Microchip Brings Sub-Femtofarad Sensitivity to Robotic Tactile Sensing
Microchip has introduced the ATFCD52, a high-speed capacitance digitizer that converts signals from fine-grid two-dimensional capacitive sensor arrays into digital data an external processor can use to calculate normal and shear forces. It supports up to 52 channels, arranged as 32 transmit by 20 receive lines, acquires all receive lines simultaneously through 20 parallel analog front ends, and reaches frame rates up to 800 Hz depending on array size. It resolves capacitance changes below one femtofarad. The application it targets is robotic touch: grippers and dexterous humanoid hands that adjust a grasp based on what they can feel rather than only what they can see, which is a harder problem than it sounds.
Why touch is harder than vision
The framing in the release is worth taking seriously. Cameras can locate an object, and locating an object is a problem the industry has largely solved. Knowing how firmly it is held, and whether it has begun to slip, is a different problem, and vision cannot answer it because the answer depends on forces the camera cannot see.
That gap matters most for the objects that are hardest to handle: soft, fragile or irregular items where the correct grip force is a narrow band. Too little and the object drops. Too much and it crushes. A system that only sees has to infer the force from a model of the object, and that model is wrong often enough to matter.
Giovanni Fontana, senior director of Microchip’s human machine interface division, makes the point in the release: robots have become increasingly capable of seeing and understanding their surroundings, while the ability to feel and respond to physical contact remains a critical challenge. That is a fair summary of where manipulation research currently sits.
Why the timing is right
The humanoid context explains the timing of the launch. Humanoid robots have moved from research demonstrations to funded engineering programmes in the past two years, and the binding constraint has shifted from actuation and locomotion to manipulation. Walking is largely a solved problem for a well-funded team; picking up an egg without breaking it is not.
That shift creates demand for exactly this class of component. A humanoid hand needs tactile sensing at every fingertip, which means many sensor sites per hand, which means the digitizer has to be small and affordable enough to repeat. Dense tactile sensing at that scale is the problem this part addresses, and it explains why the specification reads the way it does.
What the part actually does
The specification repays close reading because two of the numbers describe different things and the release does not spell out the relationship.
The device supports up to 52 channels, which are 32 transmit lines plus 20 receive lines. A capacitive sensor array is a grid: each transmit line crosses each receive line, and the capacitance at every crossing is a sensing point. Thirty-two by twenty gives 640 potential nodes, so the 52 channels describe the wiring while the array itself can carry 640 measurements across the contact surface.
That distinction matters for anyone sizing a design. A reader who takes 52 channels to mean 52 sensing points would badly underestimate the resolution available, and a reader who takes 640 to mean 640 wires would badly overestimate the routing burden. The part sits between those two numbers, and the release assumes a reader who knows how a capacitive matrix works.
The frame rate is the point
The most consequential figure in the release is 800 Hz, and the reason is timing rather than resolution.
A slip event develops over tens of milliseconds. If a gripper needs to detect the onset of a slip and correct its grip before the object falls, it needs enough samples inside that window to see the trend rather than a single reading. At 800 Hz a sample arrives every 1.25 milliseconds, which puts roughly two dozen readings inside a 30-millisecond slip window. That is enough to distinguish a developing slip from a momentary noise spike and to command a corrective grip.
The architectural choice that delivers it is the twenty parallel analog front ends. A conventional design might multiplex the twenty receive lines through a single front end, which divides the frame rate by twenty and makes 800 Hz impossible. Acquiring every receive line at once is what buys the speed, and it is the design decision the rest of the specification depends on.
The calculation the processor does
One detail is easy to pass over and worth dwelling on: the digitizer produces capacitance data, and an external processor calculates normal and shear forces from it.
That division of labour is a design choice with consequences. Normal force is the pressure of contact, and it is the easier of the two to derive. Shear force is the sideways component, the one that indicates an object sliding against the fingers, and deriving it requires comparing signals across neighbouring sensing sites rather than reading one in isolation. Shear is the harder calculation and the more useful signal for slip detection, which is why the release names it explicitly rather than mentioning force in general.
Sensitivity in a noisy environment
Resolving capacitance changes below one femtofarad is an impressive figure on its own. The harder part is doing it inside a robot.
A robot hand sits next to servo motors, power electronics and switching regulators, all of which generate electrical noise in the same band a capacitive sensor operates in. A change of less than a femtofarad is small enough that this noise can swamp it, which is why light touch is the hardest case: the signal is smallest exactly where the environment is most hostile.
The release names two mitigations, frequency hopping and oversampling, and both are the right tools. Frequency hopping moves the measurement away from a noise source that occupies a fixed band. Oversampling averages many readings to separate a consistent signal from random noise. Together they address the problem at the point where it actually occurs, and the inclusion of both in a product brief is a sign the design was tested against real hardware rather than a bench.
The development tools and the wider market
Two pieces of support material are worth noting. The EV-ATFCD52 development board lets a customer evaluate capacitive sensor designs, tactile sensor materials and stack-ups, which are the three variables that determine whether a sensing design works in practice. The ATFCD52 Configurator software handles parameter configuration and real-time data logging.
The release also names applications beyond robotics: augmented reality and gaming accessories, data-collecting gloves and foot mats, and wearable sensing such as electronic skin. Each of those is a natural fit for the same capability, because each requires dense capacitive sensing in a package small enough to wear and a rate high enough to feel responsive.
The processing story is deliberately open. The digitizer pairs with Microchip MCUs, MPUs or FPGAs depending on performance requirements, which lets a design choose between localized control and higher-throughput computation. Making the force calculation the external processor’s job keeps the digitizer general and puts the application intelligence where a customer wants it.
Electronic skin is the application to watch
Electronic skin is the application with the most room to grow. A robotic hand with tactile sensing at each fingertip is useful; a surface with tactile sensing everywhere is a different capability, and it changes what a robot can safely do near people.
That application is also the most demanding on power, because skin covers area and area multiplies sensor sites. It is why the missing power figure matters more than it might appear: for a hand, power is a design consideration, and for skin it is the constraint that decides whether the approach works at all.
What would complete the picture
Three specifications would turn this from a well-described component into a fully specified one.
Power consumption is the first, because two of the named applications are wearables and electronic skin, where the power budget decides whether a design is viable at all. Twenty parallel front ends running at 800 Hz draw current, and the figure matters more than the frame rate for anything battery-powered.
Price is the second, since the release says the part is available from Microchip or distributors and gives no indication of cost. For a component going into a robot hand with dozens of sensing sites, unit economics decide the architecture.
And the latency from sample to force estimate is the third. The digitizer produces data; the external processor turns it into normal and shear forces; the control loop acts on those forces. The 800 Hz figure describes the first step only, and the number that decides whether a grip correction lands in time is the total across all three.

What to watch
Whether a robotics customer describes a design using the part, because a named gripper or hand would convert a specification into a case. And, whether power figures appear alongside the wearables applications, since electronic skin is a power-limited problem. Whether the force calculation moves onto the device in a later variant, which would simplify integration considerably. And whether sub-femtofarad capacitive sensing at these rates becomes a common distinguishing one, because that is the point at which dexterous manipulation stops being a research problem and becomes an engineering one for the whole field.

