Hitachi and FANUC have formed a strategic partnership to develop and commercialise physical AI for manufacturing, using Hitachi factories as the first operating environment in which AI models and industrial robots will be trained against changing production conditions rather than fixed demonstration tasks. Initial validation will take place at Hitachi manufacturing sites in Japan’s Ibaraki region, with the companies aiming to establish technologies suitable for customer deployment from fiscal 2027.
The first applications include picking components with different shapes and carrying out changeovers when the product being manufactured changes. Hitachi and FANUC will measure recognition accuracy, robot movement, takt time and quality while combining Hitachi’s HMAX Industry AI technologies with FANUC industrial robots and existing robot AI functions. Hitachi’s edge AI semiconductor will also be tested alongside the robots as part of the validation programme.
The work targets tasks that remain difficult to automate when parts, positions, sequences or surrounding conditions vary between cycles. Conventional industrial robots perform consistently where tooling, coordinates, inputs and timing can be defined in advance, but mixed production and frequent changeovers still require operators to resolve exceptions, adapt setups and decide how a task should proceed when the actual workpiece differs from the programmed assumption.
Hitachi will use its plants as “Customer Zero”, allowing the AI to learn site-specific production rules from operating equipment and real process data. The companies contrast that approach with development based solely on imitation learning or digital twin simulation, arguing that factory data exposes the system to the interruptions, tolerances and operating constraints that are difficult to represent completely in a virtual environment. Simulation remains useful for training and validation, but the partnership is intended to continue learning once robots are exposed to live production.
Takt time gives the programme an unforgiving performance measure. A robot may be able to recognise and manipulate a variable component in a laboratory, yet an industrial cell also has to perceive the workpiece, select an action, execute the movement, verify the result and recover from an exception without slowing the line beyond its required production rate. Recognition accuracy therefore sits alongside motion and cycle time in the validation plan rather than acting as the sole measure of AI performance.
Production changeovers create a second test because tooling, fixtures, robot paths, quality checks and material presentation can all change when a line moves from one product to another. Skilled operators and automation engineers currently absorb much of that variability through setup work and manual intervention. An AI system that reduces the engineering required for those transitions could extend robot use into higher-mix production where dedicated automation is difficult to justify, provided the resulting behaviour remains repeatable and auditable.
The partnership also has to deal with the control problem created by continuous learning. Manufacturing equipment cannot be allowed to change behaviour unpredictably simply because an AI model has seen new data, particularly where robots operate near people, tooling or expensive work in progress. Learned behaviours need limits, validation and a route into approved production logic so that improvements do not undermine machine safety, product quality or maintenance support.
FANUC brings a large installed base of industrial robots and motion-control systems to the programme, while Hitachi contributes production engineering, operational technology and data generated through its own factories. The combination follows other work aimed at making industrial robots more adaptable, including FANUC’s September collaboration with Palladyne AI on motion planning, teleoperation, simulation and repeatable deployment across manufacturing and logistics applications.
Edge processing will be assessed as part of the new partnership because perception and motion decisions often need to occur within production-cycle times. Keeping inference close to the equipment can reduce communication delay and limit the amount of raw production data that has to leave the factory, although model development, fleet management and higher-level analysis may still rely on wider computing infrastructure. Hitachi’s edge AI semiconductor gives the companies another variable to test alongside robot control and factory data.
The companies intend to use the operating knowledge developed at Hitachi sites when deploying solutions through their global customer bases from fiscal 2027. They identify semiconductors, pharmaceuticals, healthcare, advanced materials, automotive, logistics, food, shipbuilding and agriculture among the potential markets, spanning sectors with very different regulatory, hygiene, safety and production requirements.
Transfer between those industries will depend on how much of the learning and integration method can be standardised. A model trained to handle parts at a Hitachi plant cannot simply be moved into another factory without accounting for different robots, grippers, cameras, products, safety systems and production rules, so commercial scale will depend as much on deployment engineering as on the underlying AI model.
The Ibaraki trials give Hitachi and FANUC a controlled route through that problem because they can compare adaptive behaviour against established production processes before introducing it to external customers. Recognition accuracy, changeover time, takt performance, quality and exception recovery will show whether the system can move beyond physical AI demonstrations and reduce the engineering effort required to automate variable factory work.
