01—Hitachi wants Physical AI to learn the way excavator operators actually work
Hitachi Construction Machinery announced on September 9, 2026 that it will begin a new Physical AI research program in October.
The idea is straightforward but ambitious: instead of defining every excavator action as a fixed sequence, train an AI system on real operator behavior.
Hitachi says the project will use information such as camera images, work instructions and operator lever movements. The AI will learn how experienced operators respond to changing conditions, then use that training to decide and execute appropriate machine operations.
The long-term target is autonomous construction machinery that can perceive the worksite, decide what should happen next and act through the machine itself.
02—The training data starts with human operators
Heavy equipment does not work in one perfectly controlled environment.
An excavator may face different object shapes, positions, ground conditions, weather, machine states and work instructions. Hitachi says its training data will therefore include extensive work footage and operation logs collected from human operators.
The dataset is also expected to include more than ideal examples. Hitachi specifically says the project will collect information from regular operations as well as recovery processes after operational errors.
That is important because useful physical intelligence has to learn what a good action looks like when the world is not perfectly arranged.
03—The system is built around perception, decision and action
Hitachi describes its Physical AI approach as a three-part loop.
First, the machine understands the surrounding environment.
Second, it optimizes the work plan and decides what operation should happen.
Third, it executes that operation through the construction machine.
This makes the project different from simple mechanical automation. The goal is not only to repeat a pre-programmed motion. The machine is expected to interpret changing conditions and select an appropriate response.
That perception-to-decision-to-action loop is the same broad pattern appearing across modern robotics, but here it is being applied to hydraulic construction equipment.
04—The excavator itself becomes the physical output of the model
A language model can stop at an answer.
Physical AI has to turn a decision into motion.
For an excavator, that means the model ultimately has to connect site understanding with lever-level machine behavior: boom, arm, bucket and vehicle movement have to become coordinated actions.
Hitachi is not publishing a final autonomous product yet. This is a research and development program.
But the architecture is clear. Human operational experience becomes training data, the AI learns a decision model, and the construction machine becomes the system that executes those decisions in the physical world.
05—The project brings together machinery, AI training and a robotics foundation model
The work is being developed through an industry-academia collaboration.
Hitachi Construction Machinery will coordinate the project, collect excavator operational data and build the research and verification environment.
Jizai will work on data-utilization technologies and the training environment required for AI development.
The Nara Institute of Science and Technology will research and develop a Physical AI foundation model for autonomous hydraulic-excavator operation.
That division of work shows how broad the problem is. Autonomous heavy machinery is not only a model-training task or a mechanical-control task. It needs both, connected to real operational data.
06—GENIAC brings the project into a wider Japanese robotics AI program
The project was selected for GENIAC, the Generative AI Accelerator Challenge supported by Japan’s Ministry of Economy, Trade and Industry and NEDO.
The selected work focuses on robot foundation models and the social implementation of generative AI.
For Hitachi, that framework gives the excavator project a specific research path: collect real-world machinery data, build a training environment, develop a foundation model and verify it against the changing conditions found on actual worksites.
The result is a Physical AI project grounded in one of the most demanding kinds of industrial machines.
07—Future applications go beyond one construction task
Hitachi says the future target includes autonomous excavation and loading at construction and mining sites.
It also points to debris removal at disaster sites, where machines may need to work through environments that are difficult to predict in advance.
The company lists other possible fields including material handling, forestry, mining, demolition and civil engineering.
Those applications share one requirement: the machine has to respond to the physical environment rather than operate only inside a fixed factory cell.
08—The Upgrade Feeling
The interesting part of Hitachi’s project is where the intelligence comes from.
The company is not starting with an abstract robot that has to discover heavy-equipment operation from nothing.
It is starting with humans who already know how to operate the machine.
Camera footage provides context. Work instructions provide intent. Lever movements provide action. Operational logs preserve what happened.
Physical AI can then learn the relationship between those pieces.
That is the upgrade: turning skilled human operation into a training signal for a machine that may eventually understand the site, decide on the next action and carry it out itself.