When a robot hand closes around an object, its metal fingers and the object itself can cut a wrist-mounted millimetre-wave or terahertz radio link. These are high-frequency bands that travel almost like light and do not bend around obstacles. The cut tends to happen at the moment the touch sensors start producing the data the controller most needs. The fallback route is a reflecting surface on a nearby wall, and it is much weaker, because its signal loss multiplies over two hops instead of adding. That weakness is the "double-fading" effect. Our proposal starts from one observation: this outage is not random noise. It follows from the hand's own joint angles and contact state, which the robot already knows. We propose to predict these self-inflicted outages from proprioception and tactile onset tens of milliseconds ahead. We then use the prediction to switch beams early and to send a tactile representation sized to whatever rate the reflected route can actually carry.
In two sentences: closing a hand predictably blocks its own high-frequency radio, and the backup path through a reflecting surface carries far less data. A robot that forecasts the blockage from its own body state can switch paths and shrink its tactile message before the link fails, so control does not stall at contact.
Core Claim
Prior wireless work treats blockage as caused by the environment and handles it reactively. Prior tactile compression work assumes an ideal channel. We claim that for hand- or wrist-mounted transceivers, a large share of blockage events is self-blockage that can be forecast from body state. We also claim that the right response is a joint decision over path and representation, not a better codec alone or a better beam search alone. Our concrete hypothesis is that a small predictor using joint angles, hand aperture and taxel onset forecasts link state better than one using only past channel measurements. We further hypothesize that a manipulation policy fed tactile data over the predicted-and-adapted link keeps most of the success it has on a perfect channel, while reactive schemes lose the contact-critical frames.
The rate budget on the rescue path is what makes representation choice matter. For a reflected path with transmitter-to-surface distance $d_1$, surface-to-receiver distance $d_2$, $N$ surface elements, and path-loss exponent $\alpha$, received power for a passive surface scales roughly as
$$P_{\text{passive}} \propto \frac{N^2}{d_1^{\alpha} d_2^{\alpha}},$$
rather than with $(d_1+d_2)^{-\alpha}$. This form is the product-of-distances model of Downlink Analysis for Reconfigurable Intelligent Surfaces Aided NOMA Networks, which models RIS-aided cellular coverage with stochastic geometry. We reuse that model; we do not extend the geometry. Active RIS vs. Passive RIS: Which Will Prevail in 6G? shows that adding amplifiers to each surface element largely removes this penalty. They report about 40 dB higher SNR at $N=256$ and estimate that a passive surface would need about 2.5 million elements to compete. That result constrains our design. Passive and active surfaces give rescue rates that differ by orders of magnitude, so the tactile message must come in several sizes. The two surface types will select different rungs.
Research Program
The work has three parts. The first is a self-blockage simulator driven by real robot motion. We take recorded trajectories from BAAI-Humanoid/DECO-50: 50 hours of teleoperated bimanual manipulation on a Unitree H1-2 with two Inspire hands carrying 1,062 taxels. We replay the recorded joint states through a geometric hand-and-arm model. We place candidate transceivers at the wrist and back of hand, an access point overhead, and one or more reflecting surfaces on workcell walls. A ray is marked blocked when it intersects hand, arm or grasped-object geometry. Blocked links get the 20 to 40 dB attenuation reported for quasi-optical shadowing in the file's blockage literature, and the reflected path follows the product-of-distances model in passive and active variants. The output is a time-aligned trace of link state and achievable rate for every DECO-50 frame. No measured self-blockage traces for robot hands exist in the material we surveyed. The simulator is therefore an assumption we state openly, and we test its sensitivity by varying the attenuation, the transceiver placement and the surface geometry.
The second part is a link-state forecaster. It is a small temporal network that takes joint angles, hand aperture, taxel activation and the last few channel measurements, and predicts per-path achievable rate 10 to 100 ms ahead. Two choices are deliberate. The model is small enough to run on the robot. Its inputs are split so that an ablation can remove body state and keep only channel history.
The third part is a ladder of tactile representations with an explicit rate for each rung, so the controller can pick the largest rung the forecast path can carry within the control deadline. The top rung is a full taxel frame coded losslessly. Raw taxel streams are already modest in rate, and TacCompress: A Benchmark for Multi-Point Tactile Data Compression in Dexterous Hand shows up to 200x lossless compression on a 1,140-taxel hand using screen-content video coding. The middle rung is a lossy screen-content code. TacCompress reports a 1200x reduction with only a 1% drop in SVM classification accuracy. The bottom rung is a few low-dimensional contact features. Huang et al. showed that contact centres make deep RL tractable, and Guo et al. warn that such features discard slip and edge detail, which is why this rung is the last resort rather than the default.
For optical tactile sensors, which produce video-rate frames and where bitrate really does bind, we add rungs from TaCo: A Benchmark for Lossless and Lossy Codecs of Heterogeneous Tactile Data. That benchmark compares 30 codecs and introduces TaCo-L, a neural lossy codec. Tactile Functasets offers a 512-dimensional latent code per frame from a shared implicit network. We use these codecs as they are. Our contribution is the forecast-driven choice among them, not a new codec.
The downstream consumer is the tactile plugin adapter from DECO. DECO injects taxel signals into a diffusion-transformer policy through cross-attention and tunes 7.97M of 91M parameters. We retrain this adapter with rung-dropout, randomly showing it each rung during training, so that one policy accepts any rung.
Baselines
The naive reference point is a perfect channel: the published DECO policy with tactile input, labelled DECO.p. It reports an 82.50% average success rate across 28 subtasks, against 72.25% for its vision-only version (DECO). These two numbers bracket our results. Any scheme that falls to the vision-only figure under realistic self-blockage has made tactile sensing worthless. Our measure of success is how much of the 10.25-point tactile gain survives, with the 20-point gain on assembly and waste-disposal subtasks as the sharper target.
The second reference is "do nothing" on the network side. It uses the full-rate tactile stream with a fixed beam and packet loss during blockage, and the policy holds the last received tactile frame. This shows how bad unmanaged self-blockage is.
The strongest existing networking competitor is reactive blockage handling in the style of Blockage-Aware Multi-RIS WSR Maximization. That method gives each surface its own Zadoff-Chu synchronization sequence, finds unblocked paths by energy detection, and realigns phases with a closed-form method that reaches 95% of its converged weighted sum rate in about 12 iterations. We implement its detection-then-realign logic on our traces and charge it the same per-iteration time budget as our scheme. We pair it with the best fixed codec choice tuned by grid search on the training split, using the same compute budget. Beating it means two things: higher policy success on contact-critical subtasks, and fewer control-deadline misses in the 100 ms window after contact onset, at equal transmit power. For the realignment delay itself on hand self-blockage, we found no reported result on Lacuna, so we will measure it.
The strongest tactile-side competitor is a fixed best codec over an adaptive-rate link, with no forecast. It uses TaCo-L for optical data, which keeps 62.2% grasp-and-lift success at 0.025 bits per pixel versus 63.8% uncompressed in Isaac Sim (Cheng et al.), and screen-content video coding for taxel data. This configuration is also tuned on the training split. Two ablations isolate our key ingredient. The first gives the forecaster channel history only, with no body state. The second gives it body state but runs it on a single rung, so we can separate the gain from forecasting from the gain from representation switching.
Why not just…?
Why not just use a better codec and accept occasional drops? A codec reduces rate, but self-blockage reduces rate by 20 to 40 dB for the length of a grasp. For taxel data the raw rate is already small, so a better codec barely helps. The failure is outage at contact, not bandwidth. The fixed-codec baseline tests this directly. If it matches our scheme on contact-critical subtasks, the forecast adds nothing and we will say so.
Why not just detect blockage reactively, as multi-RIS synchronization schemes already do? This is the most serious threat. If detection and realignment fit inside one control period, prediction is unnecessary. Ryu and Yang report convergence in iterations, not milliseconds, and the file reports no wall-clock figure. We will measure the deadline-miss rate after contact onset under several assumed per-iteration latencies. We will report the latency below which reactive handling wins, which turns the question into a measured crossover rather than an argument.
Why not just compute tactile features on the robot, or use a cable or sub-6 GHz radio? If the policy runs on board, the link problem disappears, and we accept that. Our scope is offloaded or teleoperated control, where dense tactile streams must leave the robot. That regime includes optical tactile video and the bandwidth demands that push systems toward high-frequency bands. We will also report results for taxel-only robots, where a low-band radio may be enough. Where it is, that is a finding about scope, not a defeat.
Why not just forecast the channel from its own history? Generic channel prediction has no warning of a sudden occlusion until the occlusion begins. The channel-history-only ablation measures exactly this. Our central bet is that body state adds lead time, and it fails if that ablation does as well.
Decisive Checks
First, on held-out DECO-50 trajectories, we measure how far ahead the body-state forecaster predicts blockage onset compared with the channel-history ablation, along with its false-alarm rate. If body state adds no lead time, the project reduces to the reactive baseline. Second, we replay held-out trajectories through the retrained adapter fed with channel-degraded tactile input. We report action-prediction error and the fraction of DECO.p's tactile gain retained on assembly and waste disposal, for passive versus active surfaces at matched element counts. We expect the passive surface to force the lowest rung often and the active surface to sustain the middle rung, which would make the double-fading gap visible as a manipulation-quality gap. Third, for optical tactile data, we run closed-loop grasp-and-lift in the Isaac Sim setup of the TaCo benchmark. The target is to stay near the 62.2% that TaCo-L reaches on a clean channel, under simulated self-blockage traces. Each check is falsifiable on its own, and the second is the one that justifies treating path and representation together.
References
- Downlink Analysis for Reconfigurable Intelligent Surfaces Aided NOMA Networks
- Active RIS vs. Passive RIS: Which Will Prevail in 6G?
- TacCompress: A Benchmark for Multi-Point Tactile Data Compression in Dexterous Hand
- Dexterous in-hand manipulation of slender cylindrical objects through deep reinforcement learning with tactile sensing
- Aerial Interaction with Tactile Sensing
- TaCo: A Benchmark for Lossless and Lossy Codecs of Heterogeneous Tactile Data
- Tactile Functasets: Neural Implicit Representations of Tactile Datasets
- DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter
- Blockage-Aware Multi-RIS WSR Maximization via Per-RIS Indexed Synchronization Sequences and Closed-Form Riemannian Updates