Read the full technical article from Jennifer Kwiatkowski on Tech Brief.
For teams building contact-rich manipulation, tactile sensing is shifting from a useful addition to a defensible requirement. Vision-only manipulation has hit a wall, tactile-augmented policies outperform vision-only baselines on contact-rich tasks, and better sensing beats brute-force data scale on cost. The reasons contact data belongs in the training pipeline are, by now, well established.
That leaves a harder question. If a tactile sensor is now a requirement, what should it actually measure, and how do you build one that survives an industrial deployment? This is the engineering problem the TSF-85 was designed to answer.
Slow industrial adoption is not a hardware-maturity problem; capable tactile hardware has existed in labs for decades. It is an interpretation problem. With cameras, resolution, frame rate, and dynamic range map predictably onto performance. Tactile sensing has no equivalent consensus on what signals a useful sensor must capture, at what bandwidth, or at what resolution. That ambiguity carries a cost: a team planning hundreds of thousands of grasps needs confidence that the sensor is capturing the right physical phenomena.
Rather than derive that specification from first principles, Robotiq reverse-engineered it from the system that already manipulates better than any robot ever built: the human hand.
Borrowing the Spec From Human Physiology
The human hand is the best-characterized model of dexterous manipulation available. Johansson and Vallbo’s 1979 study classified its mechanoreceptors into two functional modes. Slowly adapting (SA) units encode sustained pressure, edges, and skin stretch. Fast-adapting (FA) units respond to dynamic events such as vibration and contact transients. The two are not redundant: human grasp control is event-driven, with FA afferents triggering fast slip correction while SA afferents maintain the contact map that regulates grip force.
That physiology hands engineers a concrete target. A tactile sensor for dexterous manipulation must capture static pressure distribution and dynamic contact events, ideally through the same sensing element over the same region, plus a channel for fingertip orientation to interpret the pressure map correctly.
One Dielectric for Three Modalities
The TSF-85 uses capacitive sensing, chosen for the fingertip: no imaging cavity or degrading elastomer like optical sensors, no ferromagnetic constraints like magnetic ones, and manufacturable at industrial scale and cost. The engineering challenge was fitting two distinct capacitive circuits onto a single 22 mm × 37 mm PCB layer without crosstalk.
The static circuit is an array of 28 taxels in a 4×7 grid, mapping pressure across the contact surface as the SA analog. The dynamic circuit is a single taxel around the array’s perimeter, sharing the same dielectric but measuring capacitance change up to 1,000 Hz, spanning both fast-adapting bands. Running both through one shared dielectric eliminates the registration errors and inter-layer crosstalk that plague designs built by stacking separate sensor layers. An integrated IMU completes the picture, supplying fingertip orientation and an independent second source of vibration data.
Built to Survive an Industrial Deployment
Accelerated testing beyond 2 million grasp cycles on an uneven surface shows stable response with no meaningful degradation. Sensor-to-sensor and taxel-to-taxel variance is handled with a simple calibration routine that applies a known load and computes the gain that aligns each output, which brought 37 sensors into alignment at 500 counts under a 100 N load. Because the response exhibits hysteresis, the sensor is optimized for contact detection and orientation estimation rather than absolute force.
Read the Full Engineering Breakdown
The full article goes deeper, covering the complete mechanoreceptor-to-modality mapping, the layered sensor construction, the cycle-testing and calibration data, and the decade of research validating grasp stability prediction, slip classification, in-hand object recognition, and dynamic re-grasping.
Read the full article on Tech Brief.
Ready to take the next step?
Talk to our technical team about tactile integration for your manipulation pipeline and learn more about how Robotiq can enable your application.
