AIToday
RoboticsLarge Language ModelsITmedia AI+Published: Sep 4, 2026, 13:01 JST2 min read

AWS Japan's physical AI program shows data shortage solutions

AWS Japan's physical AI program shows data shortage solutions

Key takeaway

  • AWS Japan's physical AI program helped 51 companies tackle data scarcity.

  • Firms used simulation, lighter models, and shared data.

  • The challenge now is making products that truly work on site.

3 Key Points

  1. What happened

    AWS Japan's physical AI development support program ended with 51 companies presenting results, all sharing a common struggle: a lack of data for training robot control models.

  2. Why it matters

    Unlike LLMs that learn from web text, robot foundation models need camera footage, joint movements, and sometimes touch data. Companies without real-world sites struggle to collect enough data, making the wall harder to climb.

  3. What to watch

    The program ends but support continues under AWS Japan's existing generative AI program. ZEALS plans to mass-produce 100 units of its humanoid D1 and achieve 10,000 hours of on-site operation in fiscal 2026.

Ask the AI about this article →

Context & Analysis

The AWS Japan support program brought together an unusual mix of participants, from startups under one year old to a general contractor with over 400 years of history. Their common complaint was a lack of data, but their approaches varied. Some, like FastLabel and Takenaka, chose to generate data through simulation. Others, like Telexistence and Ricoh, tried to reduce the amount of data needed or make models less dependent on it.

Japan's position in physical AI is distinct. ZEALS' CEO noted that the US leads in AI models and China in hardware mass production, but Japan has real-world sites with labor shortages that need robots. This insight aligns with the program's focus: turning real-world challenges into data sources. However, as the program ends, the larger test begins: whether these prototypes can become reliable products and scale into sustainable businesses. The next phase under AWS Japan's existing generative AI program will reveal if that momentum continues.

FAQ

How did companies try to solve the data shortage?
Some generated data through simulation, like FastLabel expanding 90 action patterns to 10,000. Others reduced data needs, such as Telexistence using a lightweight model and Mameno limiting joints to 10.
What results were announced?
FastLabel improved task success by 14.2% with simulated data. Telexistence achieved 86% success in convenience store tasks. ZEALS' D1 completed a 3-day trial at a hospital.

Get the latest Robotics news every morning

For example, today's edition would include:

  • Nvidia exec: AI reshaping chip designDIGITIMES Asia · 5h ago
  • Lockheed Martin expands venture investments in AI and quantum techTop Companies AI · 18h ago
  • NVIDIA announces entry-level AI robotics computerTop Companies AI · 18h ago

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. Q&As are published on this page for other readers too.

Related Articles

Next articlePalantir Stock Surges on Expanded PwC AI Deal