On August 7, The Robot Report published a long interview with Vikram Pavate, co-founder and CEO of Tacta Systems, walking through the TactaBot the company had put on the wire from Palo Alto on July 27.
Pavate gave the company its one-line thesis: 「Humans have amazing brains, but AI is catching up. We have decent eyes, and that’s an area where computer vision, arguably, has caught up. The parts that have been missing are the hands.」
That claim is not new. What Tacta picked as the fix is.
Start with the hand
TactaBot is three things: the Tacta Hand, a Skill Capture system, and a model the company calls Dexterous Intelligence.
The hand is human-sized, five fingers, 15 degrees of freedom at three per finger. Each fingertip lifts up to 25 newtons, roughly 2.5 kg. Actuation is Tacta’s own Fluidic Tendon design: fluid runs through cabling and pushes pistons that move the fingers.
Pavate’s reasoning for that choice is mechanical, not marketing. Fluid lets the fingers be built very thin, and nothing near the hand generates heat, neither of which motor-driven hands manage. Reliability is the other argument. Conventional cable-driven hands are fast and precise, but one cable fails and the whole hand is down. Tacta claims a mechanical life of up to 30 million cycles.
Fingertips and sensors are scored separately. Both wear out, so the company designed them as consumables.
The sensor is where the engineering sits. Each one is smaller than a grain of sand and reads from 250 pascals to 700,000 pascals, which covers the full dynamic range of human touch. It also reads temperature to 0.1 degrees Celsius. These are not MEMS parts; each carries its own analog-to-digital conversion, laid out at a 1 mm pitch.
That is the hardware. It is good hardware, and it is not what makes this a labor story.
The product is the glove
Tacta put the same sensor into a glove. 256 per fingertip, with motion capture across the back of the hand, streaming wirelessly to a local host.
Two lines from the interview belong next to each other.
The first: the glove deploys immediately, without training, and workers wear it naturally as they complete complex tasks.
The second, in the company’s own words: every hour of work performed with the Tacta Glove further compounds its robot training dataset.
What the glove records is force, motion, video, and temperature, taken directly from workers on the factory floor. Tacta then takes an off-the-shelf pre-trained model, trains it on that tactile and motion-capture data, and fine-tunes on the customer’s own operations with gloves on the customer’s own line.
Line those three steps up and the displacement path has a shape nobody has drawn before. There is no moment when the robot arrives and the worker leaves. The worker keeps clocking in and keeps doing the work that needs feel in the fingers. The only change is a glove. The skill stays in the hands. The record of the skill does not.
There is no Reddit for the physical world
Pavate is blunt about the data problem: 「There’s no Shakespeare and the web and Reddit and The New York Times to train these models in the physical world, especially for high value work.」
That is the real constraint on embodied AI, and it is what separates the competing bets. Language models found a corpus lying around. Physical intelligence has to manufacture one, and the manufacturing method is the strategy.
NVIDIA open-sourced Isaac GR00T N1.7 on April 28, built on 20,854 hours of egocentric video under Apache 2.0, with the first published scaling law for dexterous manipulation attached. We wrote that one up at the time. Egocentric video is cheap, abundant, and scrapeable. It also contains no force, no slip, no temperature, and no deformation.
DeepMind’s Gemini Robotics 2 landed July 30, controlling a full humanoid end to end for the first time. Its own published benchmark carried an uncomfortable pair of numbers: 92% success unscrewing a light bulb, 36% screwing it back in. We read that chart as five-fingered hands losing to two-finger grippers across a range of tasks.
Tacta’s entire bet is an answer to where that 36% comes from. Its answer is not that the model is weak. It is that the hand does not know what it is touching. So the company went after the input layer rather than the model layer, and specifically after the one channel you cannot scrape.
The July 27 release calls this large-scale skill capture. The phrase is accurate. It just needs reading slowly.
Agility pays for its data. Tacta gets paid for collecting it.
On July 30, Agility Robotics opened a Fremont hub and started hiring roughly 200 people to teach Digit warehouse work, while going public at a $2.5B valuation through a SPAC. We logged that.
Both companies are converting human motion into robot capability. The books are nothing alike.
Agility’s training data is a payroll line. It hires the people, pays them, and generates the data on its own floor, where the cost lands on the income statement. Marginal cost of data equals marginal cost of labor, which means the ceiling on dataset size is whatever the company is willing to burn.
Tacta hires nobody. The labor generating the training data is already employed, by the customer, on the customer’s payroll, filling the customer’s orders. Tacta ships a capture system, the customer pays for the capture system, and the corpus comes out the other side.
That difference is structural, not incremental. In the first model data is a cost. In the second it is a byproduct of revenue. Byproducts do not compete for budget, so they have no ceiling. When Pavate says every hour compounds the dataset, compounding is not a figure of speech. It describes a curve that climbs without further investment.
This tier of work was supposed to be safe
Pavate frames the business as a response to labor shortage: some of the worst workforce gaps in the world are in skilled trades, and manufacturing worst of all.
In the near term that is true. There genuinely are not enough people willing to do precision electronics assembly, and the gap is real.
But this tier was considered automation-resistant for exactly the reason Tacta is targeting it. The first wave of industrial automation took the repetitive, high-volume, fixed-path motions. What survived is the work that requires judgment in the fingertips: whether a connector has seated fully, how much force a flex cable will take, whether a stuck part wants easing or twisting, what a slight deformation in a component surface is telling you. None of that runs through the eyes. It runs through touch.
Tacta names electronics, AI infrastructure, and automotive as its first industries, with reshoring as the tailwind. First TactaBots ship in early 2027, and the customer-site fine-tuning happens before that.
Pavate’s own long-range examples go further: nursing homes, hospitals, food preparation. He offers a world-class sushi chef wearing the gloves through prep, or a pianist wearing them through a piece. Both examples describe the same operation in more flattering light. Skill in the hands can be taken off the hands.
If gloves show up on your line
Three questions worth asking this quarter.
Ask who owns the capture and what it trains. Instrumented wearables usually enter a plant under one of two banners, ergonomics or quality traceability. Tacta is unusually straightforward that its device exists to train the model that will do the motion. The vendors that are not that clear are the ones to press.
Check whether your craft has ever been recorded, not whether it is hard. Touch had no dataset because nobody could afford the sensors, not because it resisted recording. Once a tactile sensor is cheap enough to sell as a consumable, that assumption is gone. Exposure tracks whether a process gets logged, not how difficult it is.
Move toward defining and adjudicating. Hands stay valuable for the next few years precisely because the corpus is thin. The durable position is owning what counts as acceptable, handling the exceptions the machine cannot resolve, and knowing which step to change when yield drops. No glove is capturing any of that.
TactaBot ships in early 2027. Every shift between now and then is feeding it.
