We sweep blocker speed and sweep period, and we will report the regime boundary where prediction stops helping. If there is no regime where it helps, the idea fails.

Why not just build a radio map, or a digital twin, and plan around dead zones? Static maps assume fixed obstacles and perfect channel knowledge (radio-map path planning). The online twin of Network Digital Twin for 5G-Enabled Mobile Robots is limited by ray-tracing latency and only routes around shadows. Our comparison gives the map baseline perfect static knowledge so that any remaining gap is attributable to moving blockers. If the map planner matches us when blockers are sparse, that is an honest and expected result.

Why not just learn a neural predictor of the link, as the beam-prediction papers do? Multimodal classifiers reach 98.19% top-13 beam accuracy (Multi-Modality Sensing in mmWave Beamforming for Connected Vehicles Using Deep Learning) and a 0.9016 distance-based accuracy score (VLM beam prediction) on DeepSense 6G. Those results come from large models trained on large datasets, in mostly line-of-sight scenes, and they predict only the base-station beam, not RIS paths or motion consequences. The recurrent-predictor MPC baseline tests whether a neural predictor beats the data-matrix one at equal data. If it does by a wide margin, the contribution narrows to the tuning and scheduling machinery, and we would say so.

Why not just tune DeePC with DeePC-Hunt? It needs an approximate plant model to differentiate through (DeePC-Hunt). Here the uncertain part is the channel under moving blockers, which has no trustworthy model. Its own result, 66% versus 22% success for mismatched MPC on the rocket benchmark, shows how much tuning matters. It does not show how to tune without a model. If a crude geometric channel surrogate turns out good enough for DeePC-Hunt, that would be informative, and we include it as a secondary comparison.

Decisive Checks

The first check comes before any control. On simulated blockage episodes and on DeepSense 6G power sequences, we measure the data-matrix predictor's power error over the horizon, conditioned on time-to-blockage-onset. The idea is viable only if error in the window just before onset is small enough to flag the coming outage before the reactive baseline would detect it. If the error spikes only after onset, the predictor sees blockage no earlier than a reactive rule and the program stops there.

The main experiment compares all baselines on outage-free task completion rate, total outage time, travel time, and per-step solve time. It runs over a factorial sweep of blocker density, blocker speed, and RIS codebook resolution (1, 2 and 3 bits), with many random seeds and confidence intervals on every comparison.

The tuning experiment plots closed-loop cost against the regularization weight for several blockage conditions. It tests whether the U-shaped optimum shifts between conditions, and whether Bayesian optimization finds the minimum within a budget of about 20 episodes while an expert grid at the same budget does not.

The last check is compute. The $K$ parallel quadratic programs must fit within one control period at the blockage timescale. We expect codebook slicing to deliver this, judging from the DeePC-GS speedups. If it does not, the codebook must be pruned and we will report the trade-off.