Ocado’s AI-powered robotic arms: levelling up efficiency in online grocery and logistics

Recent advancements in AI and robotics have transformative potential for the online grocery and logistics industries, driving efficiency, reducing costs, and fast-tracking productivity gains. At Ocado Group we have over 20 years of experience testing and innovating our solutions in one of the most challenging automation environments in the retail space: online grocery.

Now, we are applying these insights to solve challenges in global supply chains. The latest game changer in this space is our AI-powered robotic arms, which offer a radical step-change in the efficiency of our customer fulfilment centres and unlock stronger economics.

Why is picking and packing notoriously challenging? 

Picking and packing are some of the most important and complex fulfilment challenges in global logistics. They are also some of the most difficult challenges to solve with robotics. Grasping and handling items, and determining the qualities of each product (fragility, weight, dimensions) are skills that humans are innately able to deploy. So much so that we take their complexity for granted.

Robots simply don’t possess the same sensory or cognitive abilities as a person, such as the cognition humans rely on to learn and generalise knowledge in new situations.

In no industry is this more challenging than grocery retail, where a whole new host of complexities are introduced, from differing temperatures to delicate items like glass bottles, and food categories which vary in shape and size, such as soft fruit. Some of the difficulties our robots face include: 

Variety - Our robotic arms come across tens of thousands of items of varying shapes, sizes, weights and fragility. For example, how you would handle a box of eggs would be entirely different to the approach used for a bunch of bananas. In grocery, packaging can also change seasonally, making it all the more important for robots to adapt and learn at pace. 

Packaging - Imagine the dexterity and perception needed to remove an individual item, such as a jar of honey, from a cardboard tray with a plastic overwrap, and the knowledge required to select the product and not the unnecessary packaging from the tote.

Care - Robotic arms need to take care to preserve the quality of the items they pick and pack, preventing stock damage and potential downtime from spillages.
‍
Density - Packing density is crucial to maintain the efficiency of a logistics operation. At the same time, robotic arms need to make sure bags/totes aren’t overloaded.

Rigidly programmed robotic systems widely used in other industries are not fit for the unique demands of the grocery space. To address these challenges we’ve developed a robust autonomous system, powered by advanced applications of machine learning, which is capable of making smart decisions on the fly. 

How are Ocado’s robotic arms solving the problem? 

On Grid Robotic Pick (OGRP) is Ocado’s robotic arm technology, automating the picking, packing and organisation of stock within fulfilment centres, directly from the grid.

Ocado’s in-house development teams have designed OGRP to operate in real-world contexts at scale —with a wide range of products. The real-life learnings our robotic arms have gained from working on the frontlines of a fast-paced online grocery operation have been invaluable, helping us to develop a solution that can adapt to a wide variety of environments and conditions.

1. Game-changing productivity 

OGRP is propelling productivity. Ocado’s fulfilment centres are already among the most labour productive sites in the world, with a 50-item order picked in a matter of minutes and hundreds of orders picked simultaneously. 

At full capacity the introduction of OGRP will enable a step-change in the improvement of labour productivity in these sites.

These changes are improving productivity by: 

• Extending the picking window - The robotic arm works around the clock, assembling customer orders at peak hours, and consolidating and organising stock in off-peak hours. 
• Enabling existing roles in the warehouse to be redeployed in wider ways - Our technology is designed to drive higher throughput and greater efficiency whilst also supporting colleagues by reducing the most arduous and repetitive activities associated with distribution and fulfilment centres.

2. Greater economics at scale 

OGRP is helping businesses to save on resources by:

• Making better use of available space - Enabling greater throughput from existing sites without requiring additional footprint in our automated warehouses to improve the efficiency and economics of site builds.
• Managing labour shortages - As businesses scale their operations, access to labour can be challenging. OGRP can help to manage these more effectively. 
• Addressing the needs of different products and environments - In highly regulated industries, having fewer human touchpoints is preferable from a security standpoint. For example, in pharmaceutical supply chains, the more people involved the higher the risk of product contamination. OGRP ensures a higher level of security within your supply chain, while reducing the costs required for employee training and security clearance.

3. Cutting-edge innovation to address unique industry challenges

Ocado Group is employing a cutting-edge combination of machine learning and robotics, which isn’t just unique in grocery but is cutting-edge in any sector.

Through smart Machine Learning models, our robotic arms learn how to:

Recognise and handle objects 

An OGRP operation is like a game of Tetris, only with much greater complexity, variability and in three dimensions. Picking is affected by the packaging, weight and shape of products. Packing is equally challenging as robots must consider how to densely pack items without using too many bags and taking care of which products are packed first to avoid crushing them. 

Our robots use Machine Learning to pick and pack products with no prior understanding of what they are and how they should be handled. Our proprietary machine vision systems enable this, identifying the optimal grasp points of a product to help our robotic arms to pick items correctly. Intelligent built-in sensors also help to reduce the risk of crushing or damaging products during picking and packing. 

Collaborate at scale 

Where the robot needs additional support, we use behaviour cloning, a new up-and-coming field within Machine Learning. A robotic arm could be confronted with tens of thousands of products to pick, and therefore tens of thousands of choices to make on the way to pick these items. Using this approach, humans can train robots on how to pick and pack items. This is later augmented with Reinforcement Learning, a technique which reinforces “good” picking behaviour to help the robot improve even more.  

We’ve designed our remote systems to scale rapidly, gathering data throughout training, which is later incorporated into our AI and ML models to train entire fleets of robots. The unique benefit of this way of learning is that the arm can then adapt on its own and share these learnings with a group of robots across the globe, so each error it makes becomes a lesson for the entire fleet.

4. Industry-agnostic technology 

For over 20 years, we’ve been using technology and expertise to solve the most complicated supply chain challenges in grocery retail. We’ve tested our technology in the most demanding of sectors and we’ve deployed it into production at scale.

Now, we are sharing and continually improving these learnings with other operations in new industries to help increase the consistency of operations, manage labour more effectively, and optimise efficiency.

What does the future hold for OGRP?

OGRP is already providing efficiency at scale for our partners, empowering them to unlock unprecedented levels of growth. 

With over 1,100 technology patents granted and filed and over 2,500 technologists in our team, we operate at the cutting edge of software and hardware engineering.

Our  teams continue to leverage the latest breakthroughs in Machine Learning. To expand OGRP’s picking capabilities and understand how to generalise these skills beyond its current applications, we are exploring diffusion – a model which underpins the GenAI revolution. This will allow us to tap into previously unattainable efficiency levels, as we continue redefining supply chains worldwide.