Don't limit yourself to what people can do, but go beyond.
In a Massachusetts robotics lab, a machine once defined by its limitations is being quietly remade into something that exceeds the human form it was designed to approximate. Boston Dynamics' Atlas humanoid robot now moves with a fluidity and range that no biological body could match — a product not just of what was added, but of what was deliberately removed. The company is moving carefully, aware that the distance between a compelling demonstration and a reliable, affordable machine is measured not in code but in time and trust.
- Atlas can now rotate its joints a full 360 degrees, cartwheel, and rise from the floor using only its feet — movements that blur the line between machine and something stranger.
- The removal of wires crossing the robot's joints — a chronic source of failure in earlier models — unlocked a range of motion no longer constrained by engineering compromise.
- An AI brain powered by Nvidia chips learns tasks through VR-guided teleoperation, with three-fingered hands equipped with tactile sensors still being refined for precise grip and force control.
- This summer, Atlas will begin training on a real Hyundai factory floor in Georgia — the first test outside a controlled demonstration environment.
- Boston Dynamics' CEO warns that while software moves at digital speed, building machines that are truly reliable and affordable is a slower, harder discipline — and the hype cycle does not change that.
Five years ago, Atlas moved like a machine — capable but unmistakably robotic. Today, it cartwheels, dances, and twists its torso a full 360 degrees. It can rise from the floor using only its feet and stand in postures no human body ever could.
Much of the breakthrough came from subtraction. Earlier Atlas models had wires running across every rotating joint — a design that limited motion and constantly broke down. The new version eliminates those wires entirely, freeing the robot to move without compromise. When it needs to reverse direction, it doesn't turn around. It simply pivots its core 180 degrees and keeps going.
The robot's AI, powered by Nvidia chips, learns through teleoperation: a human operator wearing VR gear demonstrates tasks — stacking cups, tying knots — until the neural network absorbs the pattern and can execute it alone. The hands, each with three reconfigurable digits and tactile sensors, can shift between gripping a small object with an opposing thumb or spreading wide for something larger. Precise control of grip shape and force, the company acknowledges, remains an unsolved problem.
CEO Robert Playter is measured about the moment. The hype around humanoid robots is real, and the financial predictions are enormous — but he draws a firm line between the speed of software and the patience required to build machines that actually work reliably and affordably in the world.
Since the original broadcast, Boston Dynamics has already revealed a taller, stronger Atlas. This summer, it will begin training at a Hyundai factory in Georgia — its first deployment in a real manufacturing environment. Whether the engineering can keep pace with the ambition is the question the factory floor will begin to answer.
Five years ago, when 60 Minutes first walked through Boston Dynamics' offices, Atlas was a marvel of engineering that moved like a machine—bulky, stiff, capable of running and jumping but unmistakably robotic in its gait. The humanoid robot has since undergone a transformation so complete that it barely resembles its earlier self. The new Atlas can cartwheel and dance with fluid, almost human-like motion. It can twist its arms, head, and torso in full 360-degree rotations. It can pick itself up off the floor using only its feet. It stands in ways no human body ever could.
The breakthrough lies partly in what Boston Dynamics removed rather than what it added. The earlier Atlas models had wires running across the joints of the limbs, torso, and head—a design that limited rotation and created a chronic maintenance problem as those wires frayed and broke over time. The new version has no wires crossing those rotating parts at all, which means the robot's range of motion is no longer constrained by engineering compromise. Scott Kuindersma, the company's head of robotics research, explained the philosophy: the robot isn't built to mimic human limits. Instead, it's designed to exceed them. When Atlas needs to turn around, it doesn't walk in a new direction the way you or I would. It simply pivots its core 180 degrees and continues forward.
The second major upgrade is the robot's brain. Atlas now runs on an AI system powered by Nvidia chips, and it learns through a process called teleoperation. A human operator puts on virtual reality gear and controls the robot remotely, teaching it to perform specific tasks by demonstration. The operator might show Atlas how to stack cups or tie a knot, repeating the task until the robot's neural network absorbs the pattern and can execute it independently. This is how Boston Dynamics is building a machine that can adapt to different jobs without being reprogrammed from scratch each time.
The hands represent perhaps the most intricate engineering challenge. Human hands are extraordinarily versatile—we use the same hand to pick up a grain of rice or carry a gallon of milk. Atlas has three digits on each hand, but they're not fixed in position. Each digit can swing into different configurations, allowing the hand to function as a traditional three-fingered grip, or to reconfigure so one digit acts as a thumb for opposing grasps on small objects, or to spread wide for handling large items. The fingers are equipped with tactile sensors that feed information back to the robot's neural network, teaching it how much pressure to apply when gripping different materials. Even so, Kuindersma acknowledged that precise control of grip shape, motion, and force remains an open problem. The teleoperation systems that train the robot are still being refined.
Boston Dynamics' CEO Robert Playter is careful about the language of possibility. Yes, there is significant hype around humanoid robots right now. Financial institutions are predicting millions or billions of robots in our future. But Playter distinguishes between the speed at which software can advance and the pace at which reliable physical machines can be built and deployed. Software can move at digital speeds. Machines require time. They have to work reliably. They have to be affordable. Those constraints don't disappear because the vision is compelling.
Since the original 60 Minutes segment aired in January, Boston Dynamics has already unveiled yet another iteration—a taller, stronger version of Atlas. This new model will begin training at a Hyundai factory in Georgia this summer, marking the first time the robot will work in an actual manufacturing environment rather than a controlled demonstration space. The question now is whether the engineering can keep pace with the ambition, and whether the machines can prove themselves reliable enough to justify the investment and the hype.
Citações Notáveis
Don't limit yourself to what people can do, but actually go beyond.— Robert Playter, Boston Dynamics CEO
Building reliable machines takes time. These robots have to be reliable. They have to be affordable. That will take time to deploy.— Robert Playter, Boston Dynamics CEO
A Conversa do Hearth Outra perspectiva sobre a história
When you watch Atlas rotate its torso 180 degrees without moving its feet, what's actually different from the old design that made this possible?
They removed the wires. The earlier versions had cables running across the joints—necessary for control, but they limited rotation and constantly broke. The new design routes everything differently, so there's nothing physically stopping the rotation anymore.
So it's not smarter, it's just... unwired?
Exactly. The intelligence is in the AI brain, which is separate. The physical freedom came from solving an engineering problem that had nothing to do with learning.
Tell me about the hands. Three fingers seems limiting.
They're not fixed fingers. Each one can pivot into different positions—so you get different grip shapes from the same hand. It's versatile in a way that mimics what human hands do, but the mechanism is completely different.
How does it learn to use those hands?
A person puts on VR goggles and controls the robot remotely, showing it how to pick up a cup or tie a knot. The robot watches, learns the pattern, and eventually does it on its own. It's learning from human demonstration, not from being programmed line by line.
But the CEO said there's still a lot of hype ahead of reality. What's missing?
Reliability at scale, mostly. And cost. You can build one incredible robot. Building thousands that work reliably in factories, that don't break down, that are affordable—that's a different problem entirely. The software moves fast. The machines move slow.