Depictions of the future in books and film are usually influenced by what’s going on at the time, reflecting social malaise, impending armageddon, or economic anxieties. The robots in classic sci-fi usually resembled humans, as most authors assumed they would eventually assist us in the same tasks humans did.

Instead, however, today we find artificial intelligence doing some of our thinking for us, but it’s more often solving problems that don’t need intervention from self-contained humanoid robots. The promise of autonomous robots that matched our abilities has given way to a more specific focus on tasks that are fulfilled by armies of smaller bots controlled by machine learning and algorithms running in the cloud. Their scope is more complex, but a lot less dramatic than, say, Forbidden Planet’s Robby the Robot, or the replicants from Blade Runner.

Not so long ago Ars Technica looked at the uses of robotics in retail and the range of different approaches, from wheeling shelves around giant Amazon warehouses, to Domino’s forthcoming airborne pizza drones. The response to that story was very positive, so we decided that we would follow up with another piece that dives even deeper into the world of automation. The first story focused on what is being done with automation; now we’re going to look at how you build an automated system.

Why automate?

The assumption is that, despite the large amount of effort and cost of designing and deploying an automated system, the end result will be worth it—that the automated system does the job better and cheaper than before.

Automation is faster than the human alternative. Even if a bot is slower in absolute terms, it doesn’t suffer from fatigue, makes fewer mistakes, and can work 24/7 to make up for any tardiness. Furthermore, unlike hard automation, robots are adaptable and can be taught new tricks.

This sounds satisfactory enough, but off-the-shelf robotics will inevitably present shortcomings for those who truly want to exploit magnificent machine-learning algorithms in real-world settings. Factors such as a robot’s physical design may be less than ideal for a specific environment, or the sensing and control mechanism might not be suitable for the intended workflow.

The choices faced by any big firm that wants to go robotic, or indeed a solo inventor who wants to craft the first Cylon, are stark: redesign an existing system to try and suit your needs, or build a new one yourself. Customising someone else’s robots is cheaper, but it’s necessarily a compromise and, done poorly, could totally undermine the advantages and efficiencies that come from automation.

So just how hard is it to build an army of robots from the ground up?

That’s a good question to put to Ocado’s CTO Paul Clarke, whose division Ocado Technology, in partnership with Ocado Engineering, is developing robots that will one day allow for fully autonomous warehouses. Ars was invited to take a closer, exclusive look at the company’s homegrown robots that are currently live-trialling at its new Andover facility.

Robots, Clarke explains, are one part of a grocery retail enterprise that manages incoming goods in warehouses known as “customer fulfilment centres” (CFCs), and ferries them around to human pickers, who process orders by placing items in separate baskets to make up customer orders. Customer baskets then end up in a van and are put out for delivery.

One day the human basket-fillers will be joined by robots equipped with compliant, human-like grippers (Ocado is involved with the EU’s SoMa project), and some of the vans may be replaced with autonomous delivery vehicles, and all of these robots will be tended to by humanoid SecondHands robots that wander the CFC floor looking for machines in need of maintenance… but we’re not quite there yet.

For now, the focus is on building a hive of thousands of custom-built robots that bring products to the humans, optimising their behaviour, and reliably communicating with them so that they don’t have another Rogue Robot Situation on their hands.

But we’re getting ahead of ourselves. Before you raise an army, particularly a robotic one, you need to define the army’s operational theatre. Furthermore, you have to work out how the army will behave within that arena and how it will physically go about achieving its goals. To do all that, the first step is collecting and analysing a lot of data.

Optimisation in 4D

Ocado’s current-generation CFCs in Hatfield and Dordon are assisted by a neural network-based AI. The AI, with the help of forecasting algorithms, optimises the layout of the warehouses to improve the speed at which orders can be picked and dispatched.

These forecasting algorithms “do a lot more than just look at customer preferences or what their favourites are,” Clarke explains. “We do time-series analysis and predictions around how fast they will consume products that they’ve bought, based both on their behaviours and other people’s,” Clarke says, enthusing about an instant order feature that, with a simple click, predicts what you want in your basket based on past consumption patterns. “But our vision goes much further.  We want to build an online retail platform where the right groceries turn up at the right time, without you having to do anything, as if by magic,” he says. “The more information we have on our customers’ households, behaviours, and desires, the better we can do that.”

Ocado’s use of customer analytics isn’t just limited to preparing instant shopping selections; it repeatedly crops up in the company’s modus operandi, whether it’s forecasting the traffic congestion that van drivers might encounter in order to determine delivery schedules, or estimating order quantities of the 50,000 different items it sells to meet continuously variable demand.

At the company’s new Andover facility, where a hive of warehouse robots are being deployed, data plays an even more important role: robots, controlled by a central AI, alter the layout of the facility in real time. Companies have tried similar ideas in the past with just one or two robots, but Ocado wanted something a lot more flexible and with higher throughput—something that would allow for several thousand robots operating on the same grid system.

“The original idea for this kind of storage mechanism really dates back to container ports,” says Clarke. “In that case, you have container cranes that roam over a grid of boxes and move them around. The way these robots are choreographed on the grid is like a complex, real-time air-traffic control system which is effectively solving a 4D optimisation problem. You’ve got the 2D motion of the robots, you’ve got the third-dimension optimisation of the grid [the storage bins underneath], and there’s a fourth dimension, of doing that while looking ahead in time.”

The publicity images of the bots give no idea of their size, but a rare video of the real thing in action seen by Ars shows them to be about waist height, slightly wider than a washing machine. They’re certainly not picnic baskets; rather, they’re weighty beasts that hoist the bins of goods from this hive of storage up inside their chassis. They then move these bins to another area where they are distributed to human pickers before being sent back, ready for retrieval and repositioning to another storage location within the grid.

However, given that orders are constantly coming in from customers, the best place to return something like a bin of cucumbers, is constantly changing, Clarke says.

“A bin’s position in the grid not only depends on the current moment but also on looking ahead to future orders that you’re going to be processing and new inventory being deposited into the grid. And at the same time you want the best route-optimisation at a 2D level as the robots fly around. It’s a complex problem; it’s one of the mathematically and computationally hard problems that has led us to explore technologies such as quantum computing and massive clusters of GPUs.”

For now, as true quantum computing is still a little ways off, these decisions are guided by machine learning, but there’s a nuts-and-bolts aspect to the developers’ challenges. Simulations can be used to stress-test the design and the control algorithms, but there comes a time when you need to roll up your sleeves and get your hands dirty.

Building the robots

Hardware-wise, the hive robots were designed (and patented) by Ocado and then built by a third-party manufacturer. Getting exact details out of Ocado was difficult, but Clarke did say “they are equipped with an array of sensors, actuators, and motors connected via a CAN bus and controlled and monitored by a microcontroller-powered brain running embedded C code.”

“When we built our warehouses in Hatfield and Dordon, we bought in the automation hardware from German and Austrian companies, and we effectively discarded the software and rewrote it right from the tin upwards. In terms of the robots in the new Andover facility, the big sea change here is the hardware. It’s our design, our IP. We built them in collaboration with some external partners but it’s not bought-in hardware.”

The robots are battery powered, which simplifies some of the other issues with having hundreds of washing machine-sized bots rolling around a warehouse at high speed. The average energy consumption per robot is only about 200 watts, which is bolstered by regenerative braking. The power pack itself is about the size of a car battery and features a cell chemistry that lets it charge up in a few minutes and run for several hours. Energy efficiency was a key design consideration but not just for endurance: when the bots are charging or moving goods around, it was important to ensure they wouldn’t generate enormous amounts of heat.

Further up the stack there’s “high-performance” embedded hardware running Linux, components (presumably custom hardware and software) that implement Ocado’s custom wireless protocol (more on that later), and then a real-time control system, written in Java, that integrates data from the robot’s sensors with the rest of the hive. While Ocado is using ROS (the open source Robot Operating System) for its Horizon 2020 robotics R&D projects, “the challenges that we faced with our OSP robot hive were sufficiently complex and proprietary that we took an alternative, more integrated and computationally efficient, approach.”

Ocado went through a number of exploratory stages before hitting upon the right design, admits Clarke, but they did eventually end up with a system that can cope with quickly picking and storing goods in a highly dense and scalable manner.

Building a hive mind

Connecting all of the robots together is a low-latency communications system underpinned by a bespoke implementation of the 4G LTE air interface. The system, according to Cambridge Consultants, Ocado’s wireless protocol development partner, is “the densest mobile network in the world.”

The two companies collaborated to successfully engineer a communications system capable of controlling hundreds—and, potentially, thousands—of robots operating on a grid the size of several football pitches. Behind this effort were Ocado’s principal wireless engineer, Adam Green, and Nathan Wrench, head of industrial and energy at Cambridge Consultants, one of the guiding lights behind Silicon Fen, the startup-heavy area that’s grown up around the university.

Wi-Fi has always been inadequate for a grid like this purely in terms of scale, as a typical access point supports between 64 and 256 concurrent clients, but it did have features that intrigued Green.

“I do remember when we first came to speak to Cambridge Consultants, we were interested in using Wi-Fi because it has a point-coordination function (PCF) in the standard,” he says.

Green had hoped to use this feature to synchronise the bot cloud, so that each one would have its voice heard, so to speak. He likens the scenario to polite conversation at a dining table.

“If there are four of you,” he says, “you can have one conversation very easily; you can all take it in turns to speak. That works very well. The problem is when you have hundreds of people at a state banquet. It’s not possible to have one conversation where everyone gets to have their say—some people won’t say anything. When you’re controlling thousands of bots, and one of them doesn’t get a chance to say anything, it means the other bots don’t know where it is and there’s a collision risk.”

Scale issues aside, the point-coordination function in the Wi-Fi spec did seem to solve the problem in theory. However, putting it into practice became a challenge of a different sort.

“We couldn’t find any chipset manufacturers that would sell it to us. We were wondering if Cambridge Consultants knew of anyone. They didn’t either.”

Green admits that part of the problem was that chip makers would want to sell 10 million units—not the quantity he had in mind at this stage. He’d also discovered that Atheros (now owned by Qualcomm) sold a chip with PCF, but it had been withdrawn from production as it didn’t sell, and no-one else had filled the gap in the market. Such are the real-world problems that engineers have to work around. Keen to get fresh eyes on the problem, Green had Cambridge Consultants examine his findings, in part to confirm what he and his team already knew.

“There was a very hard bit of work just looking at the technical feasibility, just on first principles, from which emerged a pretty short list of possible routes forwards,” Wrench admits.

They soon realised that together they were looking at a clean slate, which, rather than undermining the venture, opened up new avenues. This led to the use of off-the-shelf 4G LTE components with additional modifications.

“There’s no new custom silicon that we’ve designed in this board, as we have been able to do everything with components available from one vendor or another. There’s one slightly exotic chip in there which is one of the few single-source components that we’ve had to design around. That was quite a significant choice that we made early on,” says Wrench, declining to give further details.

The system uses the unlicensed 5GHz band, and, given its original purpose was to function in warehouses, it operates over a short range. However, the Ocado wireless protocol, which has yet to receive a brand name, can be configured to suit different scenarios that go beyond the confines of a customer fulfilment centre, but more on that later; back to that noisy dinner table. What did they do to get this conversation point solved?

Green gets technical:

“The wireless protocol uses orthogonal frequency-division multiple access (OFDMA) which is used by LTE for the downlink air-interface. However, we have replaced the media access control (MAC) layer with our own implementation and entirely removed the upper layers of the cellular offering, which were not relevant for this particular project.

“OFDMA systems have the notion of a resource map—where the frequency and time allocation for each node is described. One unusual aspect of the Ocado wireless protocol is that the map is fixed for each client as we add it to the system, rather than LTE or WiMAX, where some element of scheduling is involved.”

This resource map is split up into a maximum of 1,800 tiles, with each tile potentially linking to an individual bot. The actual numbers vary because the spectrum must also accommodate a separate communications channel that handles bot maintenance, with 600 tiles being a more typical use-case. Regardless of the number of operating bots, they are synchronised so that each unit transmits and receives a message 10 times per second. If there is no data present for the control and status messages, then datagrams padded with zeroes are sent as protection against silent failure. As they all get a turn to chime in, they don’t disappear from the grid, so collisions are avoided.

And when it comes to operating at scale outside the warehouse, in say, an agricultural setting, there is potential for considerable expansion. Within the few hundred megahertz of spectrum that the Ocado wireless protocol occupies in the unlicensed 5GHz band are 24 channels, each channel being 10MHz wide.

“Because we have a good quality RF front end,” says Green, “these systems can cope with being next to each other in the RF space. So I could install a row of 24 base stations, supporting 24 lots of 1,800 clients in a 165m radius around me. Or I could install attenuators to reduce the power to bring the cell size down and then install multiples of that.”

In silico

So that makes 43,200 robots from just one 24-channel system, but even that doesn’t represent the limit. If you added more of them and lowered the signal strength, you could theoretically have over half a million bots running from just 12 base stations. As a back of a beermat calculation, that’s impressive, but the real world might make it a bit less feasible. If as a stress test you broadcast a load of RF interference to evaluate the resilience of radio transmissions on the grid, you’re likely to break the wireless communications for everyone else in the vicinity. That’s not only impractical but frowned upon from a regulatory standpoint.

If all you want to do is see how well the devices cope with degraded signals, there is a simpler way, as Green explains: “What we can do is use a conductive approach. So rather than having an antenna and a receiver radiating, you join them with a coaxial cable, which gives you all kinds of scope. That’s a very powerful and underused approach. I think what is more important is that we have this very high-quality RF chain.”

Beyond that, everything else in the Ocado wireless protocol is implemented in software, enabling simulations for test routines. Combined with simulating the design and control of the robots themselves, Ocado is almost capable of modelling the entire system in silico.

“We have all kinds of opportunities to test the hardware on a bench,” he says. “So if you can imagine a server room—loads of boards that are RF screened with our kit in—with a cable joining them all together, building the latest version of the firmware. And with our expertise, we’re able to add all the failure cases that we’ve seen so far, and all the ones we think we’re going to find, with some great tools available to purposely inject errors into interfaces between systems. And that is how we all gain confidence that we know what the limits of the system are.”

Moreover, the matter of external interference is attenuated by virtue of the design of the building. Modern warehouse construction includes a shell of Kingspan phenolic insulation panels that feature an alloy sheet in the foam to act as a shield against RF frequencies, leaving the door as the weakest point. Lest we forget, another factor that helps avoid disruption is that the range has been deliberately limited to operate within 165 metres, though Green admits they haven’t actually got a warehouse big enough to test those limits in the flesh.

Robots must be resilient

This is all well and good in a perfect world with no component failures, and just the one base station. However, the Ocado wireless team takes a more realistic view when configuring communications with the bot-hive at its customer fulfilment centre in Andover. Although one base station can cover the whole grid there, base stations are tethered in pairs, providing a hot standby that’s ready to go at a split second’s notice. Another duplicate pair is also present to test roaming features, for a total of four enabled base stations—even though only one is actually needed.

Roaming involves lots of spectrum optimisations, and a warehouse can be a surprisingly tough environment to conduct real-world tests. In a scenario where hundreds of bots travel at speeds of up to 4m/s, or 9mph, over 100m distances, the handover between base stations was an interesting problem to solve.

“When you leave your house in the morning to get to work, your phone will roam between several base stations on your journey,” explains Green. “In an enterprise Wi-Fi environment this can cause problems for real-time traffic because it takes in the order of a few hundred milliseconds, which is too slow. If our packets go missing in the middle of that, we’re not going to know where the bot is.”

As it stands, the data rate for Ocado’s wireless protocol is quite low, “but we guarantee we can speak to these bots 10 times a second,” says Green. “For the dimensions of the grid cells and the speed at which the bots go, we can receive status messages every time they traverse into a new cell, which is the important thing when you’re doing real-time control. We don’t believe there’s anything like this in the world, so far.”

And if everything does go pear shaped on the control side and communications are lost, the control algorithm has it all in hand. As with air traffic control management, the bots are given clearance to move to a specific location and are tracked until they arrive. If communications fail, the bots would finish their last command and come to a rest. What makes this possible is the fact that the bots are battery powered: even if there’s a power failure elsewhere in the system, there’s no random halting of bots across the grid.