FANUC opens its robots to ROS 2 and NVIDIA as a record iREX 2025 turns Tokyo into a physical AI showcase

FANUC, the Japanese industrial robot maker, used the 2025 International Robot Exhibition (iREX) in Tokyo to open its robots to outside software. The company released an open-source ROS 2 driver, made Python a standard programming language on its robots and presented a physical AI collaboration with NVIDIA, according to its product announcement. The show ran at Tokyo Big Sight from December 3 to 6 and drew 673 exhibitors, the most in its history, and 156,110 visitors, according to final figures published by the organisers, the Japan Robot Association (JARA) and the Nikkan Kogyo Shimbun.

FANUC said the ROS 2 driver supports the ros2_control framework for real-time motion control, runs at an ultra-high-speed control cycle of 1 millisecond and covers its line-up from 3 kg payload robots to 2.3 tonne models, including the CRX collaborative series. The source code is published on GitHub under the FANUC-CORPORATION account. The repository is licensed under Apache 2.0, with its main branch targeting ROS 2 Jazzy Jalisco and a separate branch for Humble Hawksbill.

On the AI side, FANUC said it runs NVIDIA Jetson computers on its robots and has linked its ROBOGUIDE offline simulation software with NVIDIA Isaac Sim, so that customers can rehearse production lines inside photorealistic virtual factories before touching real equipment. FANUC robots are offered as OpenUSD SimReady assets inside Isaac Sim. At the show the company demonstrated robots that take spoken instructions, avoid obstacles, route cables with two arms and track moving parts while tightening screws.

Yaskawa Electric, FANUC's closest domestic rival, staged a different bet. On December 1 it announced a memorandum of understanding with SoftBank Corp., the Japanese mobile operator, to develop physical AI for spaces shared with people. In the first use case, a robot in an office linked to a next-generation building management system finds and retrieves a specific smartphone from a shelf. SoftBank said a vision-language model it developed runs on multi-access edge computing (MEC) servers, reads sensor, camera and building data and issues task-level instructions, while a vision-language-action model developed by Yaskawa turns those instructions into motion on a MOTOMAN NEXT robot. The two companies showed the system at Yaskawa's main booth and did not disclose financial terms.

The exhibition's own numbers show where the momentum sits. Exhibitors rose to 673 from 654 at the previous show in 2023, and visitors to 156,110 from 148,125, while the number of booths fell to 3,334 from 3,508, according to the organisers' report. Overseas participation grew to 140 companies with 481 booths, from 121 companies with 443 booths in 2023, although they came from 13 countries and regions against 18 two years earlier. Attendance peaked at 47,464 on December 5, and 10,147 visitors came from 95 countries and regions. A Humanoid Robot Forum featured talks on physical AI from NVIDIA, Kawasaki Heavy Industries and China's Unitree Robotics.

For FANUC the shift is cultural as much as technical. Founded in 1972, the company built its position on tightly controlled hardware, controllers and its own programming environment, which kept integrators inside a FANUC ecosystem. Opening a supported ROS 2 path and Python as standard lets university labs, AI start-ups and system integrators drive FANUC arms with the same open tools they already use for perception and planning.

Rivals have moved in the same direction at different speeds. Universal Robots, owned by Teradyne, has long courted the developer community with its own ROS driver, and ABB and KUKA offer simulation and open interfaces of their own. Yaskawa has chosen to pair its MOTOMAN NEXT platform with telecom and AI partners rather than lead with an open driver. NVIDIA, for its part, used the show's Humanoid Robot Forum to argue the case for physical AI in front of Japanese manufacturers.

In plain terms, ROS 2 is the open-source middleware most robotics researchers use to connect sensors, planners and motors, and ros2_control is its standard layer for commanding joints in real time. A 1 millisecond cycle means an outside program can stream new joint targets a thousand times a second, which is what learned policies need to react to a changing scene. Isaac Sim lets engineers generate training data and test policies on digital copies of FANUC robots. The Yaskawa and SoftBank design takes a different route, keeping a large model on the network edge for reasoning and a smaller action model on the robot for movement.

Open questions remain. FANUC did not publish customer names, pricing or a timetable for production deployments of the voice-driven applications it showed. Its GitHub repository listed 33 open issues as of October 2026, a sign that the driver is still being hardened by its users. The Yaskawa and SoftBank arrangement is a memorandum, not a supply contract, and its reliance on MEC means that latency and coverage of the mobile network become part of the robot's safety case.

The organisers have scheduled a new event, RoboNext 2026, for December 2026, with the next full International Robot Exhibition due in December 2027. Both will show whether open interfaces and edge AI move from booth demonstrations into installed cells.