The gap between Japanese robotics and its Chinese and American rivals does not come from hardware on the robot. Kazumasa Hirano, who heads robotics at NVIDIA Japan, points to model training. The interview precedes his EdgeTech+ 2026 keynote.
Key takeaways
- NVIDIA Jetson AGX Thor uses the Blackwell architecture and supports FP4
- Two modules: Jetson T5000 and T4000, with the dev kit emulating T4000
- A classic AMR can run largely on CPU, a humanoid needs GPU
- Hirano names training, not edge, as the Chinese and US advantage
- Keynote on November 20 at 11:20, EdgeTech+ 2026
Why a robot cannot rely on the cloud
There are two reasons. The first is latency — a robot must react on the spot, not wait for the network.
The second is security. In manufacturing, logistics and healthcare, shop-floor data should not leave the site.
AMR versus humanoid: an order-of-magnitude shift
Hirano previously ran mobile robot projects in logistics, and his comparison is blunt. A classic AMR handled a limited stream, and navigation rested on SLAM. A humanoid has to run locomotion, manipulation and a decision layer in parallel.
| Dimension | Classic AMR | Humanoid |
|---|---|---|
| Input stream | camera, LiDAR, host-system signals | many cameras and sensors, plus audio |
| Motion | SLAM-based navigation | locomotion + manipulation |
| Decision layer | route planner, rules | VLA, world models, agentic models |
| Dominant compute | CPU | GPU |
Thor does not replace Orin
Hirano warns against treating Thor as the default pick. Required compute depends on model size and concurrent processes. You can use several Jetson AGX Orin modules, or pair an Orin with a Thor.
The glue is the shared software platform across the Jetson line, so changing processor does not force a rebuild.
Safety spreads across three computers
The on-board unit alone is not enough when a robot has to react to the unexpected. Behaviour is first evaluated in simulation, then deployed to the edge, with irregular field data fed back into training.
Where the difference actually sits
The overwhelming difference is training. Not edge — training.
Kazumasa Hirano, head of the robotics business unit, NVIDIA Japan
It is about the scale of investment and access to training compute, but also about narrowing a task and fine-tuning a model until it is usable. Hirano expects the differentiator to stop being the model itself and become data quality.
Why it matters
It explains the market's shift towards teleoperation and simulation. Both produce data rather than just robot motion — and both cost less than buying the next generation of compute module.
What's next
- Hirano's keynote at EdgeTech+ 2026 is scheduled for November 20 at 11:20
- The Jetson AGX Thor developer kit lets teams target the T4000 module without owning one
Sources
- ロボスタ — NVIDIA 平野氏に聞く「フィジカルAI」で変わるエッジコンピューティング Jetson Thorが拓くロボティクスの未来【EdgeTech+ 2026】
- NVIDIA Developer — NVIDIA Jetson AGX Thor
