[Paper Notes] OpenTac: An Open Ecosystem for Customizable, Scalable, and Transferable 3D Tactile Sensing
Published:
TL;DR
OpenTac combines customizable tactile hardware with tools for design, simulation, cross-sensor mapping, and adaptive readout. Its sensing mechanism places soft bumps over groups of piezoresistive taxels: normal force changes their common response, while shear tilts the bump and creates a spatial difference.
The robot experiments show why shear can be useful for insertion. On SO-101 plug insertion, the three-axis sensor’s tactile policy reaches 70% success, compared with 54% for its own proprioception-only baseline. The broader contribution is a shared sensor description that connects fabrication geometry to the software consuming its measurements.
Paper and reading version
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. These notes follow arXiv:2610.31661v1, an arXiv preprint, checked on October 25, 2026. PDF · Project
These notes follow Sections 3–5. The official repository contains the hardware and software ecosystem.
A mechanical layer makes shear visible to a resistive array
The sensor stack contains a flexible PCB with interdigitated electrodes, a piezoresistive film, and a deformable bump cover. Applied pressure changes the local resistance. Each bump can cover a $2\times2$ group of taxels.
Let $z_{uv}$ be the baseline-subtracted readings, with $u,v\in{-1,+1}$ identifying positions relative to the bump center. The raw three-axis signals are
\[h_z=\sum_{u,v}z_{uv},\qquad h_x=\sum_{u,v}u z_{uv},\qquad h_y=\sum_{u,v}v z_{uv}.\]A normal load produces a roughly common response. Tangential loading tilts the bump, redistributing pressure across the group. The sums and differences expose that change; an optional affine calibration maps the signals into force units.
This design also changes contact mechanics. Adding a bump cover affects friction and grasp geometry, so a better policy with bumps cannot automatically be attributed entirely to extra information. The paper addresses this partly by training a separate proprioception-only baseline for each hardware configuration.
One manifest connects fabrication and interpretation
OpenTac-CAD accepts drawn outlines, DXF geometry, or selected surfaces of an STL mesh. A language-model route can propose a 2-D outline from text. The compiler then determines taxel placement and routes row/column connections under geometric clearance rules, producing fabrication files and a sensor manifest.
The manifest records physical taxel centers, scan indices, footprints, boundaries, connector mappings, and bump groups. This matters because electrical scan order need not match spatial order on an irregular skin. A downstream model should receive measurements associated with physical locations, not assume that adjacent array entries are adjacent on the robot.
The paper reports 31 fabricated sensor designs. The geometry pipeline supports developable surfaces that flatten into patches. Doubly curved selections that cannot be flattened are detected; the flexible PCB itself is not stretchable.
OpenTac-Sim reads the same description. Bumped skins use a linear-elastic FEM model; smooth normal-pressure skins use a reduced foam-contact model. The current release supports stable contact with rigid objects. Its signals remain modeled deformation responses, with calibration needed to relate them to the real sensor.
Transfer preserves contact geometry under explicit assumptions
OpenTac-Transfer maps calibrated normal-contact fields between sensor designs. Given source cells $A_i$, target cells $B_j$, alignment $T$, and source calibration $C_s$,
\[M_{ji}=\frac{|T(A_i)\cap B_j|}{|A_i|}, \qquad \hat F_j=\sum_i M_{ji}C_s(s_i).\]The inverse target calibration converts predicted target force into readout values, with its noise floor and saturation. Total force is conserved when the target covers the source contact.
This requires known contact alignment and sensor calibration. The transfer experiments use four smooth skins and evaluate normal-contact shape. They do not establish automatic three-axis transfer, absolute-force accuracy, or successful transfer of a learned control policy across embodiments.
Shear helps distinguish contact states that normal load merges
The SO-101 experiments train ACT policies on 100 demonstrations per condition and evaluate 50 trials. Inputs are joint angles with or without tactile signals; vision is absent.
| SO-101 task and hardware | Proprioception only | With tactile input |
|---|---|---|
| Peg-in-hole, FlexiTac | 50% | 58% |
| Peg-in-hole, OpenTac-1D | 52% | 60% |
| Peg-in-hole, OpenTac-3D | 66% | 82% |
| Plug insertion, OpenTac-3D | 54% | 70% |
On plug insertion, neither normal-only tactile policy improves on its corresponding proprioception baseline. Normal pressure can indicate contact while leaving uncertainty about whether the plug is aligned with the socket or pressing against its surroundings. Shear provides additional directional evidence.
The YAM key-in-lock experiment shows the same qualitative preference for three-axis sensing. Its text reports 25 trials alongside percentages that do not all correspond to integer counts at that sample size. I would verify the underlying counts before using those values in a statistical comparison.
Faster readout trades away a complete spatial snapshot
On a 1,024-taxel skin, full raster scanning reaches 67 frames per second. Adaptive sampling with 16 taxels per update reaches 1,286 updates per second, a 19.2× increase. It allocates a quarter of the sampling budget to searching the sensor and the rest to previously active regions.
Those updates are partial readouts, not complete 1,024-taxel images at 1.286 kHz. Sparse, localized contact is the favorable case; broad contact needs a larger budget. The impact experiment shows the benefit for short events, but it does not remove the spatial sampling tradeoff.
The authors report hysteresis, drift, and temperature dependence, with readings settling after roughly two hours in a ten-hour endurance experiment. My priority for using OpenTac would be to keep geometry and calibration in the dataset metadata, then test shear’s benefit against the same hardware without tactile input. That comparison distinguishes sensing value from the mechanical changes introduced by the sensor cover.