Data Scarcity Challenges in Physical AI Development and Corporate Solutions

This program is an initiative to support companies and organizations developing robot foundation models and similar technologies. Following support including the provision of approximately ¥900 million worth of AWS credits, 51 companies presented their results.

Many companies cited "data scarcity" as a major challenge. Unlike Large Language Models, models for robot control require data such as camera footage, joint movements, and tactile sensations.

As a countermeasure to data scarcity, some companies are working on generating data through simulation. FastLabel stated that they expanded teleoperated motion data into 10,000 patterns within a simulation, improving task success rates by combining this with real-world machine data.

Additionally, cases of innovative data collection methods were reported, such as Takenaka Corporation, which built an environment to collect data from remote locations using Apple Vision Pro, and Mercari, which collects tactile data using specialized equipment.

Reports regarding model lightweighting and efficiency were also presented. Telexistence reported that they developed a lightweight 10B parameter model based on a video generation model, achieving a high success rate. Mamezo stated that they achieved full-body control even with limited motion data by narrowing down the number of joints to be controlled.

AWS Japan said it will integrate and continue this support through its "Generative AI Practical Application Promotion Program."