APTO Opens Toyosu Data Factory, a Real-World Data Collection Facility, in Toyosu

APTO announced Sept. 10, 2026, that it had opened Toyosu Data Factory, a real-world data collection facility for physical AI and robotics, in Tokyo. The facility collects data using physical robots and develops data needed for robot learning. An opening ceremony was also held Aug. 21, 2026, with representatives from companies and research institutions in the AI and robotics sectors attending.

APTO’s announcement reflects a view that a bottleneck for physical AI lies not in robot hardware, but in the process of collecting and evaluating real-world data. The company says the facility will support everything from collecting data for imitation learning to building datasets, giving it a role in preparing training data for companies and research institutions. Representatives from NVIDIA’s robotics division and Toyota Motor Corp.’s Future Creation Center also attended the event, underscoring how difficult it is to separate model development from preparing field data. Turning know-how from real-world operations into data is likely to be one of the conditions for implementing physical AI.

Toyosu Data Factory’s distinguishing feature is its focus not on developing robots themselves, but on providing a facility to collect real-world data and feed it into training and evaluation. APTO already centers its business on data collection, annotation and preparation, and the new facility extends that work. For physical AI, however, data on movements, sensors and interactions with the environment is needed alongside model performance. The facility’s significance therefore lies not simply in providing a workspace, but in building infrastructure that affects the quality of input data. In our related article, NVIDIA develops the SONIC humanoid control model, we reported that NVIDIA was researching a humanoid control model. Kazumasa Hirano of NVIDIA’s robotics division spoke at APTO’s new facility, showing that model development and real-world data preparation are not separate matters but successive stages of the same supply chain. Masayoshi Tsuchinaga of Toyota Motor Corp.’s Future Creation Center also attended, highlighting the question of how to turn expertise from manufacturing, mobility and robotics operations into data. From an industry perspective, the facility is intended less to directly increase the number of robot deployments or applications than to bolster the supply of real-world data usable for training. Without enough such data, VLA and imitation learning can be difficult to generalize. If data collection and quality evaluation can be handled together, robot development could shift from a focus on prototypes toward continuous learning. However, the information disclosed so far covers the opening and the services offered. The types and number of robots the facility will handle, data volumes, utilization rates and scope of external access remain unclear. Those will be the next points to watch.

According to APTO, Toyosu Data Factory handles multimodal data for physical AI and robotics, including VLA (Vision-Language-Action), beyond images and text. This includes robot movements, camera and sensor information, and interactions with the environment. The company says it will support work ranging from data collection for imitation learning to annotation, quality evaluation and dataset development. APTO plans to use Toyosu Data Factory as a base to gradually expand the robots, tasks and types of data it can handle, while developing a data pipeline to connect data collected from physical robots to training and evaluation. The opening ceremony also featured a panel discussion attended by Kazumasa Hirano, head of NVIDIA’s robotics division; Masayoshi Tsuchinaga of Toyota Motor Corp.’s Future Creation Center; Taro Yamada and others.

In Japan, a key question will be whether expertise from manufacturing and logistics operations can be converted into high-quality real-world data. The annotation, quality evaluation and dataset development cited by APTO are relevant to standardizing field data, with implications for companies that have manufacturing and logistics know-how and businesses that prepare data for robots.