Blue Jay Way

This story first appeared in A3's Automated newsletter. Subscribe here. 

Development was a big piece of last week’s Blue Jay unveil. The move was something of a shift for Amazon Robotics, which has tended to spend the majority of a product reveal discussing the technology itself. Here, however, chief technologist Tye Brady devoted a chunk of his stage time not just to how the industrial system was created — but how quickly the team was able to do so. 

Specifically, the relatively compact, six-arm system was iterated on in “just over a year.” Quoting directly from Amazon’s copy here, 

Blue Jay’s development moved from concept to production in just over a year — a process that formerly took three or more years for earlier Amazon systems like Robin, Cardinal, or Sparrow. The reason: Years of trial-and-error were condensed into months of development thanks to advancements in AI. 

It’s an interesting note, to be sure. Anecdotally, I do feel as though robotics firms these days are leading with speed of development as a key selling point. This is particularly the case among emerging humanoid firms — a technology (along with AGI) — many skeptical parties insist is much further from mainstream than Silicon Valley would have you believe.  

The real reason it jumped out at me, however, is how well the “just over a year” bit matched up with Amazon essentially hiring up a sizable chunk of Covariant’s staff at the tail end of last August. The well-funded physical AI firm was founded by U.C. Berkeley roboticists Pieter Abbeel, Peter Chen, and Rocky Duan. Their work largely centered around the development of robotic language models, which, incidentally, Amazon also licensed as part of that August 2024 deal. 

Speaking with a few folks at the event last week confirmed my suspicion that Covariant’s work played a key role in this accelerated roll out.  

“We’ve been working with the folks at Covariant on how to orchestrate our robotic work cells,” Amazon Robotics’ director of applied science Aaron Parness. “How to use their models, but also their model infrastructure to go faster in going from a prototype to a product or going from a concept to something we have in the field. This is something Blue Jay is a great example of.” 

We need to be precise when discussing scale, of course. We know that Amazon has deployed north of one million robots across its fulfillment centers, the vast majority of which are AMRs. Robin arms, meanwhile, currently number around 3-4,000. Blue Jay currently numbers in the one. There was a non-operational piece of the robot at last week’s Milpitas event, but the real thing is currently being put to work at a South Carolina fulfilment center, as the company works to expand sub-Same Day delivery in the region.  

Unlike some startups afforded long runways by VC backers, Amazon Robotics’ work lives and dies by how well it can plug into the existing matrix — and, eventually, scale. As mentioned in recent issues, we now know the company was exploring the humanoid form factor at least as far back as 2018, when Agility was first emerging on the scene. The company determined the form factor wouldn’t make its processes more efficient and scrapped the idea for the moment. 

Amazon publicly announced Agility pilots in more recent years, though the status of that project remains unclear. Either way, there’s nothing stopping the company from revisiting a technology after it’s had sufficient time to mature. Billions upon billions are currently being waged on whether that much can be said for humanoids. 

For now, we’re left to disentangle Morovec’s paradox, which is to say, if a humanoid can do a running backflip off a wall, surely moving boxes around a factory shouldn’t be a problem, right? Which brings us back to another bit of Covariant business. The weeks leading up to the Blue Jay announcement also gave us the first signs of life of Amazon FAR (Frontier AI & Robotics), a research wing inside the company that also emerged as a result of the Covariant deal. 

The fundamentals of that work are similar to those that accelerated Blue Jay’s development: simulation, synthetic data, and, of course, the use of robotic foundational models. Here, however, humanoid robots were tasked with executing a complex series of tasks in a variety of different indoor conditions. Cool stuff. Also, wall flips.