Why We're Building Dexterous Intelligence

Tacta Systems was founded to address a critical bottleneck in manufacturing: the most dexterous tasks on factory floors still rely on human hands. Connector insertion, alignment, fastening, and delicate component handling have resisted automation for decades. Solving them does not just require better hardware. It requires a new kind of intelligence, one that can reason about contact and force the way a skilled worker does.

Physical AI has made remarkable progress. Robots can now perceive their surroundings, navigate through the world, and perform basic pick-and-place. This is a genuine step change, and it is the foundation everything else is built on. But moving through the world and changing the world are two different problems. The value in a factory is created at the point of contact: the moment a connector seats, a fastener reaches tension, or a fragile part is handled without breaking. Doing that reliably requires dexterity, and dexterity is where today's Physical AI hits a wall.

We call the next step Dexterous Intelligence: the ability to not just move through the world, but to truly manipulate and change it. We believe the unlock for Dexterous Intelligence is touch. This article focuses on the intelligence layer of our full-stack pipeline, where tactile data becomes learned, autonomous manipulation.

The next evolution of Physical AI

The models driving Physical AI are built almost entirely on text, images, and video. They have never felt anything. This is not a minor gap. A connector that is misaligned looks almost identical to one that is seated, right up until you feel it click. A part that is starting to slip looks fine in a camera frame. A fastener at the correct tension is visually indistinguishable from one that is stripped. Touch is the signal that tells a skilled worker what is actually happening at the point of contact, and it is exactly the signal that vision-only models are missing.

This is why locomotion and gross manipulation have advanced quickly while fine, contact-rich manipulation has not. You cannot learn to seat a connector, gauge the tension on a fastener, or handle a part that shatters under too much force from vision alone. The research community has reached the same conclusion: across the literature on contact-rich manipulation, vision-only policies fail in consistent and predictable ways. They proceed past failed grasps without noticing, apply force in the wrong direction despite detecting contact, and cannot modulate grip force for fragile objects. The missing dimension is touch.

We believe adding tactile sensing on top of vision models is the key unlock for Dexterous Intelligence. Touch is the layer that connects perception to truly capable manipulation. With it, we believe robots can reach and ultimately exceed human-level performance on the dexterous tasks that manufacturing depends on, because our sensors offer resolution and dynamic range beyond what human fingertips can perceive.

‍

Why this hasn't been done before

If touch is the unlock, why hasn't Physical AI already incorporated it? Two things have been missing.

First, the world has not had really good tactile sensors. The sensors that exist have been too bulky, too fragile, too low-resolution, or too difficult to integrate to be practical at scale. You cannot build touch into an intelligence layer if you cannot reliably capture touch in the first place.

Second, and more fundamentally, there is no history of collecting touch data. The internet is full of text, images, and video, which is exactly why today's foundation models are built on them. There is no equivalent corpus for touch. No one has been recording what expert hands feel while they work, because no one has had the sensors or the system to capture it. The data that would teach a robot to do skilled, contact-rich work simply does not exist at scale.

Tacta is changing both. We have developed the world's best tactile sensor and, just as importantly, built the system to gather touch data in the volumes that modern learning requires.

‍

How we break the barrier

Dexterous Intelligence rests on three capabilities that must come together. Each is a hard problem on its own; together they are what make touch-driven learning possible.

1) The world's best tactile sensor

Everything begins with the sensor. Ours are high-resolution, with hundreds of sensing elements per fingertip and hand, offering resolution and dynamic range that exceed human touch. Critically, they are also small enough and robust enough to be integrated seamlessly into both a wearable glove and a robotic hand. That dual integration - congruence - is what makes the rest of the pipeline work: the same class of sensor that captures what a human feels during a demonstration is the sensor that lets a robot feel the same thing during execution.

2) A skill capture system that gathers touch data at scale

The single most important unlock is data. It is very hard to get good tactile data from teleoperation or other conventional methods, which are slow, unnatural, and struggle to capture force at all. Our Skill Capture System takes a different approach. It is a wearable, glove-based system that lets a person perform a real task with their own hands while we record the full skill: where the fingers are, how the hand moves, the surrounding scene, and exactly how hard each fingertip presses.

Because the demonstration is performed with natural human motion and without interference, the data reflects real skill rather than a constrained subset of it. And because the sensors are so small and seamlessly integrated into the glove, we capture motion, video, and tactile together in a single natural recording. This is only possible because of our sensors. It is how we build the touch dataset that has never existed before, at the scale Dexterous Intelligence requires.
‍

3) The right hand to embody the intelligence

Intelligence is only useful if it can act in the world. The Tacta Hand is purpose-built to embody Dexterous Intelligence on the factory floor: dexterous enough for contact-rich work, reliable enough to run for millions of cycles, and equipped with the same class of tactile sensors used to capture the training data. Because the hand feels the world the same way the glove does, the intelligence we train transfers directly onto the hardware that executes it.

‍

Our approach to Dexterous Intelligence

With the sensor, the data, and the hand in place, we build the intelligence itself in four steps.

Start with best-in-class robotics foundation models. We do not start from scratch. We build on the strongest available robotics foundation models. These models already encode broad visuomotor priors learned from massive text, image, and video datasets. They are the foundation we extend.

Add tactile signals as both input and output. We expand these models to incorporate touch, leveraging tactile signals on both the input side (so the policy can sense contact, slip, and force) and the output side (so it can act on force targets, not just positions). This is the frontier the research community is actively pushing on, and it is the direction our own work confirms. Touch-reactive skills can be learned at scale and even transfer across different sensors and embodiments.

Fine-tune on data from our Skill Capture System. We fine-tune these tactile-augmented models on the demonstrations gathered through Skill Capture. Beyond simply adding touch, we find that tactile data makes human demonstrations more valuable, because touch gives us a physical signal to align human motion to robot execution. Touch helps bridge the gap between the human hand that captured the skill and the robot hand that performs it.

Leverage Tactile-Aware Reinforcement Learning. Finally, we use reinforcement learning to push success rates higher still. Here too, touch is a force multiplier. We have seen that tactile data drastically speeds up the training rate for RL, because force feedback gives the policy a dense, physically grounded signal about what is working and what is not, rather than relying on sparse visual outcomes alone.

‍

What this enables

The factories building the most high-value and consequential products in the world (Electronics, AI Infra, Automotive) need work done that requires human-level dexterity, and there are not enough skilled workers to meet that demand at scale. Physical AI got robots moving through the world. Dexterous Intelligence is what lets them change it.

By pairing tactile sensing with the best robotics foundation models, fine-tuning on touch data no one else can capture, and sharpening it all with tactile-aware reinforcement learning, we are building the intelligence layer that has been missing from physical AI. Delivered through the Tacta Hand and the broader TactaBot platform, Dexterous Intelligence represents a new paradigm for industrial automation, where the most high-value, contact-rich work in factories is no longer constrained by the availability of human hands.

‍