Muso Action to Develop and Demonstrate Logistics Warehouse Picking Foundation Model After GENIAC Selection

PR Times reported on September 9, 2026, that Muso Action will develop and demonstrate a foundation model for picking in logistics warehouses after being selected for GENIAC. We have not been able to verify the primary source, so please refer to the original for details.

According to Muso Action’s announcement, the plan is to collect data at a scale of 1 million episodes a year from operating sites for warehouse picking tasks. Whether warehouse automation will extend beyond transport to grasping and sorting is likely to depend on the quantity and quality of this on-site data. The plan to use three sites operated by Askul, Nippon Express and Wakamatsu Konpo Unyu Soko can be read as an effort to build a shared foundation across different products and tasks, rather than optimize for a single warehouse. In logistics operations, implementation will depend on how broadly the same model can be applied across different products.

What is new in this announcement is that it combines equipment for collecting on-site data with training a robotics foundation model for warehouse picking. Rather than a standalone demonstration, the plan is to collect 1 million episodes a year from operating warehouses at three sites — Askul, Nippon Express and Wakamatsu Konpo Unyu Soko — and further train a publicly available VLA model in two stages. In other words, it is designed to bring data collection, model training and field demonstrations together in one framework, with an emphasis on acquiring operational data rather than research alone. An earlier report by this newspaper, AIRoA Makes About 5,000 Hours of Robot Motion Data and a Foundation Model Publicly Available, showed the development of robot motion data and foundation models from the perspective of a data ecosystem. Muso Action’s project, by contrast, focuses on ongoing collection and adaptation at actual warehouses, not on preparing public data. Taken together, the two efforts show that competition in robotics foundation models is not only about the performance of trained models; where and in what form motion data is collected is also becoming a competitive factor. Another report by this newspaper, GROUND’s AMR “PEER 100” Adopted for Nippon Express’ “Warehouse for Everyone” Project, showed how warehouse automation is expanding in stages from transport to picking. This project targets the next step: grasping, selecting and placing goods. Key questions include how well the three sites can standardize the long tail of tens of thousands of SKUs, and whether the design’s power consumption — about 60% to 70% lower than that of humanoid robots — can be maintained in actual operations. A broader industry implication is that the bottleneck in warehouse automation is shifting from hardware alone to data collection and model updates that bridge differences between sites. The figure of 1 million episodes a year points to a supply network for ongoing learning, rather than a one-off proof of concept. Because the products at the three sites differ — daily necessities, snacks, and canned goods and seasonings — the degree to which one model can be deployed across sites, and how to distinguish that from site-specific additional training, will be points of evaluation. Open questions include how many SKUs the system can handle reliably after the demonstrations, its success and uptime rates in actual operations, and what can be generalized through the two-stage training based on a public VLA model versus what must remain site-specific. The specifications of the data-collection devices and the scope of any release of data collected at the three sites also need to be clarified.

Because the project will collect data at operating warehouses run by Nippon Express, Askul and Wakamatsu Konpo Unyu Soko, Japanese warehouse sites will serve as a training base for robotics foundation models. For domestic logistics operators, a key question is how much shared data can be developed for picking, a process with significant site-to-site variation. The published announcement provides limited detail on the specifications of the data-collection devices and the scope of any release of the collected data. For Japanese robotics companies and logistics operators, operational systems that can continuously capture on-site data may prove more decisive for competitiveness than the model itself.