Sensor hardware, redefined by AI

Touch is the infrastructure of physical interaction. We build force and tactile sensors where the intelligence lives in the algorithm, not in the hardware count.

Why high-performance force sensing stayed expensive

Three structural costs have kept precision force sensing out of reach for volume robotics.

1. Dense sensor arrays. Higher spatial resolution has traditionally meant more physical sensing units, and with them more cost, more wiring and more thickness. Fingertips and joints do not have the room.
2. Manual assembly. A conventional six-axis force sensor requires 24 or more strain gauges, bonded by hand. The process has resisted automation for decades.
3. Calibration. Six-axis joint-loading calibration systems are not commercially available. Every manufacturer builds their own, and every new product requires a full recalibration cycle.

Our approach attacks all three.

Higher resolution has traditionally meant more sensing points. We take the opposite approach.

A MEMS sensing architecture paired with algorithmic decoupling. Sparse hardware, resolved in software.

Sense

MEMS elements in metal or flexible encapsulation, engineered for minimal sensing points and minimal profile.

Decode

Algorithmic decoupling separates the six force and torque components from coupled raw signals.

Data

Real-world tactile time series, the modality simulation cannot synthesise.

Our tactile sensing architecture is peer-reviewed and published at AAAI.

Thin by architecture, not by compromise

Rather than shrinking a conventional strain-gauge design, we rebuilt the measurement architecture. Three levers drive the result:

• New sensing architecture, removing the cost floor set by strain gauge material and manual bonding
• Simplified manufacturing, with 65% fewer process steps
• New calibration model, holding accuracy at 0.2% F.S.

72% cost reduction. 39% performance gain. Up to 1000% overload.

Sample units delivered at 13, 7.2, 6.5, 6 and 3 mm profiles.

Fewer sensing points, higher resolution

A human fingertip localises contact far more precisely than the spacing of its mechanoreceptors would allow. The brain interpolates. We built a sensor on the same principle.

Sensing elements are sealed inside overlapping air chambers. Press anywhere on the surface and every chamber responds, so each element carries a global receptive field of the entire contact area. A neural network resolves the signals into a contact position, using the temporal continuity of motion to reject noise and stabilise the estimate over time.

0.13 mm localisation accuracy. 2507× super-resolution factor. Four physical sensing points.

State of the art at publication.

What high-resolution touch enables

Haptic trajectory tracking

A pen writing on the sensor surface is reconstructed as a continuous trajectory in real time.

Adaptive grasping

As a held bottle fills with water, micro-slip appears as a shift in contact position. Grip force adjusts before the object moves.

Human-robot handover

A lateral displacement pattern signals human intent to take the object. The gripper releases.

Current research directions

We publish what we build, and we are open about where the work continues.

• Normal and shear force decoupling on the overlapping air chamber structure
• Multi-point contact resolution
• Higher-bandwidth signal chains for high-speed slip detection

Research collaboration welcome.