Ryosan Ryoyo launches introductory physical AI package for learning imitation and reinforcement learning on real robots

In October 2026, Ryosan Ryoyo began offering its “Physical AI Introductory Package” as a workforce development option under its corporate “AI Techmate Program.” The company offers two configurations that let engineers learn imitation learning or reinforcement learning using physical robots. The package is aimed at engineers beginning to develop robots using physical AI.

According to Ryosan Ryoyo, the package provides separate configurations for experiencing imitation learning with real robots and reinforcement learning using simulation, giving manufacturing engineers an environment in which to evaluate the differences between the learning processes. However, it is intended for research and evaluation, not mass-production equipment or commercial operation, and does not directly demonstrate the feasibility or effectiveness of deployment at work sites.

The package’s distinguishing feature is that it separates physical AI learning into “operation and data collection on real robots” and “trial and error in simulation,” presenting them as separate hardware and software configurations. The imitation learning option covers a continuous workflow, from remotely operating the SO-101 and collecting data to training and inference using models such as NVIDIA GR00T N1.7. The reinforcement learning option combines NVIDIA Isaac Sim and Isaac Lab to let users experience training that optimizes actions in simulation. Both provide learning environments; the announcement is not for operational robot systems or mass-production equipment. A related article in this newspaper reported that Forbes published an article on physical AI. However, publication of an article does not indicate that a company is offering a product. This announcement is not about discussion of the topic; it is a specific workforce development offering that brings together physical robots, computing environments, models and learning frameworks. The two should not be presented as the same kind of development: editorial coverage and a service offering are distinct. The industry implication lies in a model that bundles equipment and software for the learning process and offers them as a foundational learning environment, rather than simply selling a robot arm. In imitation learning, human operation is the entry point for data collection, followed by training and inference and then verification of the robot’s actions. In reinforcement learning, simulation provides a setting for trial and error. The two configurations distinguish a pathway using real-world operation data from one based on learning in a virtual environment. The announcement materials do not, however, confirm a mechanism for integrating data from the two or a process for transferring learning outcomes to robots at work sites. The terms of six months of email support, with up to five inquiries, individual quotes and research and evaluation use are also important for understanding the scope of the offering. It should not be confused with a deployment service designed for mass production or commercial operation, nor does it indicate uptime at manufacturing sites, connections to existing equipment or maintenance arrangements. Questions to watch include the tasks each configuration can handle, how training data is stored and reused, reproducibility on physical robots, and what support is available when moving from evaluation to commercial operation. These details are not included in the announcement materials, so outcomes beyond the package’s intended use cannot be inferred.

Package A for imitation learning covers everything from remotely operating the SO-101 and collecting data to training and inference with AI models. Its components include NVIDIA Jetson Thor, a camera, a tabletop tripod, the VLA foundation model NVIDIA GR00T N1.7, and LeRobot, a dataset management and training framework. Package B for reinforcement learning uses the SO-101 and NVIDIA DGX Spark, combined with a simulation environment in NVIDIA Isaac Sim, NVIDIA Isaac Lab and ROS 2. Both packages include email technical support for six months from the handover date, with up to five inquiries. Pricing is available by individual quote. The package is intended for research and evaluation, including proof of concept, evaluation and learning; use as mass-production equipment and commercial operation are outside its scope.

The offering presents a foundational learning option for engineers in Japan’s manufacturing sector, combining the SO-101 with training and inference environments. However, the announcement targets research and evaluation and does not indicate deployment for mass production or operational use at manufacturing sites in Japan.