Amid Amazon’s Robot Surge, Proteus Charts a New Path Forward


The robots glide across the floor, sometimes pausing to spin a quarter turn or two before resuming their route. They come close to one another but never collide. It’s not choreographed – they’re adapting on the fly – but the movement does have the feel of a ballet.

If ballet dancers were mechanized platforms on wheels, that is. Flat-topped and low to the ground, like oversized bathroom scales granted the gift of movement and the ability to navigate on their own.

These are Amazon’s Proteus robots in action. 

In a spacious Amazon warehouse in London, as in its counterparts around the world, Proteus, Titan and fellow robots are perpetually tasked with fetch quests – finding and retrieving shelving units that contain items that all of us order day in and day out and bringing them to stations where those items are picked, packed and sent on their way.

Some of those days are busier than others – Prime Day sales, for example, when Amazon orders surge. During these periods, fulfillment centers bring on thousands more workers and the robots keep pace.

We visited two Amazon locations – the LCY3 London fulfillment center and the BOS27 robot development facility in Westborough, Massachusetts – to better understand the role robots play in ensuring our packages reach us at speed, both now and in the future.

After decades of humanity’s sci-fi-inspired preoccupation with robots, advances in AI (including large language models and vision language models) over the past five years are increasingly allowing robots to interact with people in more natural ways. For the most part, these real-world robots bear little resemblance to the pop culture depictions, particularly of the humanoid variety. Humanoids are starting to spring up, but most robots around us today are much closer to the type Amazon and other companies are using in industrial settings.

A short yellow and black robot, flat but wide, shining a green light on the floor in front of it

Proteus version two — coming soon to a fulfillment center near you.

Katie Collins/CNET

In Amazon facilities, the robots range from Proteus, which could be a Roomba’s more strapping younger sibling, to Vulcan, a robotic arm with a sense of touch that can pick up objects and understand what it’s handling. Altogether, Amazon has over 1 million robots operating in fulfillment centers, handling tasks such as stowing, picking, sorting and transporting. 

Even though Amazon has been developing robots for years, it’s still only in the early stages of growing its robotics portfolio, said Tye Brady, Amazon’s chief technologist, speaking in London in early June.

What it’s learned so far is that robots make the environment safer, and therefore more efficient. In centers where robots have been deployed, Amazon has seen a 41% reduction in the number of accidents and a 40% increase in the amount of goods delivered. 

“The efficiencies allow us to pass on a low cost to our customers,” said Brady. “The robotic systems allow us to store more goods physically closer to our customers as well.”

Over time, Brady added, the gradual introduction of robots is creating a powerful cycle within Amazon. “We deploy systems, we learn from them, we improve them and then we expand on what they can do for people,” he said.

That’s exactly what it’s done with Proteus, with a new version ready and raring to replace the existing model in fulfillment centers across the globe in the next few years.

Freewheeling Proteus robot gets language skills

Proteus is Amazon’s first fully autonomous robot – a “collaborative robot” designed to work and move around in the same spaces as humans going about their normal activities, not cordoned off behind fences with tightly restricted access for employees. It’s loaded with sensing and navigation capabilities.

“You just put them where the people are, or put the people where they are, and they’ll get right around you,” said Travis Hearn, a QA engineer at Amazon’s BOS27 facility, located 30 miles west of Boston along a once rural road now lined with low-rise industrial and commercial buildings. Cyclone fencing divvies up sectors of a cavernous space, where a diverse array of mobility and manipulation robots go through their paces.

The more diminutive demo area for Proteus, by contrast, is wide open, simulating the fulfillment center terrain it’s built to traverse, potentially several hundred meters from where chutes drop customer packages to where those packages get placed into delivery vehicles.

Tall yellow racks stand side by side. Under one of them is a low-profile blue robot

In a London fulfillment center, an Amazon mobility robot has slid under a rack that it’ll lift and tote across the floor.

Katie Collins/CNET

A Proteus robot – 7.8 inches tall, 31.5 inches long and 29.9 inches wide – can carry up to almost 900 pounds. That’s modest compared to what the larger, lookalike Hercules and Titan mobile robots can carry (1,250 and  2,500 pounds, respectively). Racks holding the goods for delivery get stacked on top, creating tall rectangles that scoot from one station to another.

But Proteus can be much more freewheeling than its fellow bots. It doesn’t need markers on the floor to know where it is or what route to follow. It learns its environment over time. It also recognizes when something – or someone – unexpected is in the way.

“You could think of it like an invisible force field, a bubble around the vehicle. So if somebody stepped in the way of the vehicle, then it would come to a safe stop or slow down,” Scott Dresser, Amazon’s vice president of robotics, said in an interview this week at BOS27. “The intelligence is to find and detect people and safely avoid them.”

The first-generation Proteus has been around for several years, and Amazon has a little over 4,000 of them at 25 sites. Earlier this month, the company introduced the Proteus 2, which gains natural language processing so that people will be able to direct it with voice prompts.

“What makes this possible is a new AI architecture that allows employees to interact with Proteus through natural language using advancements in our generative and agentic AI systems,” said Brady. 

Amazon employees will be able to talk to the robot the same way they do their colleagues, including gesturing – with a casual, “Hey Proteus, could you take this to the corner of the building?” It will be able to figure out route planning and timing and then execute the task on its own.

The second-generation Proteus will be rolled out to Amazon facilities in the coming months.

Robots doing fulfillment work for Amazon orders

The new Proteus will be deployed at LCY3 in the first half of 2027. Meanwhile, Amazon robots are already an essential part of the furniture.

Situated in Dartford, right at London’s eastern-most point, LCY3 is a strategically located fulfillment center on the banks of the River Thames, serving the British capital and beyond. Here, Prime Day orders are picked, packed and shipped across the UK and Europe.

Last year Amazon invested $60 billion across Europe to grow its operations on the continent, and it has ambitious goals for improving delivery times. It’s growing Amazon Now ultra-fast delivery to 20-plus sites in the UK, and it’s accelerating same-day delivery by adding more than 25 sites across Europe this year.

“When we make delivery faster, we are not just moving boxes quicker,” said Mariangela Marseglia, vice president of Amazon European Stores, speaking at the London event. “We are giving people minutes, hours back.” 

Faster delivery, she added, comes from working safer and smarter. This is where the robots come in.

Two robot components on display

Amazon is experimenting with different robotic systems for different tasks.

Katie Collins/CNET

To hit its delivery goals in Europe, Amazon is investing more than $10 billion to expand and modernize its fulfillment network with robotics across the continent over the next few years. Some of the robotics systems it’s putting in place have been built on suggestions made by Amazon employees, said Armin Cossman, the company’s vice president of operations for Europe.

A new system called Stark, for example, was the idea of an Amazon operations employees in Spain. It picks up huge crates from conveyer belts and places them onto trolleys – repetitive work that puts an enormous amount of strain on the human body. Stark is being piloted in Barcelona, but Amazon plans to bring it to at least 15 more sites across Europe by the end of 2027.

It’s the first successful deployment of collaborative robots in Amazon’s fulfilment network, said Cossman. “Employees work side by side with collaborative technology – the same space working together on the same process.”

On both of our visits to its sites, Amazon was careful to impress upon us that this human-robot collaboration is a key part of its robotics strategy. The company, which has been repeatedly accused of unsafe work conditions in its warehouses and of looking to replace workers with machines, wanted us to know, and you to know, that its robots aren’t here to take its workers’ jobs – just to make them better.

Two low-profile robots, one under a blue rack stack, on a sprawling warehouse floor.

Amazon’s Proteus robots can navigate safely around other robots and humans.

Katie Collins/CNET

“When people have a people versus machines mentality, I find that wrong,” said Brady. “I believe that people, when they have technologies as a tool set, that there’s nothing in this world that they can achieve.”

Amazon has upskilled 700,000 workers, he added, with many more to come. He also anticipates the creation of new jobs linked to robotics as Amazon’s portfolio evolves.

“Robots create jobs. Full stop. It’s a fact,” said Paul Miller, vice president and principal analyst at market researcher Forrester. “New jobs are created to maintain the robots, to manage the robots and to do the new work that’s made possible because automation has lowered the cost, improved the consistency or accelerated the delivery of the tasks people once performed.”

Still, some individuals will be adversely affected by the disruption, Miller added. Those people will need to be supported as they change careers to ensure they’re better off.

At LCY3, there were many workers stationed across the 2 million square feet of operating space, spread out across airy halls with natural light flooding in from the Thames-view windows. Many were packing deliveries or unpacking returns, and some were working with and on the robots.

One key role is that of amnesty responder, whose responsibility it is to rescue items that have fallen from the pods the Proteus robots whisk around. Fallen items are the main point of failure in the Proteus system. When something tumbles out of one of the shelving units, the amnesty responder hits a button and the entire ballet pauses to allow the human in the loop to retrieve the offending object. Only once they’ve exited the arena does the dance continue.

On rare occasions, Amazon acknowledged, a collision occurs. Usually this will result in the Proteus needing a new camera lens, courtesy of the mechanic that’s always on hand. Then it’s back to work.

What next for Amazon robotics

Amazon’s robotics capabilities are evolving fast.

Beyond the walls of its fulfillment centers are delivery robots, such as the Amazon Scout and Amazon Prime Air drone. The latter is already live at eight sites across the US. Meanwhile, the company is testing the service in Darlington in the UK.

The MK30 drone can deliver shoebox-sized packages, allowing Amazon to deliver from a range of 60,000 items within a two-hour window. With its six propellors, a redundancy that allows the drone to continue on even if one fails, it will hover above the ground and drop packages without damaging them (it can detect obstacles on the ground).

A winged drone with six propellers

Amazon Prime Air is another of the company’s robotics projects.

Katie Collins/CNET

Meanwhile, today’s robots are the preliminaries for what comes next. No, not humanoids, like in Elon Musk’s fever dreams of swarms of Optimus robots doing factory jobs.

Amazon has more modest expectations, targeting somewhere between what it’s doing with robots today and what humanoids may eventually deliver. That could include merging the capabilities of its mobility (e.g. Proteus) and manipulation (e.g. Sparrow) robots. Dresser said Amazon sees paths to using some combination of those technologies.

“How can we move and manipulate in the same robot, and what does that look like? Because we think that that is where our operations are heading,” Dresser said. “I think we’re going to see some new, interesting form factors in the coming months that are going to be in our warehouses very quickly.”

It’s clearly a company learning in real time – designing robots to meet its specific needs, and then refining them based on how they perform when thrust into real-world situations. “The systems we’re building today,” said Brady, “are laying the foundation for what comes next.”





Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews


Explain CAP

CAP theorem is also called Brewer’s theorem, which stands for Consistency, Availability, and Partition Tolerance.

Consistency: 

This situation expresses, all nodes have similar information simultaneously. Implementing a read function will return the estimation of the latest write function making all nodes provide similar information. A framework has consistency if an exchange begins with the framework in a reliable state, and finishes with the framework in a predictable state. A framework can (and does) move into a conflicting state during an exchange, however the whole transaction gets moved back if there is a mistake during any process all the while. We have 2 unique records (“Bulbasaur” and “Pikachu”) at various timestamps given in the picture below. The result on the third part is “Pikachu”, the most recent input. The nodes will require time to refresh and won’t be available on the organization as frequently.

Consistency

Availability:

This situation provides that each solicitation gets a reaction on success/failure. Accomplishing availability in an appropriated framework necessitates that the framework stays operational 100% of the time. Each customer gets a reaction, paying little heed to the condition of any individual node in the framework. This measurement is trifling to quantify: possibly you can submit the read/write commands, or you can’t. Thus, the databases are time autonomous as they should be accessible online consistently. In contrast to the past model, we couldn’t say whether “Pikachu” or “Bulbasaur” was included at first. The result could be any one among both. Consequently, high accessibility isn’t feasible when dissecting streaming information at high frequency.

Availability

Partition Tolerance: 

This situation expresses that the framework keeps on operating, in spite of the quantity of messages being deferred by the organization among nodes. A framework which is partition tolerant can support any measure of organization failure which does not bring about a failure of the whole network. Information records are adequately duplicated across blends of nodes and organizations to maintain the framework up through discontinuous blackouts. While managing current distributed frameworks, Partition Tolerance is a requirement and not a choice. Thus, we need to exchange among Consistency and Availability.

Partition Tolerance

Enroll in our Apache Storm Training program today and elevate your skills!

Big Data Hadoop Training

  • Master Your Craft
  • Lifetime LMS & Faculty Access
  • 24/7 online expert support
  • Real-world & Project Based Learning

Distributed Database Systems 

In a NoSQL type dispersed data set framework, Different PCs, or nodes, cooperate to give an impression of a unique operating database unit to the client in a NoSQL type distributed database system. They store the information among these numerous nodes. Every one of these nodes operates an event of the database server and they converse with one another. At the point when a client needs to write to the database, the information is suitably kept in touch with a node in the disseminated data set. The client may not know about where the information is composed.

Essentially, when a client needs to recover the information, it interfaces with the closest node in the framework that recovers the information for it, without the client thinking about this. Along these lines, a client essentially communicates with the framework as though it is connecting with a solitary information base. These nodes recover information that the client is searching for, from the important node, or putting away the information given by the client. 

The advantages of a distributed system are very self-evident. The expansion in rush hour gridlock from the clients, we can undoubtedly scale our information base by including more nodes to the framework. As these nodes are commodity equipment, they are moderately less expensive than adding more assets to every one of the nodes independently. Horizontal scaling is less expensive than vertical scaling. The horizontal scaling assures that the replication of information is less expensive and simpler. It implies that now the framework can undoubtedly deal with more client traffic by fittingly appropriating the traffic among the recreated nodes.

HKR Trainings Logo

Subscribe to our YouTube channel to get new updates..!

What is the CAP Theorem?

The CAP theorem states that a distributed database system has to make a tradeoff between Consistency and Availability when a Partition occurs.

A distributed database framework will undoubtedly have partitions in a certifiable framework because of network failure or some other explanation. Along these lines, partition tolerance is a property we can’t dodge while setting up the framework. A distributed framework will either decide to abandon Consistency or Availability however not on Partition tolerance. For instance, if a partition happens among two nodes, it is difficult to give steady information on both the nodes and accessibility of complete information. Consequently, in such a situation we either decide to settle on Consistency or on Availability. A NoSQL circulated database is either portrayed as  AP or CP. CA type information bases are for the most part the solid databases which operate on a solitary node and give no conveyance. Subsequently, they need no partition tolerance.

Where can the CAP theorem be used as an example?

The CAP theorem can indeed serve as an illustrative example within the realm of distributed database systems. When setting up a distributed database framework, it is inevitable to encounter partitions due to network failures or other unforeseen circumstances. Hence, partition tolerance becomes a necessary property that cannot be avoided in such a system. In this context, the CAP theorem comes into play. It states that a distributed framework must make a trade-off between either consistency or availability, as it is not possible to achieve both simultaneously when a partition occurs between two nodes. For instance, during a partition, it becomes challenging to maintain consistent data on both nodes while ensuring complete data availability. As a consequence, in such scenarios, we are left with the choice of prioritizing either consistency or availability.

To better understand this, it is essential to consider the different types of distributed databases. NoSQL distributed databases can be characterized as either AP or CP. AP databases prioritize availability and partition tolerance over strict consistency. On the other hand, CP databases prioritize consistency and partition tolerance at the expense of availability. These distinctions become crucial when deciding the appropriate database type for specific use cases.

CAP Theorem NoSQL Database Types

NoSQL (non-relational) databases are suitable for distributed network applications. NoSQL databases are horizontally adaptable and disseminated by layout, it can quickly scale across a developing network comprising different interconnected nodes.They are characterized dependent on the two CAP attributes they uphold: 

CP database: A CP database conveys partition tolerance and consistency at the cost of accessibility. At the point when a partition happens between any two of the nodes, the framework needs to shut down the non consistent node (make it inaccessible) until the partition is settled. 

AP database: An AP database conveys partition tolerance and accessibility at the cost of consistency. At the point when a partition happens, all nodes stay accessible however those at some unacceptable end of a partition may return a more established rendition of information than others.  

CA database: A CA database conveys accessibility and consistency among all nodes. It will not be able to do this if there is a partition in between any two nodes  in the framework, in any case, and can’t convey adaptation to internal failure.

Spaces defined by CAP

CD Space: The engines of this space concentrate on accessibility and consistency, information dispersion doesn’t prevail. It is the spot where Relational Databases are placed, in spite of the fact that we can likewise discover some NoSQL engines which are diagrammatically arranged. 

ND Space: This doesn’t receive any Databases engine and is an empty set. It repudiates the CAP Theorem on the grounds that with the most recent innovation it can’t achieve with three of the Theorem features. 

DT Space: Here, the resistance of divisions and consistency are favored, leaving to the side certain degree of accessibility. Confronting a network division, these Databases couldn’t react to particular sorts of inquiries.

CT Space: Here the engines will support the accessibility and resistance of divisions, however that doesn’t mean they do not provide any consistency as it is relative and can’t ensure between nodes. 

Big Data Hadoop Training

Weekday / Weekend Batches

Conclusion

Distributed frameworks permit us to accomplish a degree of computing ability and accessibility that were essentially not accessible previously. The frameworks have better performance, lower inertness, and close to 100% up-time in servers which last till the whole globe. The frameworks are operated on product hardware which is effectively accessible and configurable at moderate expenses. Distributed frameworks are more intrinsic than their single-network partners. Learning the intricacy brought about in distributed frameworks, making the fitting compromises for the CAP, and choosing the correct apparatus for the task is essential with horizontal scaling.

 



Source link