Cheap robot skins often cover a large surface with only a handful of sensors buried inside a soft or structured body. A touch anywhere bends the whole body, so every sensor feels a little of every contact, and a learned model works out where the touch was. This works well for one touch, but real manipulation is rarely one touch. A hand holds an object firmly, and meanwhile the object starts to slip, or a second finger brushes against it. The sensors then record a sum of a large signal and a small one.
Wireless engineers face the same problem when a strong and a weak transmitter share a channel. Their standard fix is successive interference cancellation: decode the strong signal, subtract its predicted contribution, then decode the weak one from what is left. Their theory also gives a sobering result. The subtraction is never exact, and the leftover error grows with the strong signal's power. Past a certain point, making the system less noisy stops helping. This proposal carries that result, and its fix, into touch sensing.
The claim has two parts. First, in sparse-sensor skins, whether a light secondary contact can be detected is limited by a residual floor that grows with the primary load, not by sensor noise. The cause is that soft materials respond nonlinearly: the primary press stiffens and reshapes the substrate, so the secondary touch's signature is both weaker and displaced compared with its signature on an unloaded skin. Second, a decoder built around this fact does better than one that ignores it. The decoder subtracts the primary contact's contribution and then reads the secondary contact's signature conditioned on the primary's estimated state. It also raises its detection threshold in proportion to the measured floor. We expect it to detect lighter secondary contacts and to report fewer "ghost" contacts (touches the model reports that do not exist) than both naive subtraction and black-box regression.
Why this is open
Two tactile systems show both the opportunity and the gap. HapDefX instruments a robot limb with 10 strain gauges. A three-stage network reconstructs deformation over 3,211 surface points, achieving 3 mm single-contact error and 10 mm double-contact resolution with up to four contacts at 1 kHz. However, it reports ghost contacts at high forces, attributed to nonlinear structural deformation. Its feedforward stages have no mechanism for peeling away a dominant contact. Overlapping air chambers localize a single touch to 0.13 mm with four barometers and a causal convolutional network over nine frames. Multi-contact is never evaluated, even though four sensors seeing two simultaneous contacts is underdetermined in a single frame.
The communications work supplies the theory and the method. Yue et al. model leftover interference as a fraction $\beta$ of the strong user's power. They prove that outage then hits an error floor that more transmit power cannot remove. The DL-HetNet study shows that when this fraction is large or the power gap between users is small, orthogonal access (the analogue of physically isolated taxels) beats non-orthogonal sharing. That result is a constraint on this proposal: it predicts a regime where cancellation should lose, and we must find it rather than hide it.
Multiplicative-array channel sounding is the closest methodological precedent. It removes a strong path together with its deterministic ghost cross-terms before searching for weaker paths. It assumes an exact analytic ghost model, which elastomers do not offer. CIDER takes the opposite route. It avoids sequential subtraction entirely by decoding all users in parallel slots with a repulsion term, because unconstrained networks tend to predict the dominant component twice. That makes it the natural non-sequential competitor.
No work in the file measures a residual-interference coefficient for mechanical signals or asks how it scales with grasp force.
Core mechanism
Let $s \in \mathbb{R}^M$ be the $M$ sensor readings, and let $g(c \mid c')$ be the sensor response to contact $c$ (position and force) when contact $c'$ is already present. For secondary contact $c_2$ under primary contact $c_1$, define a detectability ratio by analogy with the imperfect-cancellation signal-to-interference ratio:
$$\rho_2 = \frac{\lVert g(c_2 \mid c_1)\rVert^2}{\sigma^2 + \beta\,\lVert g(c_1 \mid \varnothing)\rVert^2}$$
Here $\sigma^2$ is sensor noise and $\beta$ is the fraction of the primary's signal left after cancellation.
The first empirical question is whether $\beta$ is roughly constant as primary force $F_1$ grows. If it is, the smallest detectable secondary force scales linearly with $F_1$: a fixed force ratio, independent of $\sigma^2$, which is the tactile analogue of zero diversity order. The form follows directly from the equation. When $\beta\lVert g(c_1)\rVert^2 \gg \sigma^2$, $\rho_2$ depends only on the ratio of the two responses.
Naive subtraction uses the unloaded response $g(c_1\mid\varnothing)$ for cancellation and the unloaded template $g(c_2\mid\varnothing)$ for detection. Our decoder instead learns both the primary's cancellation term and the secondary's signature as functions conditioned on the estimated primary state. It sets its detection threshold from a $\beta$ estimated on calibration data. We therefore predict it shows a smaller effective $\beta$ and fewer ghosts.
The decoder also exploits a fact that real grasps usually provide: contacts arrive at different times. The primary is estimated while it is alone and then tracked. This is what makes the four-barometer case solvable. Frame-level underdetermination becomes a tracking-plus-residual problem. The temporal continuity idea comes from the air-chamber work, now applied to cancellation.
Research Program
We will work on two substrates.
The first is simulation. We will build soft-pad and structured-limb models in open-source CalculiX, following the FEATS labeling pipeline, and reuse HapDefX's ANSYS multi-contact data where available. This lets us sweep primary force, secondary force, separation, and arrival order densely, with exact ground truth. It also lets us switch the material between linear-elastic and hyperelastic, so we can test whether the floor comes from nonlinearity.
The second is physical. We will replicate the overlapping-air-chamber pad and add a two-indenter rig. A UR5 applies the secondary contact while a fixed, force-controlled indenter holds the primary load. The same indenter sizes and depth grid as the original dataset will be used.
On both substrates we will do three things. We will measure $\beta$ curves for each decoder. We will map the "cancellation region" of force ratio and separation where sequential decoding beats joint decoding. And, following the DL-HetNet crossover, we will map where no sparse decoder succeeds and only denser, isolated sensing would.
The decoder itself is small: a single-contact tracker, a conditioned residual network, and a floor-aware detector. The target latency is under 1 ms, to preserve kilohertz update rates.
Baselines
The naive reference is the unmodified single-contact model: the air-chamber causal CNN, and HapDefX's pipeline applied as published. It shows what happens when multi-contact is ignored.
The next baseline is linear template cancellation, which amounts to greedy matching pursuit. We subtract the unloaded single-contact response of the strongest estimated contact and repeat. This is the "perfect cancellation" assumption, and the theory predicts it hits the floor hardest.
The strongest existing tactile method is HapDefX's three-stage pipeline, retrained on our multi-contact data. The best reported number is 10 mm double-contact resolution with up to four contacts (HapDefX). For the air-chamber pad there is no reported multi-contact result on Lacuna, and its single-contact 0.13 mm error serves as an upper bound on what two-contact decoding could reach.
We add two further strong competitors:
- End-to-end set prediction. The network outputs a set of contacts and is trained with optimal bipartite matching between predicted and true contacts, on the same data and temporal window.
- CIDER-style parallel decoding. An adaptation of CIDER with parallel contact slots and a repulsion term.
We also include windowed joint nonlinear least squares over a learned differentiable forward model, a fast stand-in for Su et al.'s circuit optimization.
All learned methods get identical training data, the same GPU-hour budget, and the same hyperparameter search size.
Two ablations isolate the key ingredients: conditioned versus unconditioned cancellation, and floor-aware versus fixed detection thresholds. A third orders contacts by arrival time rather than force, to test whether the temporal ordering carries the gain.
Beating the field means both of the following at matched latency:
- a statistically lower minimum detectable secondary-to-primary force ratio, and
- lower ghost rate than the best of set prediction and joint least squares.
Why not just…?
Why not just train a bigger end-to-end network on multi-contact data? This is the real threat. HapDefX's ghosts may reflect limited multi-contact training rather than a missing mechanism. CIDER's duplicate-collapse finding suggests unconstrained decoders struggle, but that evidence comes from wireless frames, not strain fields. The matched-budget comparison settles it. If set prediction matches our decoder's $\beta$ curve, the decoder claim fails. The measured floor and the cancellation-region map would remain the contribution, and we would say so.
Why not just solve all contacts jointly by optimization? With a learned forward model and few sensors, joint fitting may be fast, unlike Ipopt's 11 samples per second on a 2×2 textile grid. Its weakness is identifiability: two contacts under four barometers are underdetermined in one frame. Windowed joint fitting is included precisely to test whether a smoother recovers what sequential tracking does.
Why not just use optical tactile sensors or isolated taxels? Optical sensors cannot cover finger links and palms, and Yin et al. show whole-hand coverage matters more than fingertip resolution. Isolated taxels bring wiring and fabrication costs. We do not assume sparse skins always win. The crossover map will show where isolation is necessary.
Decisive checks
The decisive result is the curve of minimum detectable secondary force against primary force.
- Floor exists. A linear curve for template cancellation on the hyperelastic model, and a flat curve on the linear-elastic model, would confirm that nonlinearity creates the floor.
- Conditioning works. A shallower slope for the conditioned decoder, holding across physical data, would confirm that conditioning lowers it.
- Floor does not exist. If the curve is flat for every method, the imperfect-cancellation framing does not apply to touch, and the project stops there.
We will additionally report ghost rate per frame, localization error for both contacts, latency, and the force-ratio boundary where sequential decoding stops beating joint decoding.
References
- Machine Learning for Haptics: Inferring Multi-Contact Stimulation From Sparse Sensor Configuration
- Improving Robotic Tactile Localization Super-resolution via Spatiotemporal Continuity Learning and Overlapping Air Chambers
- Outage Performance of a Unified Non-Orthogonal Multiple Access Framework
- Distributed Cluster Formation and Power-Bandwidth Allocation for Imperfect NOMA in DL-HetNets
- Channel Sounding Using Multiplicative Arrays Based on Successive Interference Cancellation Principle
- Structured Masked Diffusion for Joint Multiuser Decoding
- Learning Force Distribution Estimation for the GelSight Mini Optical Tactile Sensor Based on Finite Element Analysis
- Optimizing Multi-Touch Textile and Tactile Skin Sensing Through Circuit Parameter Estimation
- Rotating without Seeing: Towards In-hand Dexterity through Touch