AIV: How AI Is Changing the Way AGVs Behave
What can artificial intelligence actually bring to industrial mobile robotics? After two and a half years of development, Alstef Group is introducing a new approach to how AGVs perceive and respond to their surroundings.
We sat down with Matthieu Crosnier, Alstef Group Warehouse Automation Sales Director, and Oilid Harbouche, Alstef Mobile Robotics Technical Director, to discuss the thinking behind the AIV (Autonomous Industrial Vehicle), the technology developed by our teams and the first applications taking shape.
What Does the AIV Change on the Factory or Warehouse Floor?
The AIV project started with very practical challenges encountered in the field. How can AGVs (Automated Guided Vehicles) coexist more effectively with operators and other equipment? How can unnecessary stops be reduced? And how can systems become more flexible and easier to implement?
The objective was never to develop artificial intelligence for its own sake, but to understand how it could help address these challenges.
“AI is not an end in itself. The idea was to build on our experience, as well as feedback from our existing installations and customers, to see how we could evolve our solutions and develop the features of tomorrow,” explains Matthieu Crosnier.
This thinking gradually led our teams towards a more fundamental evolution of the AGV: moving from an automated vehicle that responds to information and instructions from the system to a vehicle capable of analysing its surroundings and adapting its behaviour accordingly.
This is the principle behind the AIV, or Autonomous Industrial Vehicle.
What does this change on the factory or warehouse floor?
An AGV already uses safety devices to detect an obstacle and stop when necessary. But detecting something does not necessarily mean understanding what it is.
With its perception dome and artificial intelligence, the AIV can recognise different elements in its surroundings and determine their position, whether an operator, a pallet, a forklift, a load to be transported or a pedestrian crossing.
“The machine needs to be able to analyse and recognise its environment, and adapt its behaviour according to what it sees,” says Matthieu Crosnier.
This understanding makes it possible to develop behaviours suited to different situations. When approaching a pedestrian crossing, for example, the vehicle can recognise it and automatically slow down. During pallet pick-up or drop-off, it can interrupt its movement if a person is nearby and resume once the operator has moved away.
The aim is to avoid situations systematically resulting in a vehicle stop when an intermediate response, such as slowing down or signalling its presence, may be sufficient.
AI designed specifically for warehouse automation
Behind these capabilities is a form of artificial intelligence very different from the generative AI models now commonly associated with the term.
“We’re not talking about AI like ChatGPT. This is highly specialised AI that turns pixels into intelligence about the vehicle’s environment,” explains Oilid Harbouche.
The onboard cameras and electronics feed the perception dome, giving the vehicle a 360-degree understanding of its surroundings. The images are analysed to identify the different elements around the AGV and associate them with specific functions.
Today, the model can recognise around ten different categories. And these are not simply obstacles. A pedestrian crossing is information that may cause the vehicle to slow down, while a pallet may be the load it needs to pick up. This distinction is precisely what allows the vehicle to move beyond simple obstacle detection.
AI Developed In-House and Embedded in the AGV
To develop this capability, Alstef Group made a strategic choice: to control the entire AI development chain.
“We decided to control the complete AI chain, starting with the acquisition of data specific to warehouse and industrial environments,” says Oilid Harbouche.
It starts with data. AGVs from our installed base were fitted with dedicated kits to collect images from a variety of customer environments. The data is then anonymised and annotated before being used to build and train the models.
This means working with situations the vehicles actually encounter in operation: different types of racks, loads, equipment and traffic configurations. Data collection continues today to further enrich the dataset and expose the models to the diversity of real industrial environments.
But developing the model is only part of the challenge. It also has to run onboard an AGV. While some AI models require significant computing resources, here the system needs to operate on electronics suited to the constraints of an industrial vehicle.
This is where inference comes into play: our teams work on making the AI models compatible with the onboard AI accelerators, while meeting the robustness requirements of industrial operations.
AI and AGV Safety: Two Distinct Functions
Using artificial intelligence on an industrial vehicle naturally raises questions about safety. On the AIV, the functions enabled by AI do not replace the vehicle’s onboard safety systems.
“The machine maintains the same level of safety. The objective is to avoid unnecessarily triggering these safety systems and bringing the machine to an emergency stop when, for example, simply slowing down would be enough,” explains Matthieu Crosnier.
Understanding its environment therefore gives the vehicle an additional layer of information. Depending on the situation, it can slow down, stop or signal its presence while retaining its dedicated safety systems.
In the longer term, using AI for safety-related functions presents another challenge. While regulations allow such applications to be considered, the industrial certification ecosystem does not yet have sufficiently established methods to certify them. This is an area that is expected to continue evolving over the coming years.
From vehicle intelligence to system intelligence
After two and a half years of development, the first perception dome functions are now being demonstrated on Alstef Group AGVs in Rennes. But this is only the first step.
“The first step is to bring intelligence and autonomy to the machine itself. The second will be to feed all this information back into the supervisory system to further enhance and simplify its operation,” explains Matthieu Crosnier.
One example gives an idea of where this could lead: an AGV approaching a rack could identify available locations or the pallets currently stored there and feed this information back to the supervisory system.
The potential therefore goes beyond the behaviour of an individual vehicle. Information gathered in the field could ultimately enrich the operation of the system as a whole and support the development of new functions.
This is also what lies behind the name AIV, Autonomous Industrial Vehicle: the evolution of the AGV towards a more autonomous vehicle, capable of better understanding and adapting to the reality of its environment.