Exotec and AI in the Warehouse: A Comprehensive Analysis

How is AI, with its promises of enhanced performance but also its share of myths and misconceptions, finding its way into the supply chain? In modern warehouses, automated logistics solutions are increasingly powered by cutting-edge technologies. However, pinpointing or naming these technologies with precision has become challenging, given how overused the terminology around AI has become. From artificial intelligence to deep learning and machine learning, we delve into the topic with insights from Paul Vanhaesebrouck, Machine Learning Team Leader at Exotec.

Before exploring the practical applications of AI in supply chains and warehouse automation, it is crucial to define AI and its subfields. “The general public does not always use the term artificial intelligence correctly,” explains Paul Vanhaesebrouck. He elaborates: “This confusion often arises from its overlap with machine learning. Once we clarify this, we can begin discussing more concrete concepts. ”. Let’s start with a few definitions.

What is the Difference Between AI and Machine Learning?

The European Parliament defines Artificial Intelligence (AI) as “the ability of a machine to replicate human behaviours, such as reasoning, planning, and creativity.” Machines that employ AI are capable of analysing their environment and making decisions accordingly.

However, as Paul Vanhaesebrouck points out, “most of the time, when people refer to AI, they are actually talking about machine learning.” Is machine learning a form of AI? “Yes,” he says. “Machine learning is a subset of artificial intelligence.” This discipline, grounded in mathematical models, involves building systems that can learn from data collected from their environment.

What is the Difference Between Machine Learning and Deep Learning?

Within machine learning lies an even more specialised field: “In machine learning, we pay particular attention to deep learning, which involves training artificial neural networks.”

What are the practical applications of deep learning? Broadly speaking, when we discuss breakthroughs such as GPT-based language models or image generation systems, we are, in fact, referring to deep learning.

At Exotec, the focus is on deep learning — not to generate images but to analyse them. The goal: to better interpret visual content by recognising objects, counting them, identifying their shape, and determining their position within a storage container.

How does Exotec apply machine learning — and specifically deep learning — to warehouse automation and supply chains?

A prime example is Exotec’s Skypod® system robots, which are equipped with cameras to detect obstacles. “The robot analyses the image captured by the camera to determine whether there is an obstacle or a person and estimate the distance,” explains Paul Vanhaesebrouck.

The picking robot developed by Exotec demonstrates this concept further. Designed for goods-to-person logistics, the robot retrieves items from a storage container and places them into preparation bins. Deep learning comes into play here: a camera positioned above the container captures images, enabling the robot to precisely identify the item to be picked and deposit it accurately.

However, this process poses a challenge: “It’s impractical to write code that explicitly explains what to do with every pixel in an image — it’s too complex. Instead, we take a reverse approach: rather than designing the algorithm, we define the objective, meaning the desired behaviour based on the image. The algorithm must then learn to replicate it.” The challenge lies in diversifying the data: “The more varied the types of items, quantities, and positions, the better the algorithm can generalise and produce accurate results.”

The machine learning process unfolds in two main phases. First, “we collect data (or generate it artificially) in a warehouse”, Paul Vanhaesebrouck explains. Then, “we develop the programmes that will train the model.”

What does this involve? “We start by defining a neural network architecture tailored to the problem we’re solving. Using machine learning tools, we determine what constitutes a correct or incorrect response, helping the neural network make decisions.” At this stage, the neural network is not yet operational: “The objective is to determine how much importance each neuron assigns to the information it receives from others — essentially solving a mathematical optimisation problem by assigning weights in a graph.”

To train these models, datasets are crucial — and, in the case of Exotec, these datasets comprise images. Generating ‘synthetic data’ (artificially produced images) can complement real-world data. Synthetic data offers two significant advantages: it can be produced rapidly in large volumes and comes perfectly annotated, unlike real-world data, which requires manual annotation.

Exotec’s deep learning team leverages both methods.

What advantages does machine learning bring to the supply chain? Paul Vanhaesebrouck summarises: “It enables us to design algorithms that would otherwise be impossible to articulate in human-readable code.” The result: highly adaptable systems capable of handling diverse scenarios without added complexity or maintenance costs.

The robots are programmed to respond optimally to new situations, enriched by the experiences artificially generated through machine learning.

“It’s easier to develop algorithms that train neural networks than to design algorithms to process images directly,” he continues. The outcome is both efficient and scalable: a system capable of recognising and responding to a wide range of items to meet Exotec customers’ needs.

“The ultimate goal is to develop ultra-efficient picking systems that can handle any item, regardless of its shape, material, or position.” (For further insights, see our interview with Louis Esquerre-Pourtère, R&D Manager at Exotec.)

A final advantage: “Machine learning often allows us to develop simpler, more cost-effective solutions, while also reducing the environmental impact of our systems.”

Simple, efficient, and economical — machine learning is already transforming warehouse performance. And what lies ahead? Paul Vanhaesebrouck concludes with a look to the future: “Advances in hardware running neural networks will be pivotal. As equipment becomes more compact and efficient — as seen with smartphones — we’ll unlock new possibilities across industries, integrating AI into smaller, smarter systems. ” To be continued…

In the meantime, discover how our Skypod automated warehouse solution works.