Deciding Before Locating: Near-Field Radio Sensing That Indexes Interception Plans Directly

Fast robot planners now pick a trajectory in well under a millisecond. Sensing is what makes them slow. A robot arm blocking a thrown ball can retrieve a precomputed, torque-feasible trajectory in 0.109 ms, yet it spends about 83 ms of a roughly 350 ms reaction window waiting for a stereo camera to report where the ball is. Camera noise and dropouts account for most of its failures. Large antenna arrays at millimeter-wave frequencies offer a different sensor. When the array is physically large compared with the wavelength, nearby objects sit in the "near field," where radio waves arrive as curved spheres instead of flat planes. That curvature encodes both the distance and the full velocity of a target in a single snapshot across the array. The catch is that the best existing near-field motion estimator still takes 0.1–1 s, which is far too slow for a robot.

This note proposes that the robot should not ask the radio sensor where the projectile is. It should ask which precomputed plan to run. We train a fast estimator to map raw near-field channel snapshots directly to a probability distribution over the entries of a precomputed trajectory library. An anytime stopping rule then commits to a plan once the expected success of acting now beats the expected success of waiting for more snapshots. The claim is that at tight decision deadlines, this decision-level inference achieves higher interception success than the usual pipeline of estimating the state and then planning, at equal compute. The reason is that the planner only needs to separate trajectories that lead to different plans. It does not need to resolve the full position-velocity state, whose near-field uncertainty is curved and non-Gaussian.

Where this sits

The planning side supplies the method. Preprocessing-based Kinodynamic Motion Planning Framework for Intercepting Projectiles using a Robot Manipulator divides the workspace into inner and outer "domes." Each pair of cells defines a tunnel a projectile can pass through, and the authors precompute one blocking trajectory per tunnel offline, which takes about five hours of CPU time. Online, the robot looks up a trajectory in 0.109 ms and blocks 78% of real throws and 68.9% in simulation. Think Fast precomputes a tree of dynamically feasible actions instead. It chooses actions in under 10 ms by projecting a Gaussian belief about the target onto an intercept plane and reaches 74% on physical throws, against 63% for a naive baseline. Both papers motivate this proposal, because both report that optical perception is the bottleneck. Both also constrain it. Each assumes a state estimate arrives first, and Think Fast assumes that estimate is Gaussian.

The sensing side motivates the proposal and also defines the gap. Near-Field Integrated Sensing and Communications shows that curvature gives range and angle without synchronized time-of-flight measurement. Its 28 GHz, 65-element array has a Rayleigh distance of 46.73 m, which comfortably covers a 6–12 m throwing range. However, it handles stationary targets with a slow spectral search. Near-Field Motion Parameter Estimation: A Variational Bayesian Approach recovers 2D position and 2D velocity from one array. It splits the array into subarrays, each locally far-field, and fuses their angle and Doppler estimates by message passing over a factor graph. This achieves centimeter and sub-m/s accuracy in 0.1–1 s. That is an order of magnitude faster than grid search but still too slow for a reaction window, and the method never feeds a planner. A Spatio-Temporal-Frequency Transformer Framework for Near-Field Target Recognition shows that a small network can read raw near-field channel tensors with very few frames and subcarriers, reaching 93.9% classification accuracy with a single frame and a single subcarrier. We borrow its input design but replace object category with plan identity as the output. Finally, Safe and Efficient Path Planning under Uncertainty via Deep Collision Probability Fields supplies a methodological precedent: it moves Monte Carlo collision probability into an offline-trained network. We do the same for the sensor-to-decision step.

The closest combination would be "run the near-field motion estimator, then look up a tunnel." This proposal differs in two ways. First, it removes the explicit state from the online loop. Second, it makes the moment of commitment an explicit, deadline-aware decision.

Core Claim

Let $y_{1:t}$ be the channel snapshots received up to slow-time frame $t$. Let $k \in \{1,\dots,K\}$ index the precomputed blocking trajectories, and let $S_t(k',k)$ be the simulated probability that trajectory $k'$ blocks a projectile whose true tunnel is $k$, given that the robot starts executing at frame $t$. A network $q_\phi(k \mid y_{1:t})$ is trained offline on simulated spherical-wave channels to approximate the posterior over tunnels. Online, the robot computes

$$ V_t = \max_{k'} \sum_{k} q_\phi(k \mid y_{1:t})\, S_t(k',k), $$

the expected success of committing now. It commits at the first frame where $V_t$ exceeds a learned estimate of $\mathbb{E}[V_{t+1} \mid y_{1:t}]$. Because $S_t$ shrinks as the remaining execution time shrinks, waiting has a real cost. Uncertainty that does not change the argmax costs nothing. That is the source of the expected advantage over state-then-plan.

Research Program

We will build three pieces.

The first is a channel simulator. It models a planar near-field array and computes exact spherical-wave returns from ballistic projectiles over the throwing volume used by Natarajan et al. We first model projectiles as point scatterers, then as extended scatterers following the extended-target model of Cramér-Rao Bound Optimization for Near-Field ISAC with Extended Targets. The simulator includes optional impairments: phase noise, mutual coupling, and per-subarray gain errors. These are the mismatches that the ISAC and motion-estimation papers name as open problems.

The second is the decision estimator. It uses a lightweight convolutional encoder over antenna-domain slices followed by attention across slow-time frames, in the style of the STF Transformer. It outputs a distribution over the $K$ tunnels at every frame, and its inference cost is profiled on an embedded GPU.

The third is a closed loop in MuJoCo with the dome-tunnel library. Simulated channels drive tunnel selection, and the precomputed B-spline trajectories are executed, so every executed motion keeps its offline torque guarantees. Separately, we compute the induced posterior over tunnels to measure how far it departs from Gaussian, which tests Gap 4 directly.

Baselines

The naive reference has two parts. One is a robot that never moves, which sets the floor. The other is the optical pipeline: a tunnel lookup fed by camera-like measurements with the reported ~83 ms delay and a noise level calibrated so that the closed loop reproduces the published 68.9% simulated block rate from Preprocessing-based Kinodynamic Motion Planning Framework for Intercepting Projectiles using a Robot Manipulator. We cannot model the published camera stack itself, so this calibration is the honest proxy.

The strongest existing competitor is state-then-plan with near-field sensing. We run VMP-NMPE from Near-Field Motion Parameter Estimation on the same simulated channels. We warm-start it across frames and cap its iterations so it fits the same per-frame compute budget as our network. Its Gaussian output feeds both the tunnel lookup and the QMDP action-tree selector of Think Fast. It is the strongest competitor because it uses the same physics, keeps calibrated uncertainty, and has a published accuracy record. Its reported latency is 0.1–1 s and its accuracy is centimeter-level in position and sub-m/s in velocity. No reported block rate exists for it, because it was never run in a closed loop; no reported result was found on Lacuna for VMP-NMPE interception success.

Two ablations isolate the key ingredient. In the first, the same network architecture, trained on the same data, outputs a Gaussian state instead of a tunnel distribution, and that state is then looked up. Any difference between this and our method is due to decision-level output alone. In the second, the same pipeline uses features from a planar-wave model, which isolates the value of curvature. All learned methods get the same training set, the same tuning budget of hyperparameter trials, and the same inference hardware. Beating the competitors means a higher block rate at matched deadline and compute, over at least 200 projectiles per condition with confidence intervals. The minimum bar is to exceed both the 68.9% optical-proxy rate and the VMP state-then-plan pipeline at deadlines of 10 ms or less.

Why not just…?

Why not just use a high-speed or event camera? This is a fair competitor for latency. It does not address depth: stereo depth still requires solving pixel correspondence, and depth noise and blur were the dominant failure sources in both interception papers. The file gives no reported interception numbers for event cameras, so we include a camera proxy with 1 ms latency and the stereo depth noise the optical baseline is calibrated to. If that proxy matches the radio pipeline, the result is that latency rather than depth was the bottleneck, and we will report it as such.

Why not just use a wideband far-field radar with time-of-flight range and Doppler? Such a radar measures range and radial velocity well. Interception, however, needs lateral velocity, and a far-field array sees lateral motion only as slow angular drift. Near-field curvature recovers the full 2D velocity vector from one array, as shown in Near-Field Motion Parameter Estimation. We test this directly with a simulated wideband planar-wave radar of matched aperture and bandwidth. This is a real threat for throws aimed mostly toward the array, and we expect its block rate to fall mainly for oblique throws.

Why not just make the variational estimator faster and keep state-then-plan? This is the most serious alternative. Warm-starting and GPU message passing might bring VMP-NMPE under 10 ms. That is why it appears as the main baseline at equal compute rather than at its published latency. If it matches our method at all deadlines, the decision-level claim fails. The remaining contribution would then be a near-field perception front end for existing planners, and we would say so.

Why not train end-to-end from radio signals to joint torques? Precomputed library trajectories carry offline torque-feasibility guarantees, and a learned torque policy would lose them. The diffusion planner in Keypoint-based Diffusion for Robotic Motion Planning on the NICOL Robot reports dataset bias in its collision behavior without explicit checking layers. Our network chooses among verified plans, so it cannot output an infeasible motion.

Proposed Experiments

The decisive experiment plots block success against decision deadline, swept from 2 to 100 ms, for every method on identical projectile sets drawn from the 6–12 m regime. The proposal succeeds if the tunnel-posterior method dominates both the state-output ablation and budget-matched VMP below roughly 20 ms, with the curves converging at long deadlines, where accurate state recovery makes the two approaches equivalent.

Secondary checks address four questions. First, we measure how far the induced tunnel posterior departs from the Gaussian assumed by Think Fast, and whether that gap predicts where our method wins. Second, we measure robustness to injected coupling and phase errors. Third, we measure sensitivity to the number of frames and subcarriers, mirroring the low-budget results of the STF Transformer. Fourth, we test whether the stopping rule's commit times concentrate on easy throws.

All results are simulation-only, and hardware validation is explicitly out of scope. We state this limitation plainly, since the near-field literature shares it.