Physical AI

Fujitsu’s approach to Physical AI

Fujitsu is developing Fujitsu Kozuchi Physical OS, a platform designed to enable robots to operate autonomously in environments where robots and people work together. The platform integrates data on activities occurring throughout the physical environment with data on the actions of individual robots, allowing both types of data to be analyzed and utilized in a coordinated manner. Built on a consistent foundation spanning cloud to edge, Fujitsu Kozuchi Physical OS is designed to ensure reliability and safety while addressing data sovereignty and governance requirements.
Through Fujitsu Kozuchi Physical OS, Fujitsu aims to automate and optimize operations across a wide range of fields, including manufacturing, logistics, construction, infrastructure, and healthcare. In doing so, Fujitsu will contribute to building a sustainable social infrastructure where people and robots can work together safely and with confidence.

Technologies powering Fujitsu Kozuchi Physical OS

Fujitsu Kozuchi Physical OS enables multiple robots and systems to work together in accordance with operational instructions by combining brain intelligence and spatial intelligence. Brain intelligence enhances robots’ ability to adapt to tasks by drawing on past experience and learning from human demonstrations, while spatial intelligence provides robots with information about their physical environment in a form they can understand and use.

A diagram of Fujitsu's technical platform, showing the integration of IT, AI, brain, environmental functions, and robots.

Physical AI Technology Strategy

[Brain Intelligence]
Imitative Learning Using Pseudo-Haptic Feedback to Recreate Human "Trial-and-Error" Exploration

Fujitsu’s pseudo-force and tactile imitation learning technology enables robots to reproduce the human ability to feel their way through a task by analyzing the current values of their own motors.
While conventional robots are highly capable of processing visual information, they have difficulty perceiving non-visual information, such as the amount of force being applied. This can result in damaged components or difficulty performing tasks that require fine adjustments. When a robot touches or pushes an object, subtle changes occur in its motor current. By analyzing these changes with AI, the robot can estimate how much force it is applying without the need for dedicated force or tactile sensors.

Four examples showing robots performing parts retrieval, cable connection, stacking, and heavy load transport.

Sim2Real: Sim2Real adaptive control - Infer their own context -

One of the challenges of AI is that it performs effectively only within the conditions covered by its training. Because a gap exists between simulated and real-world environments, often referred to as the Sim2Real gap, AI-powered robots trained solely in simulation may not perform as expected when deployed in the real world.
Conventional robot development has addressed this challenge through costly and labor-intensive approaches, such as reorganizing physical environments to make them easier for robots to operate in, building highly accurate simulators, and training robots with real-world data. Sim2Real adaptive control takes a different approach. It enables AI-powered robots to estimate their own state and surrounding conditions in real time, adapt to changing situations, and balance multiple tasks. This allows robots to respond flexibly to new and unfamiliar situations.

A comparison between traditional factory automation and delivery by an adaptive quadruped robot.

Loading component...

[World intelligence]
Spatial world model: The model the interactions between people, robots, and objects within a space to envision future behaviors.

In offices, commercial facilities, and other dynamic environments where people move freely, the unpredictable movement of people and objects makes autonomous robot operation challenging. Fujitsu’s Spatial World Model integrates data from cameras installed throughout the environment and cameras mounted on robots to create a 3D scene graph representing interactions among people, robots, and objects across the entire space. By modeling these interactions and predicting future actions, the technology enables autonomous robots to avoid potential collisions and operate in coordination with one another.

A diagram showing the process of spatial recognition, from information acquisition to future prediction.

Loading component...

Partner

Advancing Physical AI requires collaboration with partners that bring diverse strengths and expertise. Fujitsu is collaborating with NVIDIA, conducting joint research with Carnegie Mellon University, and working with robot manufacturers and other partners to accelerate real-world deployment.