Magnetic tactile skins are soft silicone pads filled with magnetized particles that sit over a small grid of magnetometers. Pressing or dragging the pad moves the particles, and the magnetometers see the field change. These skins are cheap, durable and easy to swap. Today they are read the expensive way: every magnetometer reports all three field axes at a fixed 100–400 Hz, whether the hand is touching something, holding still, or idle. Neuromorphic sensing does the opposite. A sensor emits a discrete "spike" only when its signal changes, and a spiking neural network (a network whose units communicate only through such sparse binary events) does work only when spikes arrive. We propose to give magnetic skins a spike-based readout. The key step is that the spikes must reflect contact rather than anything magnetic.
That distinction is the whole proposal. Applying "emit a spike when the value changes" naively to raw magnetometer readings produces a stream dominated by the robot's own motors, nearby steel and slow elastomer drift. Each of these changes the field just as contact does, so the sparsity that makes event-driven processing worthwhile disappears exactly when the robot is moving. We will build and test an encoder with three properties. It cancels field changes that are shared across the whole sensor board, the way a distant motor's field would be. It splits each sensor's response into fast-adapting and slow-adapting channels, loosely modeled on skin receptors. And it emits signed spikes along each field axis and along neighbor-to-neighbor field differences, so the sideways (shear) deformation that signals incipient slip survives the conversion to spikes. We will measure whether this keeps event rates near zero during contact-free motion while matching dense-polling accuracy on slip, localization and material tasks. We will also measure what fraction of total sensor-plus-compute energy is actually saved.
Core Claim
On a magnetic skin, a change-triggered code is only sparse and only informative if it first removes field changes common to the whole board and then encodes direction-preserving differences between neighboring sensors. With those two steps, a spike stream that is mostly silent during idle, steady holds and contact-free arm motion can match a densely polled recurrent network on slip detection and contact localization. Because adaptation subtracts each skin's static magnetization offset, it should also reduce the accuracy loss when a skin is swapped for a new one. The claim is about the encoder. Whether a spiking network or an event-gated conventional network should sit downstream is treated as an open question that the experiments settle.
Position Relative to Prior Work
The research file contains no spiking or event-driven readout of a magnetic skin, but the obvious version of this project exists for another sensor type. An Event-Driven E-Skin System with Dynamic Binary Scanning and real time SNN Classification builds a 16×16 piezoresistive array. It is scanned slowly while idle and quickly once touched, and pressure changes are converted into ternary spikes for an FPGA spiking convolutional network. The authors report 38.2x data compression and 99% stream sparsity. Similarly, ACES in ST-MNIST - The Spiking Tactile MNIST Neuromorphic Dataset emits asynchronous ON/OFF pressure events from each taxel.
These systems supply our scanning and delta-modulation machinery. They do not face our problem. A piezoresistive taxel responds only to pressure on itself, so a change almost always means contact. A magnetometer responds to every field source in the room. Their one-dimensional pressure events also carry no shear direction. Porting their encoder per axis is therefore our main ablation, not our method.
The magnetic-skin line motivates the work and supplies the hardware. ReSkin: versatile, replaceable, lasting tactile skins establishes the particle-in-silicone design with 1 mm localization at 90% accuracy. It shows that cross-skin generalization collapses to about 25% without self-supervised recalibration. AnySkin: Plug-and-play Skin Sensing for Robotic Touch uses pulse magnetization to make skins uniform enough for zero-shot policy transfer, with a 13% drop on plug insertion, and reports 92% slip accuracy on unseen objects. All the Feels: A Dexterous Hand With Large-Area Tactile Sensing scales the approach to 56 magnetometers on a hand processed by a 2-layer LSTM. All three poll synchronously. The file also notes that their vulnerability to stray fields is handled only by calibration, with differential sensing left as future work. Our common-mode rejection is that differential idea, made the front end of an event code.
Deep Learning Classification of Touch Gestures Using Distributed Normal and Shear Force supplies the physical intuition for the shear channel. Normal pressing changes neighboring sensors symmetrically, while shear produces antisymmetric gradients between them. Signed neighbor-difference spikes are a direct encoding of that antisymmetry.
Electronic Receptors for Tactile/Haptic Sensing motivates the adaptation step. Andreou observed that synthetic magnetic receptors, unlike biological fast- and slow-adapting mechanoreceptors, report raw static levels and push all baseline removal downstream. We do not build his subthreshold CMOS receptors. We implement the adaptation he found missing, in the readout microcontroller, as a step toward eventually moving it into silicon.
On the downstream side, TactileSGNet: A Spiking Graph Neural Network for Event-based Tactile Object Recognition provides a spiking graph network for irregular taxel layouts, which we reuse over the magnetometer adjacency graph. Event - Driven Tactile Learning with Location Spiking Neurons shows that neurons accumulating over spatial position reach 100% slip accuracy on NeuTouch with an estimated 10–400x energy advantage over dense networks. Both used event-native sensors and estimated rather than measured energy. Our study inherits their models but measures energy end to end, including the sensor reads that their accounting omits.
Research Program
The encoder runs on the microcontroller that already reads the magnetometers. Let $B_i(t)$ be the three-axis field at sensor $i$. We first subtract the board-wide common component, estimated either from a reference magnetometer shielded from the skin or from a robust median across sensors. This leaves residual fields $\tilde B_i(t)$. Distant sources such as motors and tables vary slowly over the few centimeters of a pad, so they appear mostly in the common component. Contact is local and appears mostly in the residuals.
The residuals then feed two channels. The fast channel runs signed delta modulation on each axis of $\tilde B_i$ and on each axis of the neighbor difference $\tilde B_i - \tilde B_j$ for adjacent sensors $i$ and $j$. A $+1$ or $-1$ spike is emitted when the change since the last spike exceeds a threshold. The slow channel emits sparse level events encoding the adapted residual relative to a slowly tracked baseline. This keeps sustained grip force available for softness and material judgments, which a pure change detector would discard.
Sampling is gated, as in the piezoresistive system. Sensors are read at a low idle rate and switched to full rate only in regions where fast-channel spikes occur. This gating is where hardware energy is actually saved, because commercial magnetometers spend energy per conversion and per bus transaction.
Energy is accounted end to end:
$$E = N_{\text{read}}\,e_{\text{read}} + N_{\text{bit}}\,e_{\text{bit}} + N_{\text{op}}\,e_{\text{op}}$$
Here $N_{\text{read}}$ is the number of magnetometer conversions, $N_{\text{bit}}$ the number of bits sent over the bus, and $N_{\text{op}}$ the number of downstream additions or multiply-adds. Each lowercase $e$ is the energy per item. We measure $e_{\text{read}}$ and $e_{\text{bit}}$ with a current probe on our boards, and count $N_{\text{op}}$ exactly. The point of this equation is to stop the project from claiming compute savings that are negligible next to sensing cost.
Downstream we train two models. One is a spiking graph network in the style of TactileSGNet, with surrogate gradients. The other is an event-gated GRU (a conventional recurrent network that updates its state only when an event arrives). Comparing them separates the value of the spike code from the value of spiking computation.
Baselines
The naive reference is the deployed practice. All magnetometers are polled at the full native rate, 400 Hz as reported for ReSkin, with a dense recurrent network on every frame. The reported targets are 92% slip accuracy on unseen objects from AnySkin: Plug-and-play Skin Sensing for Robotic Touch and 1 mm localization at 90% accuracy from ReSkin: versatile, replaceable, lasting tactile skins. If a multi-finger hand is available, we add the 2-layer LSTM numbers from All the Feels: A Dexterous Hand With Large-Area Tactile Sensing: 71.24% material identification, 76.17% softness and 59.03% texture on unseen objects.
This is the strongest competitor on accuracy because it sees every sample. It defines the accuracy ceiling and the energy cost we are trying to cut. We re-run it on our own recordings rather than relying on the published figures, since our setup differs.
The strongest existing event-driven method is the direct port of An Event-Driven E-Skin System with Dynamic Binary Scanning and real time SNN Classification. It applies their gated scanning and ternary delta modulation independently to each raw magnetometer axis and feeds a spiking convolutional network. Its reported figures on piezoresistive digits are 99% sparsity and 38.2x compression. No reported result was found on Lacuna for any event encoding on a magnetic skin.
The fair comparison works as follows. Every encoder is applied offline to the same raw high-rate recordings. Each encoder's threshold is swept over an identical grid, and every downstream model gets the same hyperparameter search budget and parameter count range. Results are compared as accuracy-versus-energy curves rather than single points. Beating this baseline means one of two things on the contact-free-motion and slip test sets: a strictly better curve, or at least a 5x lower event rate at equal accuracy.
Two ablations isolate the key ingredients. The first removes common-mode subtraction while keeping everything else. The second removes the neighbor-difference shear spikes, keeping per-axis spikes only. We predict the first ablation fails mainly on event rate during arm motion, and the second fails mainly on incipient slip and shear-direction tasks.
Why not just…?
Why not just poll with wake-on-change and keep the LSTM? This is the most serious threat. It plausibly captures most of the sensing-energy savings, because those savings come from gated reads, not from spikes. We therefore made the encoder, not the spiking network, the core claim. The event-gated GRU arm runs on exactly this design, and our energy accounting will show whether downstream compute is a meaningful share of the total. If it is not, we will report the result as an encoding contribution and drop any claim that spiking computation matters for this sensor.
Why not just lower the uniform polling rate? Slip onset is brief. The location-neuron work defines slip detection within a 0.15 s window (Event - Driven Tactile Learning with Location Spiking Neurons), and a low fixed rate trades away exactly that latency. We include fixed-rate polling at several rates on the same accuracy-energy plot, together with a detection-latency metric. If a low uniform rate matches our curve, the gating adds nothing and we will say so.
Why not just calibrate out ambient fields in software and then apply delta modulation? Static calibration, which is what deployed magnetic skins rely on, removes fixed offsets. It cannot remove fields that change as motors draw current or as the arm moves past steel. Our contact-free motion recordings test this directly. They compare static calibration followed by delta modulation against live common-mode subtraction, measuring false events per second.
Why not just use an event-native piezoresistive skin? These skins already produce sparse events. The ST-MNIST work reports only one-dimensional normal-pressure events (ST-MNIST - The Spiking Tactile MNIST Neuromorphic Dataset). They also place the conductive layer at the wear surface, whereas ReSkin survived more than 50,000 indentations with a passive silicone interface. Wear is not tested here. Shear is, through the slip and shear-direction tasks, which are the tasks where a pressure-only code should fall short.
Proposed Experiments
We build AnySkin-style pads (under $30 each per ReSkin) on a parallel gripper, and on a multi-finger hand if available. We record raw streams at the magnetometers' maximum rate with synchronized force and torque ground truth. Every encoder is then evaluated offline on identical data, and the selected configurations are finally run live with current measurement.
The data collection reproduces three published protocols. The first is ReSkin-style indentation localization. The second is AnySkin's slip protocol: 40 objects, split 30 for training and 10 unseen for testing, with six trajectories per object. The third is a contact-free arm-motion set recorded near running motors and a steel table, with no contact at all. If a hand is available, we add the D'Manus material and softness tasks. We will also add a shear-direction classification task with controlled lateral loading.
Metrics are accuracy, slip detection latency, false events per second without contact, events per second during a steady hold, and measured end-to-end energy per second and per decision.
A cross-skin experiment trains on several skins and tests zero-shot on new ones. It asks whether the adapted residual code narrows the gap that ReSkin needed self-supervised recalibration to close. The reference point is AnySkin's 13% policy drop.
Three outcomes would be decisive. The first is the contact-free motion test. If common-mode-rejected events stay near zero while raw per-axis delta modulation fires continuously, the central premise holds. If both are quiet, magnetic interference is not the bottleneck we claim, and the work collapses to the piezoresistive port. The second is the shear ablation on the slip and shear-direction tasks, which shows whether direction-preserving neighbor spikes carry information that per-axis spikes lose. The third is the energy breakdown. If sensor reads dominate and the event-gated GRU matches the spiking graph network, the conclusion is that magnetic skins need event encoding but not neuromorphic processors, and we will report it that way.
References
- An Event-Driven E-Skin System with Dynamic Binary Scanning and real time SNN Classification
- ST-MNIST - The Spiking Tactile MNIST Neuromorphic Dataset
- ReSkin: versatile, replaceable, lasting tactile skins
- AnySkin: Plug-and-play Skin Sensing for Robotic Touch
- All the Feels: A Dexterous Hand With Large-Area Tactile Sensing
- Deep Learning Classification of Touch Gestures Using Distributed Normal and Shear Force
- Electronic Receptors for Tactile/Haptic Sensing
- TactileSGNet: A Spiking Graph Neural Network for Event-based Tactile Object Recognition
- Event - Driven Tactile Learning with Location Spiking Neurons