UK-based grocery retailer Ocado is ahead of the game mainly because it has no brick-and-mortar outlets or legacy tech. At its Customer Fulfillment Centres (CFCs), robots currently pick and pack the approximately 40,000 orders Ocado handles each day in the UK via a unique AI-powered grid system.
During a recent visit to the UK, PCMag went out to Ocado’s HQ in Hatfield, 30 minutes north of London, to get a tour of what’s known as CFC1—including a sneak peek of its “robotic playpen” where robots “learn” how to pick and pack.
To give you a visual: the 65-foot CFC1 is equivalent to 11 soccer pitches, with 21 miles of track. Since 2001, Ocado has built three more CFCs, all of increasing complexity. Its latest two, CFC3 and CFC4, were built on the Ocado Smart Platform. It’s also powering other global retailers around the world and is now on our shores, in partnership with Kroger.
Ocado did not reveal how many robots it currently has buzzing around its CFCs, but when fully ramped, CFC3 in Andover will support 1,100 and CFC4 in Erith will have 3,500.
In the spirit of full disclosure, Ocado’s CTO, Paul Clarke, was my boss in 1997. So it was deeply cool (for me) to return to the kind of tech geek-outs we used to have back in the day. Here are edited and condensed excerpts from our conversation.
Paul, what made you decide to join Ocado?
When I first got the call, they said the vertical was retail, and I said I wasn’t interested. But then they clarified it by saying: “It’s not normal retail, it’s 100 percent online, come and spend a couple of hours with us and see.” I came up here, was blown away, ended up spending the whole day, signed on for a year’s consultancy, went on staff, became CTO, and am still here, 13 years later. Essentially, I love the intersection of hardware and software that controls things that move. It’s so much more interesting than desktop applications. At least, that’s what floats my boat, so to speak.
As I walked around CFC1 with you this morning, I could see this is definitely not ‘normal retail.’
It’s funny, the outside world just likes the fact that it’s very easy and convenient to shop from Ocado, as well as our very long tail of 50,000 available items, from many different suppliers. They don’t know how the magic happens under the hood and they don’t need to.
So how does Ocado’s tech work?
Picture a huge chess board. Under every chess square is a stack of storage bins, containing groceries, which can be up to 21 bins deep. Now picture robots living on top of this vast 2D dimensional grid-world. The robots stop on a chess square, grab the bin from the top of the stack and bring it up into the body of the robot system.
Then the robot has a set of choices: drop the bin off, onto another stack, or bring it to a machine, around the edge of the grid, to a pick station—where a human or soon another kind of robot picks items into a customer order—or a decant station—where goods coming in from suppliers are deposited into bins and then put away in the hive. We can pick a 50-item order in five minutes because the robots are collaborating with each other, orchestrated by a machine-learning air traffic control-like system.
You certainly couldn’t compile a weekly grocery order in five minutes by picking in a regular store.
But these are not autonomous robots, right? You mentioned the word ‘orchestrated’?
Our robots collaborate with each other, so they are orchestrated. They come at each other at a closing speed of 8 meters per second [18mph], with less than the width of your finger gap between them. Going back to the chess board analogy, the control system—aka chess master—is playing chess many moves ahead. It knows which bins needs to be where, and which robots it needs to collaborate with in order to complete the order. Just as in air traffic control: each robot is looking for permission/clearance to move from those ahead, and behind, so they don’t hit the other robots. We have a “fail safe” where if the clearance isn’t given, the robot comes to a graceful stop. You couldn’t achieve this density and efficiency if the robots were autonomous.
Which is why you rely heavily on simulation to plot it all out.
Yes, simulation is a fundamental part of what we do here. In fact, that was the first team I built when I arrived. You can’t buy a simulation system that will cope with the number of moving elements that we have. Most systems come from automotive, where just hundreds of things are moving, but we have thousands.
We used gaming technology for visualization, originally building the simulation system in Java, and now we also use VR to allow us to be immersed inside the “hive.” Also, like AlphaGo Zero, which learned how to master the game “Go” by playing itself, we are embarking on connecting machine learning to our simulations in order to come up with new ways to optimize our systems.
That is the powerful thing about simulations as a sister technology to machine learning. I believe these systems will help us evolve as a species, beyond what we already know, think or can dream up.
But groceries are very seasonal. Do you use machine learning to handle that?
Yes, because 80 percent of sales comes from 20 percent of goods so each of our 50,000 products has a velocity—not how fast is it moving on a grid, but how many are we selling per hour. We use machine learning to create a system that moves the products around the warehouse based on each item’s velocity, and then constantly tweaks it based on what the ML knows about seasonal items, celebrity chef programs, promotions, social media trends, and so on.
The opposite of warehousing, this is just-in-time grocery handling.
Exactly, and the system is constantly adjusting. It’s massive. It’s one of the reasons why we keep bumping up against Moore’s Law.
Does that mean you’re going quantum?
I can’t say too much about it at this time but yes, we’re playing with things like quantum computing because the hive lends itself to a quantum expression. We need to be ahead of the curve.
Are your robots self-diagnostic?
That’s another level of our ML we are working on. We’re allowing them to self-test and do exception handling. For example, if the robot finds a fault, it will make a decision about whether it’s fit to join the swarm again, or take itself to the “pits” to get fixed.
Are your robots built in Shenzhen like everyone else’s?
No. Our Ocado-branded robots are proudly stamped: “Made in Britain.” We own the IP and we are designing and manufacturing them. If we aren’t already, we soon will be, the largest robotic manufacturer in Europe. And these aren’t small robots, they’re the size of domestic washing machines, upturned on their side, and a complex beast.
How do you handle comms? Wi-Fi?
No, there would be too much latency with that and it wouldn’t handle the density. We utilize the mobile technology LTE in the unlicensed spectrum. Each robot has its own cell in a time/ frequency matrix. That’s crucial because it’s a time-critical control system. Our grid is now the densest mobile system in the world that we know of. It’s built for serious scale—we can now “talk” to thousands of robots from a single base station 10 times a second without contention.
How much data do your robots generate?
About 5,000 data points, 1,000 times x second. Which means each robot generates a gig of data a day or a total of 4 terabytes a day for an entire swam within one CFC.
Wow. Can you talk about your next-gen robots?
We are partners in two Horizon 2020 projects with universities across Europe. One is the Second Hand project, where we are building a humanoid maintenance robot that we’re gently torturing with DARPA-style challenges. The other is around SoMa (soft manipulation), mimicking the capabilities of the human hand to do robotic picking of different items and put them into bags. This is nothing like other industrial tasks such as putting a windscreen into a car, or spray painting it. It’s all about making smart decisions on the fly, grippers, machine learning and vision systems to handle different form factors, crushability, 3D orientation, food segregation rules, grasping strategies, competencies. It’s a difficulty curve. We started with a simple suction cup and are moving up to dexterous manipulation.
I saw in the “robotic playpen” inside CFC1 you also had robots “figuring out” how to pick.
Yes. We cage the robot and give it access to our full range of 50,000 items to see how it does, while we film it, monitor it, and see how it learns. Essentially that’s a living lab for robotics where we combine reinforcement learning, teaching, learn by doing, and so on. We need robots that have acquired a set of strategies.
Now you’ve proved this massive hardware/software/robots/grid system works, you’re taking this platform global by powering your competitors.
Yes, at Ocado Technology, we’re now putting some of the world’s largest grocery retailers online using the cloud, robotics, AI, and IoT with our Ocado Smart Platform. We’ve now become a managed services business—SaaS—but with hardware/RaaS (robotics as a service) bundled in. The grocery partner tells us the dimensions of a brownfield site/warehouse. We simulate what it would take to run their online grocery business from there. Then we provide everything from robotics to supply chain forecasting, real-time warehouse control systems, telemetry for the delivery van fleets, simulation, and big data. We’ve signed deals with companies in the UK, Spain, France, Canada, Sweden—and now Kroger in the US. Our stock price doubled overnight when we announced the partnership with Kroger.
I wouldn’t be doing my job if I didn’t ask about Amazon.
Of course. Okay. Here’s the thing: Amazon has obviously moved into grocery and so could be seen as a competitor. But what that has done is create a sense of urgency for retailers to move online fast and, essentially, “up their game.” But, guess what? The only company who can give you the solution out of the box as a one-stop-shop is us. Groceries are a hard business to make money in, at scale. The average item price is 2 Pounds Sterling ($2.63). On that you’ll make a 60p (79 cents) gross margin. Therefore the only way to do this business, at scale, is with a high degree of automation. Essentially, what we’re doing today, at Ocado, is massively parallel edge computing in order to achieve this.
This sounds like it’s all about online grocery, so where do the other atoms come in?
The funny thing is our underlying technologies know almost nothing about grocery, they don’t need to. So they have all sorts of spin-off applications in other sectors. Having recently re-organized Ocado Technology, I have more time for exploring those spin-off opportunities, some of which are very far from the online grocery tree. You ain’t seen nothing yet.
Good answer. Okay, final question: what’s your favourite item in the Ocado store?
I love dips; our Tzatziki is exceedingly good.