1 TRILLION YEN TO TEACH ROBOTS TO THINK: JAPAN'S BIG BET

In this talk

1. 0:00

   Kaleido’s role and gradual rollout

   Kawasaki aims to start using Kaleido around 2030 for patrols and moving parts in factories and warehouses, with assembly work considered later. The company is discussing possible deployments with automotive and semiconductor firms, and the broader plan includes tens of thousands of robots across several designs by 2040. Early operation will progress from remote human control to shared tasks and, eventually, greater autonomy.

2. 1:33

   A robot built to withstand failure

   Kaleido’s early design focused on durability: Kawasaki wanted a humanoid that could fall without breaking, so researchers could spend less time repairing machines. At its 2017 debut, the robot stood up and did pull-ups while tethered to power. By 2025, a later model could navigate a rescue demonstration and retrieve a stuffed cat, with an operator guiding its head and hands remotely.

3. 5:12

   Japan’s investment in physical AI

   Physical AI must sense and act in real surroundings, where errors can damage equipment or injure people. Japan has committed up to 1 trillion yen over five years to AI development, with 387 billion yen in the first tranche and later funding subject to review. Nvidia is supplying computing capacity for a data center planned to begin operations in 2028, while the program targets reasoning models first, then systems that combine text, images, video, and audio.

4. 7:48

   Training and current hardware limits

   Kawasaki plans to refine robot movements through reinforcement learning in simulated environments, then work with partners including Nvidia, Analog Devices, Microsoft, and Fujitsu at a physical AI center in San Jose. Kaleido 9 has 30 movement axes and uses LiDAR mapping, stereo cameras, and balance adjustments to navigate. Demonstrations included sweeping and moving a heavy shelf, but Kawasaki has not confirmed key operating specifications such as maximum load, speed, or battery life.

5. 11:37

   Why 2030 still needs people

   The roadmap places real-world-native AI, which is intended to understand objects and physical properties, in fiscal 2030, the same period Kawasaki targets for factory work. That timing makes remote supervision central to early deployments, rather than a temporary detail. The robot’s battery endurance, fine hand control, touch sensing, and environmental resistance also remain limitations, with some capabilities scheduled for development well beyond 2030.