Adaptive Compressive Tactile Subsampling: Enabling High Spatiotemporal Resolution in Scalable Robotic Skin

Robots require full-body, high-resolution tactile sensing to operate safely in unstructured environments, enabling reflexive responses and closed-loop control. However, the pixel counts needed for dense, large-area coverage limit readout rates of most tactile arrays to <100 Hz, hindering their use in high-speed tasks. We present Adaptive Compressive Tactile Subsampling (ACTS), a scalable and data-driven method that greatly enhances traditional tactile matrices by leveraging adaptive sensor sampling and sparse recovery. By adaptively allocating measurements to informative regions, ACTS is especially effective for spatially sparse signals common in real-world interactions. Tested on a 1024-pixel tactile sensor array (32x32), ACTS achieved frame rates up to 1,000 Hz, an 18X improvement over conventional raster scanning, with minimal reconstruction error. For the first time, ACTS enables wearable, large-area, high-density tactile sensing systems that can deliver high-speed results. We demonstrate rapid object classification within 20 ms of contact, high-speed projectile detection, ricochet angle estimation, and soft deformation tracking, in tactile and robotics applications, all using flexible, high-density tactile arrays. These include high-resolution tactile gloves, pressure insoles, and full-body configurations covering robotic arms and human-sized mannequins. We further showcase tactile-based closed-loop control by guiding a metallic ball to trace letters using tactile feedback and by executing tactile-only whole-hand reflexes on a fully sensorized LEAP hand to stabilize grasps, prevent slip, and avoid sharp objects, validating ACTS for real-time interaction and motion control. ACTS transforms standard, low-cost, and robust tactile sensors into high-speed systems enabling scalable, responsive, and adaptive tactile perception for robots and wearables operating in dynamic environments.

💡 Research Summary

The paper introduces Adaptive Compressive Tactile Subsampling (ACTS), a novel framework that dramatically increases the temporal resolution of large‑area tactile sensor arrays without requiring any hardware redesign. Conventional tactile skins, typically based on resistive or piezoresistive matrices, suffer from raster‑scan latency: as the number of taxels grows, the time needed to read the full frame scales linearly, limiting frame rates to well below 100 Hz. This bottleneck prevents their use in high‑speed manipulation, reflexive control, or rapid impact detection.

ACTS tackles the problem from two complementary angles. First, it exploits the intrinsic spatial sparsity of tactile data. By collecting a large corpus of contact patterns (single‑point presses, lines, shapes, deformations) the authors train an over‑complete dictionary Ψ (size N × K, with K ≫ N). Any tactile frame x can be approximated as a sparse linear combination of a few dictionary atoms: x ≈ Ψα, where α is S‑sparse. Second, instead of measuring all N = 1024 taxels, ACTS uses an adaptive binary‑search‑inspired sampling strategy. Starting with the whole array, the algorithm probes rows and columns, recursively halving the search space until a pressure above a threshold is detected. This “smart” subsampling concentrates measurements on the active region and discards the vast inactive background.

The measurement matrix Φ is therefore a binary one‑hot matrix: each row selects a single taxel. The hardware implementation is deliberately simple. A standard 32 × 32 resistive array (Velostat on a flex‑PCB) is driven by a Teensy 4.1 microcontroller. The MCU toggles digital row lines and multiplexes column readouts to an ADC, acquiring M ≪ N samples per frame. After the M measurements y = Φx are collected, a fast greedy sparse recovery algorithm called FastOMP estimates the sparse coefficient vector α, and the full frame is reconstructed as x̂ = Ψα. FastOMP’s computational complexity (≈ O(KM)) fits comfortably on the MCU, enabling real‑time reconstruction at kilohertz rates.

Experimental evaluation used a library of 30 everyday objects and several 3D‑printed shapes. The objects typically occupied less than 10 % of the taxels. By varying M from 56 to 256, the authors achieved frame rates from ~550 Hz up to ~1 kHz (M ≈ 56 yields ≈ 1 kHz). Reconstruction quality was quantified by “support accuracy” (the fraction of correctly identified active pixels) and by object classification accuracy using a Sparse Representation‑based Classifier (SRC). Adaptive sampling consistently outperformed uniform and random subsampling, especially at low measurement budgets. With M = 88 (≈ 636 FPS) the system classified objects with 99 % accuracy; even with M = 56 (≈ 999 FPS) it retained 90 % accuracy. Crucially, a rapid‑classification test showed that after only 20 ms of initial contact (≈ 64 measurements) the adaptive scheme achieved 88 % correct identification, whereas raster scanning lagged at 51 %.

Beyond perception, the paper demonstrates closed‑loop tactile control. In one scenario a metallic ball is guided to trace letters on a surface using only tactile feedback from the ACTS‑enhanced skin. In another, a fully sensorized LEAP robotic hand employs ACTS to execute reflexive behaviors: stabilizing grasps, preventing slip, and avoiding sharp objects, all based on sub‑millisecond tactile updates. These demonstrations confirm that the high‑speed tactile stream can be directly fed into control loops without additional filtering or latency compensation.

The authors acknowledge several limitations. The dictionary must be pre‑trained on a representative set of contact patterns; novel or highly complex multi‑touch configurations may exceed its expressive capacity, leading to reconstruction errors. The current implementation is limited to single‑channel voltage readouts; extending to multimodal sensors (e.g., temperature, shear) would require richer dictionaries and possibly more sophisticated measurement matrices. Finally, while the binary one‑hot Φ is hardware‑friendly, it does not exploit more advanced compressive sensing matrices that could further reduce M at the cost of circuit complexity.

Future work is outlined as follows: (1) online dictionary adaptation to continuously incorporate new tactile experiences, (2) integration with FPGA or GPU accelerators for even faster sparse recovery, (3) exploration of structured random sensing matrices that remain implementable on flexible PCB traces, and (4) validation on larger‑scale skins (full‑body mannequins) and on soft robotic platforms.

In summary, ACTS provides a practical, firmware‑only upgrade path that lifts the frame rate of existing high‑density tactile skins by an order of magnitude while preserving spatial fidelity. By marrying data‑driven sparse representations with an intelligent adaptive sampling policy, the method enables real‑time tactile perception and reflexive control for robots and wearable devices, opening the door to truly dexterous, skin‑based interaction in dynamic, unstructured environments.