Alberto Rodriguez from Boston Dynamics joins us to explain what makes the new Atlas humanoid robot such a major leap forward … and why the biggest challenge in robotics is no longer just hardware.
In this conversation, Alberto breaks down Boston Dynamics’ approach to “physical intelligence” versus “reasoning intelligence,” how Atlas has been redesigned with dramatically fewer parts, and why simpler hardware is critical for reliability, lower costs, and eventually mass manufacturing.
We also dive into:
- Why AI is now the primary bottleneck for humanoid robots
- How Atlas learns and adapts to new factory tasks
- Physical AI vs. reasoning AI
- Foundation models and robot training
- Factory integration and fleet management
- Why Boston Dynamics believes legs still beat wheels for many real-world applications
- The path toward scalable humanoid robots in manufacturing
Transcript: Boston Dynamics … 1 on 1 with Atlas director of robot behavior
Alberto Rodriguez:
If you look at this Atlas compared to the previous generation, there’s almost an order of magnitude reduction in complexity. It has way, way fewer parts and way fewer unique parts. The manufacturing process is much faster and simpler, which leads to higher reliability and lower cost.
John Koetsier:
So many things are happening right now in humanoids in general, and at Boston Dynamics as well. Hyundai obviously took full control. You’ve released significant capabilities. Arguably, your robot is the most capable humanoid on the planet right now.
And I say arguably because it’s really hard to tell. Specs are not all out there. Demo videos are great, but there are some really impressive things you’re doing. It’s probably expensive still right now, but tell me the state of the union—where you guys are.
Alberto Rodriguez:
Yeah. That’s definitely along the lines of how we feel. We’ve really brought out a new generation of Atlas that puts us at the forefront of capability.
It’s interesting when we mention capability because it’s really the combination of hardware and software.
When you bring out a new robot, you have in your head the capability and potential that you’ve designed it for, but it’s not really until you figure out how to get it to do those things that you verify it. We’re really happy that we’ve been able to do that quite fast after designing this new robot.
John Koetsier:
It is amazing, right? You want the hardware to be able to do the sort of tasks that you want the robot to be able to do, but without the AI, without the software—the physical AI behind it—it’s kind of useless.
I’m seeing some robot manufacturers over-engineer, in some sense, if that’s possible in humanoids, the hardware they’re initially shipping. Then they’re going to continually build the AI and software updates, like modern products.
Mm-hmm.
Alberto Rodriguez:
Yeah, absolutely. In many respects, I would say that the things we can show with existing hardware are still limited by our ability to know how to control it.
I think that Atlas, and I’m sure it’s the same for many other robots, is capable of much more than we’ve been able to squeeze out of it today.
So I would say that the AI capabilities and the control algorithms are still one of the main bottlenecks in getting value out of the hardware.
Maybe just to highlight one thing on the hardware side that we’re very excited about: in the design of this new generation of Atlas, we’ve been very explicit about focusing on simplicity and reliability.
If you look at this Atlas compared to the previous generation, there’s almost an order of magnitude reduction in complexity. It has way, way fewer parts and way fewer unique parts. The manufacturing process is much faster and simpler, which leads to higher reliability and lower cost.
So we’re happy that we’ve been able to demonstrate the same level of performance—or higher—with a robot that is fundamentally way, way simpler, which we think puts us in a really good place for the next step of mass manufacturing.
John Koetsier:
Order of magnitude—that’s a big statement.
I know you said “almost,” so you’re not saying it’s a full order of magnitude, but wow. That’s super impressive, and that’s obviously a critical step toward manufacturability, lower cost, and being able to price it appropriately.
Talk about the AI side a little bit and the physical AI side.
It’s a good time to talk about it. Apptronik just released news this morning. I literally just published a story on Robot Park. It’s a 90,000-square-foot facility. They’re doing several around the world as well, and it’s a data factory for robotics. You’re going to have robots doing many, many tasks in many different environments and learning from that.
They obviously have a partnership with Google DeepMind as well. Talk about your AI and your foundation models. How are you training Atlas?
Alberto Rodriguez:
I like thinking of the AI layer that controls the robot as a two-system approach.
On one hand, you have what I call physical intelligence.
It’s the core component of the control algorithm that worries about things like balance, physical skill, and agility—jumping, grabbing things, and moving them with speed and generality.
Then there’s the reasoning intelligence. It’s the part that looks around and understands, “Okay, to do this task, first I have to do step one, then step two, then step three.”
It also understands things like, “This object looks heavy,” or “This one looks lightweight,” or “I’m going to have to push this way.”
So we’re investing on both sides.
I think we’re uniquely positioned to have a core strength in the physical intelligence part of the behavior stack.
For many years, we’ve demonstrated how we can push forward the frontier of what these robots can do.
In the past, we’ve been inspired by areas of human agile skill. It could be dancing, parkour, or gymnastics.
Now we’ve done it in the context of soccer, using the World Cup as motivation, which has always forced us to have a very strong grasp on how to squeeze physical strength and agility out of the robots we build.
We think that’s clearly a competitive strength we have.
Over the last couple of years, we’ve been investing more in the second kind of intelligence—the reasoning intelligence—because it’s also essential if you want these platforms to have some degree of ease of retasking or generality.
If you want to deploy a robot in a factory and it has to do a certain job, but two weeks later that job changes because the workflow changes or there’s a new exception it needs to handle, you want to make sure the robot can adapt without having to invest another few months of effort validating and creating new programming.
You want the robot to learn through either experience or demonstration in a way that’s much more natural.
So we’ve been heavily investing in that second kind of intelligence over the last couple of years.
John Koetsier:
That’s super interesting because that’s the human corollary, right?
You get a new job, and somebody—your supervisor or a coworker—says, “This is what you need to do. This is how you do it.”
If your robot can learn like that on the fly and not spend two months—or even two weeks—learning the new job, that’s generally a good thing.
Are you also working on a layer of intelligence for interoperability?
I was just at Humanoid.ai’s offices in London, UK, and they’ve got three or four levels of intelligence. One of them is an agentic overseer. Others, like tutor intelligence, connect into a warehouse management system and receive tasks from above.
Then there are even layers where you think about robots working together on a task. Nobody explicitly told them to, but one robot recognizes that something is heavy and needs help.
What about those layers of intelligence?
Alberto Rodriguez:
Yeah, definitely.
There are certain things that we think are essential.
For example, being able to understand and communicate with factory manufacturing systems.
A big cost today when you’re deploying new technology in a factory has to do with integration.
You have to get your system to speak the factory’s language and communicate with the management system that’s orchestrating everything.
Getting these robots to speak that language and either receive commands from a higher-level system or provide feedback is important.
That’s actually an area where we have significant experience.
From our other products that are deployed with hundreds of customers—Spot, for example—we have a fleet management system where you don’t program each individual Spot robot. Instead, you distribute a set of inspection points throughout a factory.
Then there’s a coordination and fleet management system that decides, “Should I send this robot or that robot?” and distributes the workload accordingly.
That’s something we’ve been doing for quite a while with some of our other products.
John Koetsier:
Mm-hmm.
Do you have any thoughts around legs versus wheels?
Jeff from Apptronik, when I talked to him yesterday, said, “We see wheels as doing most of the job, but legs have a higher ceiling.”
Thoughts on that?
Alberto Rodriguez:
Yeah, lots of thoughts, of course, and many internal discussions.
We’ve actually tried both.
One of our products, Stretch, is a big wheeled base with a giant arm. Atlas and Spot are legged.
With Atlas, one of the main drivers for investing in legs is that the mechanical complexity involved in building legs today is actually not that different from building an omnidirectional mobile platform.
You usually end up with four wheels, and each wheel typically has two actuators because you want it to drive and steer.
So you end up with eight actuators, which is almost the same as you have on two legs.
The second thing is that legs allow you to reach more places.
Many factories may look flat at first inspection, but then you realize you also want the machine to cross the gap between the loading dock and a trailer.
Or there may be a mezzanine where you have to go upstairs, or other situations where legs give you the extra reach that becomes necessary.
The other important thing about legs is that they allow for a slimmer form factor.
You can squeeze into narrow spaces, which are very common in factories and warehouses because space is at a premium.
For example, if you have a mobile base and you don’t know where your arm is going to reach, you have to make the base large in every direction so it can remain stable no matter where the arm moves.
With two legs, you can choose which direction the legs face and remain much thinner in the other direction, allowing you to squeeze between tight spaces.
So there’s an accumulation of reasons why legs actually end up making sense.
Maybe the last thing is that balance and locomotion used to be a difficult problem, but it’s actually not that difficult anymore.
We’ve figured out the right recipes for how to do it reliably.
John Koetsier:
There are some humanoid startups that are still working on that problem, of course.
But you’ve been doing it for a while, so that makes a ton of sense.
Well, great to meet you, and have a wonderful day.
I might have a few more things, and we’ll chat about that later.