FastLabel Inc. (Headquarters: Shinjuku-ku, Tokyo; CEO: Takeshi Suzuki; hereinafter “FastLabel”), a company that supports the development of robot foundation models and domain-specific VLA (Vision-Language-Action) models, has released the results of its research and development on robot foundation model development using simulation data, conducted under the “Physical AI Development Support Program by AWS Japan” implemented by Amazon Web Services Japan G.K. (hereinafter “AWS Japan”).Through this program, FastLabel showed that adding simulation data to real-robot (real) data for mixed training raises the model's task success rate by an average of 14.2%. The results suggest that this approach can reduce dependence on real-robot data, which is costly to collect. BackgroundResearch in Japan and overseas in recent years has reported that “scaling laws,” under which model performance improves as the amount of training data grows, are beginning to hold in the field of physical AI, including robotics. This makes securing enough training data essential for improving model performance. However, collecting real data with physical robots comes with many constraints, such as setting up operating environments and the time and operational costs of collection, so it is not easy to scale. For companies developing physical AI, collecting massive amounts of training data in the real world is a major burden in both time and cost.To address these challenges, expectations are growing for simulation data, meaning training data generated in virtual space. In a simulation environment, it is possible to reproduce motions equivalent to those of a physical robot while generating data at low cost and on a large scale. However, the optimal training ratio of real data to simulation data (the data recipe) has not yet been established across the industry, and companies are still working it out through trial and error. Simulation data raises concerns about discrepancies with physical robots (the Sim-to-Real gap), while real data is costly to collect. As a result, the question of how to combine the two to improve model performance efficiently has become a common issue in physical AI development.Research and Development DetailsTo address the industry-wide lack of an established data recipe, FastLabel was selected for the “Physical AI Development Support Program by AWS Japan.” Over about six months of research and development, the company quantitatively verified the effectiveness of mixed training with real data and simulation data.1.ApproachThis initiative used the ugo Pro R&D model, a Japanese-made robot. FastLabel built a simulation environment on NVIDIA Isaac Sim in which the robot can perform motions equivalent to the physical unit, and established a data pipeline that efficiently generates and augments robot trajectory (motion) data. The company then built a dataset that combines the generated simulation data with real data collected from the physical robot, and verified the effectiveness and mixing ratio of simulation data by training models on AWS instances.[Image caption] Augmenting robot trajectory data2.Establishing the Simulation Data PipelineThe pipeline consists of the following four steps.① Creating trajectory dataAn environment in which the robot can perform motions equivalent to the physical unit was built on Isaac Sim, and source trajectory data for augmentation was created through teleoperation.② Trajectory augmentationMimicGen was applied to randomly placed objects, augmenting a small amount of trajectory data roughly 100-fold. From 90 source episodes, 10,000 episodes of trajectory data were obtained, excluding invalid motions and failures.③ Domain randomizationDomain randomization was applied to the augmented trajectory data on Isaac Sim. This produced a dataset that is larger in scale and also more varied (diverse) in environmental color variations such as lighting, backgrounds, and desks.④ Building a mixed datasetA mixed dataset was built by combining real data collected from the physical robot with simulation data generated through the process above.3.Verification Results① Improved task success rateTraining on a 19:1 mix of simulation data and real data raised the model's task success rate by an average of 14.2%. For example, in an accuracy comparison in the environment where the real data was collected, the task success rate rose from 26.7% with real data alone to 56.7% with mixed data.② Improved generalization in unseen environmentsIn environments different from those used during data collection (unseen conditions), the model trained on mixed data also outperformed the model trained on real data alone in robot motion accuracy, confirming improved generalization.③ Reduced data collection costsAn analysis of the relationship between dataset construction costs and success rates showed that the model trained on mixed data achieved low cost and high performance, while the model trained on real data alone had high cost and low performance. This shows that mixed training with simulation data also performs well in terms of cost-effectiveness.Details are available on our tech blog:https://fastlabel.hatenablog.com/entry/20260831/1788166800 Future OutlookGoing forward, FastLabel will support companies working on physical AI development in building simulation data pipelines, and will share the technical knowledge gained through this initiative.FastLabel will also continue to use cloud technologies, drawing on AWS Japan's expertise in physical AI, to enhance its data creation and management infrastructure. In doing so, the company will contribute to improved research efficiency in robotics AI development and accelerate its practical implementation in society. About FastLabelFastLabel Inc. is a data solutions company dedicated to supporting Data-Centric AI development by providing high-quality data and building data pipelines and data management infrastructure. Focusing on the field of Physical AI, including robotics, FastLabel supports a wide range of companies developing robot foundation models through services such as data collection, assistance in building simulation environments, model development support, and data pipeline construction.Company Name: FastLabel Inc.Headquarters: Shinjuku Sumitomo Building 24F, 2-6-1 Nishi-Shinjuku, Shinjuku-ku, Tokyo, JapanRepresentative: Takeshi Suzuki, President and CEOBusiness: Provider of professional services and products supporting Data-centric AI developmentWebsite: https://fastlabel.ai/service/robotics Media ContactCorporate CommunicationsFastLabel Inc.E-mail: pr@fastlabel.ai