THE WORDS “TROPHIC CASCADE” – a series of powerful indirect interactions that can control entire ecosystems – stream out of Alex Harvey’s mouth about halfway through our conversation. Alex is chief of advanced technologies at Ocado Technologies and trophic cascade encapsulates the quest the company’s researchers and engineers are pursuing. They’re combining an array of AI, robotics, 3D printing and ingenuity in a live experimental grocery business, set to transform not only the groceries industry but ecommerce in general.

Alex joined Ocado Technology in 2010 after nearly a decade in the aerospace industry. These days you can find him and his team fine-tuning the connection between automation and intelligent systems using any and pretty much every technology available, from advanced mechatronics to machine learning. Working on DeepPack, an AI packing and load planning system, I sometimes think back to when Ocado Technology came to my university to do a demo, showing how a warehouse can be a whirlwind of activity with robots passing and bagging goods. So when the opportunity to chat with Alex about packing, AI and robots came up, I couldn’t resist.

Andy: You mentioned you’re moving towards building robots in-house. How important is it to have control over hardware and software?

Alex: For the robots we’re deploying today, at scale, the robot arm itself is an off-the-shelf UL10 robot arm from a company called Universal Robots. We buy cameras and sensors and we package it and integrate them with our own bespoke end effector, the device at the end of the arm, which interacts with the environment.

Our plan is to deploy robot arms that we 3D print and build internally by the middle of next year, leveraging the intellectual property of a company that we acquired called Haddington Dynamics. The real reason we need control and ownership over the hardware is cost. I could buy an amazing end effector off the shelf that has lots of capabilities. But it’ll be at a price point that makes robotic picking for e-commerce and logistics uneconomical.

A human picking station has a certain price point, for argument’s sake let’s say £100,000. To man a station you have to pay someone to work shifts throughout the year. Actually, a human is an incredible performer. Once you have this initial spend to build the station and you accept the fee you’re paying the person each year, you know with 100% confidence that person can achieve everything you need in the world of picking and packing.

This sets a price envelope for the robots, which is incredibly tough. That’s why I’m so sensitive to the cost of the UL10 robot arms at about £30,000 retail because it’s a significant chunk of my overall budget for the entire robot cell.

“Lots of people are doing robotic item handling – but the real challenge is access to the data and to the real world problem. This isn’t something you can create in a lab…the real world has so many more variables.”

You mentioned matching human performance. When you first come into this kind of problem setting, for me at least with DeepPack, everyone’s talking about how we will outperform humans. But the reality is this is a problem that humans are really good at solving.

So it’s more about scale. In your context, for example, automating robots across the warehouse and coordination becomes very difficult. When it comes to a task where the human is standing at a picking station, for example, and just has to put stuff into a bag, it’s very difficult to beat. For us, the problem has transferred more into: We want to match human performance, or get as close as possible, but then add our AI to everything else around it

There are things that humans struggle with, for example, weight balancing in a container can be quite hard to do in your head. You can follow simple heuristics, which is what people do in industry. In the container loading, placing large, heavy items on either side, and then kind of starting to ‘feel’ towards the middle. But the reality is you cannot visualise the centre of gravity of a container in your head.

For us, we’ve arrived at the same conclusion: it’s more about reducing costs and making an attractive system rather than simply saying: ‘We’re going to beat your human operator.’

That’s exactly the challenge. Humans are annoyingly adaptable. (Laughs.) When it comes to picking and packing, no one really received training and our adaptability is phenomenal. But we have limitations. We are quite expensive and when we are underutilised, that’s where we tend to automate first. Look at storage retrieval. It’s a simple task: moving a box from A to B. That’s underutilising a person. And so we can automate that in a very economic way by putting robots on a grid to do that.

It’s a much simpler challenge than tasks that require adaptation on a second-by-second basis. We might’ve picked an apple before, but this apple we’re picking now has a subtly different size and shape. It’s the uncertainty within a challenge where humans excel. Part of the reason we focus so much on the economics is exactly that, Andy, you’ve hit the nail on the head: to deploy robots into production is a commercial endeavour. If you’re the retailer, putting in a person is very low risk.

In essence, we’ve also concluded that a person can pick and pack densely and do all the necessary quality assessment at a rate that is currently hard to achieve with a robot. So what we do is play a little game. Looking at the levers we have to solve the problem, what we do is deploy more robot arms than human stations. Now the robot doesn’t need to go as fast as the person.

If I’m going to deploy more robot stations than human stations, then I need it to net out. Not net out to zero, but still be better. So in the Ocado model, we provide a managed-service, end-to-end logistics platform for international retailers. Take Kroger in the US, it operates a number of our warehouses and Kroger employs people for picking. We’ll be rolling out robotic picking next year at scale into Kroger’s warehouses. For Kroger, there’s only one benefit and it’s reducing operation costs.

But the overall new product development model needs to factor in all that. Other technology we use is machine learning, which we improve over time by using human teleoperators. We augment our robot systems in real-time with human-in-the-loop supervision. This was part of the reason that we acquired a company called Kindred Systems.

Let me tell you a story about Kindred. When they first deployed an application called ‘SORT’ for apparel picking, they were doing 99% of picks manually through teleoperation. The robot was deployed but someone was using an interface to help the robots do the picking. They leveraged reinforcement learning so over time, the robot got better at using the data from the human teleoperators. Today, the SORT system achieves 99.5% autonomous picking.

Robots aren’t perfect in the early days. Their learning and improving is a process – and there, human operators help. We have tens of cells live at the moment and are planning to roll out hundreds of cells next year. Those cells are live doing real robotic picks in production. At the moment, it’s about improving the robustness before we scale tens and tens. We are very close to that point now.

One last economics question for you: Do your clients see this as a cost savings investment purely, or do they see it as an investment in the future?

Both. The goal is to be able to deliver Customer Fulfillment Centres more cheaply because we’ve taken out so much of the capital cost. With robots being on top of the grid, we don’t need the pick tunnel. There are so many benefits in terms of the warehouse design that we need fewer human operators and that transforms the economics of the warehouse itself. So there is a capital cost advantage, there is an operational cost advantage.

As a percentage of sales, pick and pack equates to about 50% of a fulfillment centre’s labour, which is about 3% of sales overall. And then if you look at other operations in the warehouse, such as decanting (unloading arriving goods), ultimately item-handling in a warehouse adds up to about 75% of the warehouse operational costs. So deploying intelligent systems that can ultimately do the item handling for everything in the warehouse takes out a significant portion of the operating costs. So we’re planning to save the retailer capital costs upfront, allow them to deploy the same volume and the same efficiency in a smaller space.

One of your comments about packing for weight balancing resonated with me. We have this process called tessellation as a customer adds to their shopping basket in our webshop. First, what the customer sees as being available on webshop is a real-time view of what we know we can deliver from our warehouse so we’re never over-promising. Andy, you look like a man who enjoys fine things – so when you’re ordering your third bottle of champagne, if we realise we’ve already sold out of Veuve Clicquot, we will say we can’t offer you a fourth bottle. But also, as you’re adding your third bottle of champagne to your basket, we’ve already begun this game of tessellation; playing Tetris in real-time. 

We’re trying to make sure that we can pack your order into the smallest number of containers. Because the number of containers that we use then drives the logistics fleet. The other thing we do in real-time is play a traveling salesman game to try to have the fewest number of drivers with the fewest number of hours, driving the fewest number of miles, in order to deliver all of the containers our customers need. Optimising to have the fewest number of containers per customer is an important step. It’s part of that overall optimisation and how the economics stack up.

Consequently, we have an algorithm that says how our customer orders should be packed ideally, we run this tessellation algorithm as you’re adding to your basket. There’s no way that we can convey that level of sophisticated information to a human as they are doing a pick and pack once every four or five seconds. 

This is a great opportunity for the future. The robots could pick using the tessellation algorithm and therefore get better packing density, and also improve the quality of the items once they are packed better. For instance how you avoid your avocado from coming into contact with your tin of beans. It has benefits with capital costs and deployments at the front end.

I imagine the supermarket industry has very low margins in general. At some point, what if everyone has this kind of smart platform? Do you think you’ll be in the same situation as you are now and just having to reduce costs?

Ocado is the only model for online grocery today that is profitable – others offer it but operate at a loss. But it comes at somewhat of a price premium when compared to a member of the public walking into a bricks and mortar store. So when we kind of consider the overall economic headwinds that everyone faces with rising inflation and cost of living, people are very sensitive. Even if the proposition is better, they get better range, better freshness, people are very sensitive to price ultimately. So today, our model is economic. But people pay a little bit of a premium for it. And when we’ve finished rolling out these technologies, our solution will enable retailers to offer groceries at a lower cost than bricks and mortar stores.

At that point, we have fundamentally disrupted the grocery industry. The technology we have is incredibly adaptable, not just for grocery, but for many other domains. In deploying this technology for our current retailers, any retailer that doesn’t have access to this technology should be pretty desperate to want to gain access to it, because it will enable them to deliver at a lower cost than they can today with their regular bricks and mortar. That’s the inflection point that really drives this kind of trophic cascade.

“Opportunities for optimisation are everywhere. And things like data science and machine learning are not held in my area alone. We try to embed them in every single team.”

Lots of people are doing robotic item handling – but the real challenge is access to the data and to the real world problem. This isn’t something that you can create in a lab. I mean, we did create it in a lab. But then when you move it to production, the real world has so many more variables. And so there are some other great robotics companies out there. But you need a way of learning in the real world and improving the system over time. The way that we operate our warehouses means, again I’m simplifying, but on day one, we had one robot and we only sent it a few items. We made sure that the robot on day one was really good at dealing with a few items. Then we told the high-level orchestration system, ‘Send it more items’. And now we’ve got a robot that’s got to pick more items than one robot can. Now let’s deploy many robots. And then we keep incrementally adding more items, once we’ve proven that it can pick those items, and ever increasingly shifting the operation from human to robot, but doing it incrementally. It’s really hard if you don’t have access to an environment like that, not just the data, but a friendly production environment where we are allowed to learn and sometimes we get it wrong. Sometimes we break an egg.

You mentioned fleet optimisation, routing and the travelling salesman problem. What other aspects of the logistics industry are you trying to optimise?

Everything. That’s a slightly glib, unhelpful answer. Ocado is an end-to-end platform so everything from e-commerce through to supply chain management, all the warehouse technology, and then last mile delivery. We leverage a lot of machine learning when it comes to supply chain. Our machine learning-based systems decide what to order – as in how many pallets of vitamins or how many cases of Coke – which far exceeds the human operators’ ability to order the right amount. Everything is an optimisation end-to-end.

Opportunities for optimisation are everywhere. And things like machine learning and data science are not held in my area alone. We try to embed them in every single team. Take our Mercedes Sprinter vans as an example. In the UK, there’s a weight limit of 3.5 tons before you get to a heavy goods vehicle license. So we take the Mercedes Sprinter unpainted and we paint it with only one coat of very lightweight paint. We take out the spare seat for the passenger. We take out all the insulation. We take out all the additional unneeded wiring. Where you have the window beside the driver, I know you don’t really wind it down these days because they’re electric, but we drill holes in a glass below the sill because there’s normally a little bit extra glass that goes below the door. And so we drill holes in the glass to reduce weight. We take out the spare tyre. So we put it on an F1-style crash diet, which reduces the weight of the vehicles by about 75 kg. This means we can get another two drops of orders on the vehicle. Optimisation is everywhere. It’s part of our DNA.