While consumers may think of Ocado as an online grocery store best known for its fleet of 1,700 fruit and veg themed delivery vans, its other business is its UK-based technology division, Ocado Group.
As Alex Harvey, chief of advanced technology at Ocado Group acknowledges, to make its grocery delivery business sustainable, you simply can’t afford to have one of these enterprises operate without the other.
“Operationally grocery is very challenging,” he explains. “It requires frequent order updates, multiple lead times; and now we’re up against new speedy delivery start ups like Gorillas, burning their way through their VC’s cash!”
“It’s also a high cost to fulfil orders with short profit margins and a short shelf life – so it’s a hard problem to solve and for us to be able to do this economically, at scale.”
Which is where the technology comes in. Ocado’s mega-sized automated warehouses are a far cry from your typical supermarket’s regional distribution centre.
It’s largest, in Erith, North London, is 563,000 sq ft and is populated by a hive of 3,000 remotely controlled cartesian (linear) robots working across a series of grids Each grid is the equivalent of about three football pitches.
The bots – which are controlled by the company’s machine-learning ‘air traffic’ control system DASH sort through items according to the orders logged then, off grid, work alongside human workers at the pick and pack stage to place bags into crates, ready for delivery.
These bots can pick over 50 items every 5 minutes, a process that took over an hour using the standard conveyer belt method. Food can also be moved in and out of warehouses faster, reducing errors and cutting food waste.
According to Harvey, Ocado has 0.5% food waste as a percentage of total products, compared to the industry average of 3 to 5%. However, Harvey adds that the group is always looking at ways to improve the efficiency of its operation.
At the AI Summit, held earlier this summer, Harvey talked delegates through some of the work it was currently doing in this area.
The occasional piece of trash might come downstream from the supply chain, as robots are picking and packing – but it’s essential that this doesn’t end up in a customer’s shopping. And while the machines have been trained to identify items – they can struggle with visually similar ones – such as smooth orange juice or the one with bits in it, for instance, or the difference between a 60 and 120g bag of wild rocket lettuce.
Harvey explains how through supervised machine learning –the positive and negative ‘cats and dogs’ machine learning model – Ocado trains its bots around what can and can’t be picked allowing it to generate the correct grasp points for the right items. “This is able to generate effective masks allowing the machine to identify the correct item to be picked thereby not picking up trash or incorrect items,” he says.
Typically, when mistakes are made on the production line the machines do not have the autonomy to deal with them. However, a team of engineers can control the bots via teleoperation so that they can remotely recover errors that occur in the warehouse.
However, in addition to teleoperation, as part of the recovery process Harvey reveals that Ocado is also using a behaviour clone AI system to create an end-to-end machine learning model whereby the robot is learning from the person who is remote controlling the robot.
“It has been has been trained on over 200 hours of engineers remote controlling picking and packing, using robots and suction picks to pick and pack,” he explains.
“The emerging behaviour is incredible: they have learned to adapt. It’s like watching someone fly fishing: When it gets something wrong, it realises it has got something wrong, puts it back and recovers. Or it might fail to pick up an item so it will move to go and try and pick it up,” he adds.
Harvey observes that the machines have even learned how to adjust putting tricky items into carrier bags and uses a large size of Celebration chocolates as a case-in-point.
“The robot tries to nudge the box into he plastic bag if the item is protruding outside the container because over heights are an essential thing we have to avoid.
“So the robot has learned to pick the box up off centre so that it’s already dangling down and it uses the carrier bag to help place it into he container effectively.”
For heath and safety reasons, humans can not enter the grid: these heavy cartesian machines whizz around at a closing speed of 4m/s with 5mm between them, and a collision would be harmful.
However, this makes detecting robots that are about to fail challenging. To manage their health, the robotics team relies heavily on the huge amounts of data they produce each day (approximately 1GB per bot per day. Or, as Harvey puts it “if these were floppy disks each bot would fill two shipping containers daily”).
So when the team looked at methods for predicting failure for motion it started sifting through the timed motion data on each robots’ drive.
Aware of the “crazy amount of data” they were dealing with, the robotics team explored a number of solutions to help them, and found a novel reuse for Google Deepmind’s speech-to-text algorithm WaveNet.
Harvey explains: “Speech is another time-based signal, so we thought, ‘why not reuse it to analyse the movements of other time based signals? And it turns out WaveNet can characterise our speed profiles incredibly accurately, allowing us to detect bad moves and separate the good bots from the bad. So this kind of speech recognition tech can be redeployed in an automation environment.”
While the team had figured out how to keep their robots healthy, failures also occasionally occur with the track the robots are on. The difficulty here lies with the fact that the track is not intelligent; it has no sensors it’s just cheap extruded aluminium.
The team also needed to identify whether the bad movement was coming from failing bot or a failing piece of track that has a gap, a hole, or food or rubbish logged into it.
“Again, using AI and ML techniques we were able to spot characterises far beyond what a human would be able to detect with the data. This is where the power of AI and ML really comes into its own.
“We had a data scientist who joined the team six months out of university and had no knowledge of hardware or bots themselves. But just using well-curated data sets and some very good techniques she was able to provide a list of locations on the grid and identified anomalies for us to check.
One of these turned out to be a broken piece of track that created a step between two points that caused the bots to take incorrect moves. The data set from these bot outliers was able to differentiate between the bots and the bad grid locations.
Subscribe to our Editor's weekly newsletter
Share
