Intelligence is the means, not the end.

Robots have never had more data. They still do almost none of the world’s work.

In under a year, open datasets of human video for robot learning grew from 10,000 hours to millions. Generalist alone trains on more than 500,000 hours of real-world manipulation data. Yet robots are cost-competitive with people for just 0.3% of job tasks in the US, and a quarter of physical tasks are still beyond them entirely.

What’s missing is the step from intelligence to machines people can actually rely on: machines that do meaningful work, at useful speeds, with the precision and reliability the real world demands.

Raising the floor is not enough

General-purpose models are raising the floor of robot intelligence. Tasks that once required carefully engineered systems can increasingly be attempted by a single learned policy. Robots can generalize across objects, scenes, and instructions in ways that would have seemed implausible only a few years ago.

But much of the work we actually want robots to do lies far above that floor. Moving an object from one place to another is only a small part of physical work. People fasten, drill, cut, grind, finish, assemble, inspect, repair, and operate machinery. Each of them depends on sustained interaction with the physical world.

In many of these tasks, the hardest part begins at contact. Take threading a lug nut. A robot can line the nut up with the stud perfectly and still cross-thread it. Whether the threads catch or bind is decided in the first turn, while the motion looks nearly identical either way. The same is true of a drill that cuts or slips, or a grinder that removes material cleanly or begins to chatter. Most of what decides the outcome lives in force, resistance, vibration, sound, compliance, and the evolving dynamics of the interaction.

The next frontier is the work where contact decides the outcome.

Skill should outlive the embodiment

Today, robot skill is often entangled with the embodiment that produced it. Change the end effector, tool, controller, or robot, and much of the data may no longer transfer cleanly.

Much of that entanglement lives in a layer most robot learning treats as fixed: the controller, which turns a robot’s decisions into motion and force. The same command can be a gentle press on one machine and a hard shove on another. Engineers tune this layer by hand, for every robot and every task. And skilled operators adapt to it. On a compliant arm they search into contact; on a stiff one they align first and push carefully.

Grippers, hands, and tools move differently, and controllers respond differently. Underneath, the skill keeps the same structure: what successful contact looks like, how the process evolves, what failure looks like, and how the interaction should change in response.

We believe robot learning should preserve that structure across embodiments and controllers. A demonstration collected on one embodiment should help another. A skill learned under one controller should become a starting point under the next.

The embodiment changes. The skill should persist.

Collecting the signals that matter

The harder problem is signal. Hours of video and teleoperation can show where a hand went, yet say almost nothing about how hard it pressed, or the moment a thread began to bind.

It is increasingly easy to collect more demonstrations, more video, and more trajectories. But the easiest data to collect is often the least informative about the parts of physical work that matter most for learning.

Hand-pose tracking, for example, can capture human motion with almost no interference, but breaks down under occlusion and says little about the forces, contact, or tool state that made the motion successful. At the other extreme, heavily instrumented gloves, teleoperation rigs, and force-feedback systems expose much richer signals, but at the cost of changing how people perform the task and making collection harder to scale.

Better supervision usually comes with more invasiveness; more natural collection usually comes with less information.

We are building data interfaces around that tradeoff: lightweight enough to disappear into real work, but rich and reliable enough to capture the physical signals that make that work learnable. The goal is to capture what a robot needs to reproduce the skill: the forces and contact behind each motion.

Team

• 
• 
• 
• 
• 
• Stanley
  Black & Decker