Source-linked AI summary
OpenTac: An Open Ecosystem for Customizable, Scalable, and Transferable 3D Tactile Sensing
Ariel Slepyan, Geeta Chandra Raju Bethala, Jiafeng Gu, Anagha Rajendra Dangle, Yinzhe Zhou, Hongyuan Fu, Jade Sim, Tuhin Basu, Xinyi Yin, Md Faizal Karim, Gregory Hager, Homanga Bharadhwaj, Anand Bhattad, Krishna Murthy Jatavallabhula, Nitish V. Thakor
TL;DR
Tactile research is constrained by sensors that are difficult to customize for varied robot geometries and by raster readouts that can miss brief contacts. OpenTac combines a customizable shear-and-normal tactile skin with design, simulation, transfer, and adaptive-acquisition tools. Its adaptive scans reach 1286 Hz, and the ecosystem is evaluated in real-robot control and policy learning.
Problem
Tactile research is constrained by fixed sensor geometries or the need for design and fabrication expertise, while conventional 30–100 Hz raster scans can limit response to contacts lasting only milliseconds.
Source: §1 · p. 2Method
OpenTac combines a piezoresistive skin that estimates local normal and shear forces with tools for sensor design, simulation, cross-sensor transfer, and adaptive acquisition.
Source: §1 · p. 2Results
1286 Hz adaptive acquisition on a 32 × 32 skin exceeds conventional full-array scans at 67 Hz by more than 19×; the ecosystem was evaluated on over 30 layouts and four real robot embodiments.
Source: §1 · pp. 2–3Takeaways & Limitations
Sensors with different shapes, resolutions, and electrical layouts can share acquisition and processing tools and a common interface for downstream control and learning.
Source: §1 · p. 2Flexible PCBs are not stretchable, mesh-derived layouts require developable patches, and adaptive acquisition trades spatial completeness for update rate.
Source: §5 · p. 12
Abstract
from arXiv · showhide
We present OpenTac, a novel sensor and a full-stack ecosystem for tactile skins that are highly customizable, scalable, and readily deployed in downstream robot control and learning applications. Our core contribution is twofold. First, we develop a scalable resistive tactile skin with a structured deformable overlay that converts tangential loading into spatially varying pressure. Differential responses across neighboring "taxels" support decoding of local shear vectors, and the common-mode response encodes normal force. Second, we provide a highly customizable sensor design tool that is able to automatically generate fabrication files for these tactile skins from 2D outlines and developable 3D surfaces, including from abstract descriptions such as text (e.g., "design a sensor for a size 9 shoe insole"). We develop an ecosystem around the OpenTac sensor, including the CAD design tool, a sensor simulator, cross-sensor transfer, and fast/adaptive data acquisition. We demonstrate the broad applicability of the OpenTac-CAD sensor design tool by designing and fabricating more than 30 sensors across a variety of robot embodiments. We characterize the three-dimensional tactile fields produced by OpenTac with a commercial, high-precision force/torque sensor. We demonstrate the broad applicability of 3D tactile sensing across contact-rich feedback control and manipulation policy learning on four real-world robot embodiments. Additionally, adaptive acquisition enables kilohertz-rate partial readouts, exceeding full-array raster scanning on the same sensor by 19.2X at 16 sampled taxels per update for a 1024-taxel skin. We provide all hardware, software, and result videos are at opentac.github.io
1 Introduction
OpenTac addresses the difficulty of fitting tactile skins to varied robot geometries while adding local shear sensing to normal-force sensing. Its ecosystem combines automated design, simulation, transfer, and adaptive acquisition, with evaluations across fabricated layouts and real robots.
1 Introduction
- Conventional raster scanning at 30–100 Hz can limit response to contacts lasting only a few milliseconds, while geometry changes also require compatible acquisition and processing software.Researchers otherwise must adapt work to fixed sensor geometries or have the expertise to design and fabricate custom sensors.Source: §1 · p. 2
- OpenTac’s piezoresistive skin decodes local shear from pressure redistributed across neighboring elements while retaining normal-force information.Its sensing layout, electrode geometry, and mechanical overlay can be customized to the target surface and force range.Source: §1 · p. 2
- OpenTac-CAD generates fabrication files from 2D outlines, selected developable robot surfaces, or verbal descriptions, while manifests support simulation and cross-sensor transfer.OpenTac-Sim uses the manifest to simulate sensors and interactions; OpenTac-Transfer maps calibrated observations using physical coordinates and local surface frames.Source: §1 · p. 2
- 1286 Hz adaptive acquisition on a 32 × 32 skin exceeds conventional full-array scanning at 67 Hz by more than 19×.The method uses adaptive partial scans for sparse contact events.Source: §1 · p. 2
- The ecosystem was evaluated on more than 30 fabricated sensor layouts deployed across four real robot embodiments, including contact-rich feedback control and manipulation policy learning.The tactile force estimates were also evaluated against a commercial force/torque sensor.Source: §1 · p. 3
- OpenTac hardware, software, and firmware are released under permissive licenses for replication, modification, and redistribution in research and commercial use.Source: §1 · p. 3
2 Related Work
Prior tactile systems use varied transduction methods and address customization, simulation, and acquisition in different ways. OpenTac combines a single-FPCB sensing architecture with robot-geometry-driven design and shared tools for simulation, transfer, and adaptive sampling.
2 Related Work
- Mechanical structures can encode shear as differential pressure, while OpenTac uses a single-FPCB architecture designed for automated customization and shared readout.The cited examples include optical, magnetic, capacitive, and resistive sensing, as well as mechanical approaches to estimating shear.Source: §2 · p. 3
- Prior customization approaches automate selected components or provide conformable coverage, whereas OpenTac jointly generates the sensing layout, routing, connectorization, mechanical structure, and metadata from robot geometry.Source: §2 · p. 3
- TACTO and Taxim simulate vision-based tactile imaging, while DiffTactile, TacEx, and Taccel also model surface deformation; OpenTac-Sim uses compiled geometry and bump layouts for its sensors.HydroShear additionally models shear and stick-slip effects for sim-to-real learning.Source: §2 · p. 3
- OpenTac-Xampling adapts sampling across compiled layouts, and OpenTac-Transfer maps calibrated contact fields through physical coordinates and local frames without paired cross-sensor data.Prior approaches include event-driven hardware, adaptive subsampling, and sensor-agnostic representations learned from aligned synthetic and real-world data.Source: §2 · pp. 3–4
3 The OpenTac Ecosystem
OpenTac combines a customizable three-axis tactile skin with shared design, simulation, transfer, and adaptive-readout tools. Its modules use a common sensor manifest to support deployment across varied robot geometries and applications.
3 The OpenTac Ecosystem
- The shared-manifest ecosystem comprises CAD fabrication, MuJoCo simulation, cross-sensor geometry transfer, and adaptive readout modules for robotics applications.The sensor stack combines a flexible PCB, piezoresistive film, and structured deformable overlay.Source: §3 · p. 4
3.1 OpenTac: Normal and Shear Sensor design
- Structured bumps redistribute tangential loads across taxels, whose differential signals encode shear, while normal loads produce a common-mode response that can be calibrated to force units.The manifest defines taxel groupings and signs for each compiled array.Source: §3.1 · p. 4, §3 · p. 4
3.2 OpenTac-CAD: Automatic sensor design
- OpenTac-CAD compiles 2D outlines or developable 3D surfaces into routed boards, recording physical taxel geometry separately from scan order in a shared manifest.Doubly curved selections that cannot be flattened are detected and reported.Source: §3.2 · p. 5
3.3 OpenTac-Sim: Simulating OpenTac sensors
- OpenTac-Sim instantiates arbitrary sensor shapes from the manifest, using FEM for bump-covered normal-and-shear skins and a reduced foam-contact model for smooth normal-only skins.The current release supports stable sensor-object simulation only with rigid objects.Source: §3.3 · p. 6
3.4 OpenTac-Transfer: Cross-sensor transfer
- OpenTac-Transfer maps calibrated normal-contact fields between skins in physical coordinates, using contact alignment and per-sensor calibration; force is conserved when the target covers the contact.Predicted target counts apply inverse target calibration, including its noise floor and saturation.Source: §3.4 · p. 6
3.5 OpenTac-Xampling: Adaptive Acquisition
- OpenTac-Xampling uses a limited adaptive sampling budget to search the full mask and follow previously active regions, working across layouts and degrading to raster-scan complexity in the worst case.The method relies on contacts being sparse, local, and temporally coherent.Source: §3.5 · p. 6
4 Experimental Evaluation
The evaluation reports results across design generation, sensor characterization, simulation, transfer, control, and adaptive acquisition. Fabricated designs were functional, and shape-aware packing improved coverage on some irregular regions relative to rectangular grids.
4 Experimental Evaluation
- Experiments report clean CAD layouts, accurate three-axis estimates, faithful simulation, downstream control and policy results, cross-sensor force transfer, and kilohertz-rate Xampling.The evaluation covers each component of the OpenTac ecosystem.Source: §4 · p. 7
4.1 Evaluating auto-generated robot skin designs
- OpenTac-CAD designs had zero design-rule violations and functional fabricated boards; its coverage reached 75%, 90%, 89%, 87%, and 93% across five regions, versus 57%, 90%, 79%, 87%, and 93% for rectangular grids.Coverage gains were 18 percentage points for YUBI and 10 for LEAP, with near-identical results on nearly rectangular regions.Source: §4.1 · p. 7
- OpenTac-Sim was evaluated by comparing simulated and real-world force readings and contact depths for a calibrated shear sensor pressed with 1–5 N.The experimental findings describe the simulation as faithful.Source: §4 · p. 7, Figure 8 · p. 8
- All nine language-agent design sessions produced DRC-clean layouts with 76–254 taxels, recording dimensional sources and design assumptions.The median session took 6.5 minutes.Source: §4.1 · pp. 7–8
4.2 Shear, material, and geometry Characterization
The characterization examines how material and electrode geometry affect force response, then measures OpenTac’s force estimates against a reference sensor.
Choice of piezoresistive film
- Chibitronics saturates at low force, LINQSTAT avoids the saturation criterion through 60 N, and Yunjia has the lowest ADC response over the tested range.These comparisons use 2.0 mm pitch and a 0.2 mm electrode gap.Source: §4.2 · p. 8
Electrode geometry
- A 2 mm Chibitronics sensor has a 60 N force range, while increasing pitch to 4 mm causes saturation at about 20 N.Larger pitch increases per-taxel signal, trading spatial resolution against sensitivity; electrode gap has a modest response effect.Source: §4.2 · p. 8
Force estimation accuracy
- R² reaches 0.95, 0.93, and 0.96, with RMSE of 0.11, 0.16, and 0.15 N for Fx, Fy, and Fz, respectively, against OptoForce.The estimate uses a 4×4 bump-strip array with LINQSTAT film, 4.0 mm taxel pitch, and 0.2 mm electrode gap.Source: §4.2 · p. 8
4.3 Shear sensor simulation
The study compares simulated and real shear-sensor responses under matched force sequences and tests simulation for policy learning.
Simulation validation
- Simulated measurements are extremely well correlated with mean real-sensor data, while real measurements show signal-dependent noise that generally increases with force.The matched pressing sequence spans 1–5 N in 0.5 N increments.Source: §4.3 · p. 8
- Contact depth and force are strongly correlated in both real-world measurements and OpenTac simulation.A simulated peg-insertion policy is also demonstrated qualitatively.Source: §4.3 · p. 8
4.4 Tactile Policy Learning
OpenTac tactile policies are evaluated on contact-rich insertion tasks, where shear-sensing variants outperform normal-only alternatives in the reported comparisons. The experiments span SO-101 and bimanual YAM robots.
SO-101 robot arm
- OpenTac-3D-full has the highest observed success rate on both SO-101 tasks; plug insertion reaches 70% versus 54% for its pose-only baseline.On peg-in-hole, success rises from 66% to 82% with OpenTac, compared with 50% to 58% for FlexiTac and 52% to 60% for OpenTac-1D.Source: §4.4 · p. 9, Figure 11 · p. 9
- For plug insertion, normal force may not distinguish socket alignment from surrounding-surface contact, while shear estimates vary between airborne and inserted states.Neither normal-only tactile policy improves on its corresponding proprioception-only baseline in these trials; the authors suggest shear is essential, as also reported in [40].Source: §4.4 · pp. 9–10
Bimanual YAM robots
- OpenTac 3D achieves 62% success on key-in-lock insertion, versus 43% for OpenTac 1D.The 25-trial bimanual task requires twisting and inserting a key that fits only in a specific configuration; shear helps detect whether it is in the slot, on the edge, or out of contact.Source: §4.4 · p. 10
4.5 Contact-rich control tasks
Two contact-rich control studies compare shear feedback with normal-force control or other feedback modes.
Ball manipulation
- The UR5e shear controller more reliably recovers ball rolling from the “Borderline” to the “Valid” region than normal force alone.The 32 × 32 OpenTac sensor regulates toward a target shear vector without ball-position feedback.Source: §4.5 · p. 10
Bimanual grasp stability under impulsive loading
- In the cup-stability task, the shear controller raises the nominal grip target before slip under impulsive loading.The Dexmate Vega-1 Pro comparison holds trajectories and gains constant; Table 2 reports 40-second holds and net cup displacement.Source: §4.5 · p. 10
4.6 Cross-Sensor Transfer
OpenTac-Transfer maps pose-conditioned normal-contact fields between sensor designs, outperforming two resizing baselines. Its evaluation concerns contact shape rather than force magnitude or sim-to-real policy transfer.
4.6 Cross-Sensor Transfer
- Median F1 was 0.750 over 5,290 predictions, versus 0.714 for resize + manifest and 0.634 for resize without a manifest.Four LINQSTAT skins were tested on five objects at 95 locations; predictions received the contact pose.Source: §4.6 · pp. 10–11
- On support covered by both skins, transfer reached 0.788 F1, close to the 0.800 same-skin repeatability.Coarser 4-mm targets scored 0.84–0.88 versus 0.67–0.78 for 2-mm targets; the foam model reached 0.727 on three skins excluded from its fit.Source: §4.6 · p. 11
- The evaluation measures pose-conditioned normal-contact shape, not force magnitude or sim-to-real policy transfer.Source: §4.6 · p. 11
4.7 High-Speed Adaptive Sampling
Adaptive partial scanning raises update rates and captures brief impacts across sensor layouts. Endurance testing found stabilized readings after an initial response increase, with relative force magnitudes remaining separable through ten hours.
4.7 High-Speed Adaptive Sampling
- 1286 frames/s at M = 16 was 19.2× the 67 frames/s raster rate on a 32×32 skin.Raster scanning reads all 1024 taxels; the adaptive rate rose as the sampling budget shrank.Source: §4.7 · p. 11
- At budgets M = 16–256, partial scans detected all three tested tennis-ball impacts per condition, while raster scanning detected two of six.At M = 512, two of three impacts were detected; the fastest settings recorded several frames of the roughly 3 ms event.Source: §4.7 · p. 11
- The same firmware ran across square, L-shaped, circular, and gripper-pad layouts without a hand-coded raster or neighborhood map.At M = 64, it read 6% of a 1024-taxel square and 63% of a 102-taxel pad per frame.Source: §4.7 · p. 11
4.8 Endurance (duty cycle) testing
- After a gradual force-response increase during the first two hours, readings stabilized and remained clearly separated in relative magnitude throughout the 10-hour test.Testing applied 3–5 N every 30 minutes; the authors recommend a warmup phase or relative rather than absolute ADC readings.Source: §4.8 · p. 11, Figure 14 · p. 11
5 Conclusion
OpenTac combines sensor geometry, fabrication, simulation, transfer, and high-speed acquisition in an open ecosystem. Its limitations concern material and surface constraints, sensing behavior, sampling trade-offs, and the scope of transfer evaluation.
5 Conclusion
- Flexible PCBs are not stretchable, and mesh-derived layouts require surfaces that can be flattened into developable patches.Source: §5 · p. 12
- Piezoresistive sensing introduces hysteresis, drift, and temperature dependence; adaptive acquisition trades spatial completeness for update rate, especially over larger contact regions.Source: §5 · p. 12
- Simulated signals are deformation proxies, while transfer evaluation covers normal-contact shape rather than force accuracy or sim-to-real policy performance.Source: §5 · p. 12
Contribution Statements
The contribution statements assign overall project leadership and core sensor and software development to Ariel Slepyan, alongside collaborators’ experimental, tool, and advisory contributions.
Contribution Statements
- Ariel Slepyan initiated and led the project, developed the sensing architecture, hardware, electronics, firmware, CAD, Transfer, and Xampling, and co-developed Sim.Slepyan also coordinated the collaboration, designed experiments, and led manuscript preparation.Source: Contribution Statements · p. 14
- Geeta Chandra Raju Bethala contributed to OpenTac-CAD and ran DexMate Vega Pro experiments, while Jiafeng Gu led OpenTac-Sim development and its efficacy and accuracy evaluations.Source: Contribution Statements · p. 14
- Yinzhe Zhou and Tuhin Basu led sensor characterization and calibration; Basu also led contact-rich control, while Hongyuan Fu and Xinyi Yin led SO-101 policy learning.Md Faizal Karim led YAM bimanual manipulation experiments, and Jade Sim designed and fabricated a tactile glove.Source: Contribution Statements · p. 14
- Krishna Murthy Jatavallabhula advised on experiment design and framing and contributed to manuscript writing; Gregory Hager and Nitish V. Thakor provided scientific supervision and guidance.Anagha Rajendra Dangle coordinated sensor design and experiment consistency; Homanga Bharadhwaj and Anand Bhattad contributed to advising, planning, and project discussions.Source: Contribution Statements · p. 14