KDDI and Japanese startup RLWRLD have begun a collaboration to develop foundation AI models for robots, in a move focused on shifting artificial intelligence from processing digital data to operating robots capable of learning from and interacting with real-world environments.
The two companies revealed the collaboration on September 18, 2026, as part of a research and development project aimed at building robot models that combine training data captured from the real world with simulation data. KDDI is participating in the project through the GENIAC program, a program for developing generative AI supported by NEDO, an organization under Japan’s Ministry of Economy, Trade and Industry. KDDI’s project was selected on September 9, 2026.
What Is Changing in Practice?
The project focuses on two main areas: collecting robot-learning data automatically and developing foundation models with greater capabilities for performing tasks in real-world environments. RLWRLD provides a foundation model for robots, while KDDI is leveraging its expertise in data collection and its partner network to expand the scope of training and testing.
KDDI says that data collection should not be limited to recording movements performed by humans, but should also include the perspective of the person performing the task and its context. For this reason, the company uses an approach based on recording human movements from a first-person perspective, with the aim of providing data more closely connected to how tasks are actually performed.
KDDI also plans to use data from KDDI Group facilities, such as Lawson stores, along with other data that can be collected directly from locations. According to the material, this makes it possible to improve the construction of AI models by combining real-world data with synthetic data generated through simulation.
Proof-of-Concept Tests and International Expansion
KDDI’s collaboration with RLWRLD began in April 2025 through an investment made by the company’s venture capital unit. In August of the same year, a proof-of-concept test was conducted at a Lawson store using a prototype intended to handle specific tasks. The project then moved in December to collecting simulation data for the store environment.
KDDI says that the number of people working on the project has increased from 20 in April 2025 to 126 people in Japan at present, with work focused on data collection, file engineering, and model development. It also indicates that the collaboration aims to form a global organization capable of expanding the use of the RLWRLD model beyond Japan.
Why Does This News Matter?
The significance of this development lies in its addressing one of the practical problems facing AI robots: the need for diverse data linking perception, movement, and context, rather than merely text data or separate images. Combining recordings of human behavior with simulation may help reduce reliance on field testing alone, but it does not by itself prove that the models will operate efficiently in all environments.
The material does not present quantitative performance results or a timetable for launching a final commercial model; what has been announced so far is a research collaboration and proof-of-concept tests under the GENIAC project. Therefore, open questions remain concerning the extent to which simulation results will transfer to the real world and how the data and models will be made available to partners or the broader community.