Robots with touch sensors are hard to train because touch data is rare. Unlike images, it cannot be scraped from the web. Every useful frame comes from a physical contact, and the frames that matter most (a slip starting, a gentle grasp, a material the model has never felt) are a small fraction of what a robot records. When a fleet of robots shares one wireless network and one edge learner, a second bottleneck sits on top of the first: the radio uplink cannot carry every high-resolution tactile frame from every robot. This note proposes treating the radio network as the instrument that decides which tactile data reaches the learner.
Each robot attaches a cheap, label-free estimate of how much a tactile clip would teach the shared model. A scheduler inside the radio access network then allocates airtime and compression strength to clips according to expected learning value per bit, given each robot's current channel quality, instead of by fairness or by reconstruction quality. The claim is that, at a fixed uplink budget, this recovers more downstream manipulation performance after a distribution shift, and recovers it sooner, than sending everything through the best existing tactile codec.
Core Claim
Two findings in the literature point the same way, but nobody has connected them. First, TaCo, a benchmark of 30 lossless and lossy codecs for tactile data, shows that signal fidelity is not semantic fidelity. YCB-Slide tactile frames survive 190x compression with over 98% classification accuracy. Its lossy codec TaCo-L keeps a 62.2% simulated grasp-and-lift success rate at 0.025 bits per pixel, against 63.8% uncompressed. Second, the AI-native network literature argues that intelligence should sit inside the radio stack rather than on top of it. The intelligence-plane and endogenous-intelligence visions say this (Data Science and Artificial Intelligence for Communications; 6G Wireless Systems: A Vision, Architectural Elements, and Future Directions), and Open RAN gives the concrete hooks: a near-real-time controller that hosts small apps ("xApps") acting on sub-second radio decisions, and a slower controller that hosts training and policy apps (A Comprehensive Tutorial and Survey of O-RAN).
TaCo's result implies most tactile bits are worthless for inference. Our hypothesis goes further: most tactile clips are also worthless for learning, and the few that are not are exactly the rare contact events that data scarcity is about. A network that knows this can multiply effective tactile data per bit well beyond what any codec achieves alone.
What is new is not a better tactile codec and not a better slicing algorithm. It is a joint decision over three things: which clip to send, at which rate, and on which radio resources. That decision is driven by the learner's current uncertainty and the physical channel.
Positioning Against Closest Work
The closest pieces each hold one part of the problem. TaCo supplies the method for rate control: we reuse its lossy codec at several operating points. However, TaCo evaluates static pipelines on fixed datasets with no channel and no fleet, and it compresses every frame equally. CoMP-Enabled RAN Slicing for Tactile Internet motivates the network side. It allocates radio resources for tactile slices under a 1 ms budget, but it models touch as Poisson packet arrivals with no content, and it names edge intelligence over tactile feedback as future work. The semantic 6G air interface of Way to Build Native AI-driven 6G Air Interface shows that task-oriented transmission avoids the "cliff effect" of separate codec-plus-channel-code pipelines at low signal-to-noise ratio. That work is on video for inference and never addresses which content is worth sending for training.
On the learning side, DECO supplies the learner. Its small plugin tactile adapter, 7.97M trainable parameters on a frozen 91M backbone, means the edge can update touch skills cheaply and push them back to robots. TACTFL shows label-free federated training with similarity-weighted aggregation on wearable sensors, and it is the main alternative design we must beat. The "Reasoning Plane" of Less Data, More Knowledge shares our premise that networks should move knowledge rather than raw data. It remains conceptual and flags the computational cost of causal models on high-dimensional sensor feeds, so we deliberately avoid structural causal models and use a cheap learned score instead.
Research Program
The system has three parts.
The first part is an on-robot value score. Each robot runs a frozen copy of the shared tactile encoder, pretrained with masked reconstruction in the style of AnyTouch 2. For each short tactile window $x$ it computes a score $v(x)$, the self-supervised prediction error of the current shared model on that window. This needs no labels and costs one forward pass, and it rises on contact transients and unfamiliar materials. A known risk is that broken or saturated sensors also produce high error. We therefore gate the score with a cross-modal consistency check: a high-error tactile window must coincide with a change in proprioception or vision, borrowing TACTFL's use of temporal co-occurrence as supervision.
The second part is a learning-value scheduler, implemented as a near-real-time xApp. Every scheduling interval, each robot $i$ reports the scores of its queued clips, and the base station reports that robot's achievable rate $c_i$ from channel state. The xApp picks, for each clip $j$, a codec operating point $r_{ij}$ (in bits) from TaCo-L's rate ladder, or skips it, to solve
$$\max_{r}\ \sum_{i,j} v_{ij}\, q(r_{ij}) \quad \text{s.t.} \quad \sum_j \frac{r_{ij}}{c_i} \le \tau_i,\ \ \sum_i \tau_i \le T .$$
Here $q(r)$ is the fraction of a clip's learning value retained at rate $r$, measured once offline from TaCo-style rate-versus-task curves. $\tau_i$ is the airtime given to robot $i$, and $T$ is the interval's uplink airtime. This is a multiple-choice knapsack problem, so a greedy solution by value per unit of airtime is fast enough for the near-real-time loop. Robots on poor channels are not starved by rule; they win airtime only when their clips are rare enough to justify the cost.
The third part is an edge learner and adapter push-back, running in the non-real-time controller. Received clips train the DECO-style tactile adapter, and updated adapter weights are broadcast back to the fleet. The refreshed weights also update the encoder that computes $v$, so the score tracks what the fleet has already learned. This closed loop is the AI-native part: the network's scheduling objective is a quantity defined by the learner, not by bit error or delay.
Baselines
The naive reference sends all tactile data through TaCo-L at a single fixed operating point, scheduled by a standard proportional-fair scheduler, with the edge training on whatever arrives. It shows what the best existing codec buys with no content awareness. There is no reported fleet-learning number for this configuration. The closest anchor is TaCo-L's 62.2% simulated grasp success at 0.025 bpp versus 63.8% uncompressed (TaCo).
The strongest data-moving competitor is TaCo-L plus channel-adaptive rate selection plus an on-robot event filter. The filter sends only windows where tactile intensity changes by more than a threshold. This is what a careful engineer would build without our scheduler, and it captures most of the "rare events matter" intuition with no learned score.
The strongest model-moving competitor is federated adapter training with TACTFL's temporal contrastive objective and similarity-guided aggregation (TACTFL). Robots upload DECO adapter updates instead of data. Its strongest reported result is 68.48% top-1 accuracy on UCF101 with 10% labels, against 35.35% for FedOpt. There is no reported result on Lacuna for TACTFL on tactile arrays or manipulation.
All methods get the same robots, the same channel traces, the same total uplink bits per hour, the same frozen backbone, and the same edge compute and hyperparameter search budget. The threshold, the aggregation interval, and our score temperature each get an equal grid of trials.
The ablations isolate the key ingredient. One replaces $v$ with random scores while keeping the channel-aware knapsack; this tests whether learning value matters at all. A second keeps $v$ but ignores channel state; this tests whether network integration matters. A third removes the consistency gate. Beating the field means reaching the target success level with substantially fewer uplink bits than the event filter and the federated baseline. We set the bar at 2x fewer bits, chosen by us rather than drawn from any paper. Beating the field also means recovering faster after a shift, with no loss in gentleness metrics.
Why not just…?
Why not just compress everything with TaCo-L and send it all? At 0.025 bpp, uniform compression may already fit the uplink, and if so our scheduler is pointless. The proposal tests this directly by sweeping the uplink budget. We expect the advantage to appear only when fleet size times frame rate exceeds capacity even at TaCo-L rates, or when channels are poor. If the curves meet across realistic budgets, the honest conclusion is that tactile codecs alone have solved the problem, and we would report that.
Why not just store data on the robot and bulk-upload it at the charging dock? This is a real threat whenever adaptation can wait. Our claim is restricted to regimes where the fleet must adapt within a shift, such as new objects, new materials, or a worn elastomer. We measure time-to-recovery against a dock-upload baseline with the same daily bit budget. If dock upload matches us on final performance and the delay does not hurt, our method's value reduces to faster adaptation, and the experiment states how much faster.
Why not just federate the adapter and never move data? A 7.97M-parameter adapter at 16-bit precision is about 16 MB per update. That can be cheaper than raw tactile video, but not obviously cheaper than a few selected, heavily compressed clips. Federated rounds also average over mostly uninformative local data, while rare events are diluted. The bits-to-performance comparison against TACTFL-style aggregation settles this. A hybrid, sending adapter updates from robots on poor channels, is a natural follow-up if neither approach dominates.
Why not just use simulation instead of fleet data? Cross-Tactile Sensor Representation Learning reports that simulators miss fine textures and friction transients. Tabero stayed in simulation precisely because the tactile reality gap is hard. Real contact data stays necessary for the shift regimes we target, and we pretrain on simulated ToucHD data in every arm so that simulation's contribution is held constant.
Proposed Experiments
Offline replay on real tactile data. We split DECO-50 into virtual robots, each replaying its own trajectories over recorded or standard fading channel traces. The 1,062 contact points per robot come from Inspire tactile pads. Pick and Place and Material Sorting serve for pretraining. Assembly and Waste Disposal arrive as the shift, chosen because DECO reports a 20% absolute tactile gain on exactly those tasks. The metric is held-out action prediction error versus cumulative uplink bits. The ceiling reference is DECO.p's reported 82.50% average success with full tactile data, against 72.25% vision-only (DECO).
Optical-sensor replay. We repeat the replay with TaCo's Touch and Go and YCB-Slide streams, where the codec rate ladder is defined. The metric is material and object classification accuracy versus bits.
Closed-loop simulation. We run fleets in the Tabero Isaac Lab pipeline with simulated GelSight sensors. We inject shifts in object compliance and friction, and we report success, maximum transient grip force, and average applied force, following Tabero. Tabero reports 87% firm and 79% gentle success. This tests whether value-driven selection over-samples slips and inadvertently teaches firmer grasps.
Decisive checks. Three results would decide the question. The first is whether the random-score ablation closes most of the gap to the full method; if so, the network gain is mere channel awareness. The second is whether the event-threshold filter matches the learned score; if so, the learned score is unnecessary. The third is whether the advantage survives at the generous end of the budget sweep. The proposal stands only if learning value, not channel adaptivity or a simple heuristic, accounts for most of the bits saved.
References
- TaCo: A Benchmark for Lossless and Lossy Codecs of Heterogeneous Tactile Data
- Data Science and Artificial Intelligence for Communications
- 6G Wireless Systems: A Vision, Architectural Elements, and Future Directions
- A Comprehensive Tutorial and Survey of O-RAN: Exploring Slicing-Aware Architecture, Deployment Options, Use Cases, and Challenges
- CoMP-Enabled RAN Slicing for Tactile Internet
- Way to Build Native AI-Driven 6G Air Interface: Principles, Roadmap, and Outlook
- DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter
- TACTFL: Temporal Contrastive Training for Multi-modal Federated Learning with Similarity-guided Model Aggregation
- Less Data, More Knowledge: Building Next Generation Semantic Communication Networks
- AnyTouch 2: General Optical Tactile Representation Learning For Dynamic Tactile Perception
- Cross-Tactile Sensor Representation Learning
- Tabero: Learning Gentle Manipulation with Closed-Loop Force Feedback from Vision, Touch, and Language