Training AI patents
- publish258
- Dec 4, 2024
- 2 min read
Updated: May 22
It’s the origin of the data that matters when training AI patents, says Robert Klinski, as in the case of autonomous forklifts in the warehouse
Robert Klinski
In warehouses, as elsewhere, AI lets you react superfast to changing environments without the necessity to run complex algorithms. When an autonomous forklift can calculate its optimum route, it can steer around the collisions to which it might otherwise be prone.
However, for anyone leading the transformation of our warehouses, it is less than certain how they can fully secure the rights in forklifts for speed stacking our deliveries. How much confidence can they have when filing a patent? It is a question being widely asked about AI in board rooms and research departments everywhere.
The characteristics of the data for training the AI model are decisive at the EPO. It may not consider AI models as such patentable, however, it is more open to training an AI model with specific data for a specific technical task, ie, specifically provisioning an AI model, which is considered a technical purpose and therefore accessible to patent protection.
A distinction is made about where the data for training an AI model originates. For example, a model for controlling a forklift’s operations can equally be trained with real-world data originating from sensing the environment or from the simulations of a digital twin.
Such a computer simulation is, according to the EPO (decision G 001/19), not sufficient to overcome its patentability hurdle. Conversely, a physical interaction with the real world, such as an environment in which an autonomous forklift operates, may qualify. The implication is that the data for training an AI model shall directly originate from physical measurements, as gathered by a forklift’s sensors.
In order to provide the data for training the AI model, the forklift can be equipped with a positioning system indicating its current position and movement. Upon the basis of such information, the AI model can dynamically be trained.
In this example, the dynamically trained AI model is specifically deployed to control operations of the forklifts to determine the most efficient movements within the storage area, based on the position of the trucks and the disposed storage goods. This interaction with the real world may render the AI implementation patentable.
• Robert’s Klinski article on AI patents appears in the 2024 edition of ‘Winning with IP: Managing intellectual property today'. See here for details.
