Boston Dynamics cut Atlas complexity by 'almost an order of magnitude'

Boston Dynamics' new all-electric Atlas has been around for half a year now. But the biggest news about the fifth-generation Atlas humanoid might not have been its super-capable robot body.

Perhaps it was the bill of materials.

In an interview, Alberto Rodriguez, Director of Robot Behavior for Atlas at Boston Dynamics, told me the fifth-generation Atlas represents "almost an order of magnitude reduction in complexity" versus the outgoing platform. Fewer parts, and critically fewer unique parts, with a faster, simpler assembly process. He was deliberate about the qualifier — "almost," not a full 10x — but even that is massively significant for the manufacturing math.

Costs will be way down, which means prices can be more competitive. Here's our chat:

"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," Rodriguez said. "Which puts us in a really good place for the next step of mass manufacturing."

Why unique-part count is the metric that matters

Total part count drives assembly time. Unique part count drives supply chain complexity, inventory, tooling, and failure modes. And, obviously, more parts generally means more cost.

Collapsing all of these simultaneously turns a hand-built research platform into something you can produce at volume with high levels of quality. Boston Dynamics plans to build at least 30,000 Atlas units a year, with 2026 production fully committed to Hyundai's Robotics Metaplant and Google DeepMind. 

Hyundai Mobis supplies the actuators, which taps into the same vertical-integration playbook that lets Hyundai Motor Group ship ~7M vehicles annually. 

Historically Atlas ran north of $200K. This simplification is the clearest signal yet that the curve is going to bend down significantly from here.

The two-system control stack

Rodriguez splits Atlas's intelligence into physical intelligence — the control loop for balance, agility, and dynamic manipulation (jumping, grabbing, moving objects with speed and generality) — and reasoning intelligence, the task-decomposition and affordance-judgment layer (is this object heavy or light, how do I sequence the steps).

Physical intelligence is the company's legacy strength, from parkour to gymnastics, as we've seen for over a decade going back to the hydraulic version of Atlas. But the company's recent investment is on reasoning, because that's what enables general intelligence -- being able to do many different things -- as well as retasking.

When a production line's workflow changes two weeks in, you want quick learning, ideally from demonstration or experience, not another multi-month reprogram-and-revalidate cycle.

Software is the bottleneck, Rodriguez told me. 

"Atlas — but I'm sure it's the same for many other robots — is capable of much more than we've been able to squeeze performance out of it today," Rodriguez said. 

Fleet orchestration is a big bonus

Integration with factory/WMS systems is a major hidden deployment cost. 

But Boston Dynamics already runs fleet management across hundreds of customers for Stretch, a wheeled warehouse robot, and Spot, a quadruped guard dog. That distributes tasks and coordinates units en masse rather than programming each individually. 

Enabling that orchestration layer for Atlas too is a head start most humanoid entrants don't have.

Legs vs. wheels, from someone who ships both

My assumption has been that wheels are cheaper than legs.

But that might not actually be true.

Rodriguez's actuator math undercuts the "legs are an indulgence" assumption: a 4-wheel omnidirectional base typically needs two actuators per wheel (drive + steer) = 8, roughly parity with two legs, he told me. 

"The mechanical complexity that goes into building legs today is not that different," he said. 

Legs also add reach (dock-to-trailer gaps, upstairs to mezzanines) and a slimmer footprint. Where a mobile wheeled base needs to be wide in all directions to stay stable under variable arm loads, a legged robot can re-orient its body to fit in narrower spaces where aisles are tight. 

And the classic hard problem of legs?

"Balance and locomotion... it's actually not that difficult anymore. We've figured out the right recipes," he says. 

Of course, plenty of newer entrants are still solving those initial problems today.

The World Cup angle

As part of Hyundai's "Next Starts Now" FIFA World Cup 2026 campaign, School of Football is a five-part series of Atlas learning soccer, culminating in a "Ghost Rabona" cross-leg kick performed by Atlas.

No CGI needed. 

It's marketing, sure, but it's also a pretty sweet demo of the physical intelligence that will also do far less telegenic work on factory floors.

Bottom line

The new Atlas is even more impressive than we initially knew.

It's got match-or-beat capability, a fraction of the complexity of old Atlas, higher reliability and lower cost because of that reduction in complexity ... and it's got increasingly better reasoning AI for retasking with a proven fleet layer, all on legs. 

In a field of 400+ companies, the order-of-magnitude simplification is an important transition between a compelling demo and a manufacturable product.