Yaskawa Electric and SoftBank Corp. validated a vision-language-action driven system on Yaskawa's MOTOMAN NEXT AI robot, announced July 13, 2026. The task: picking wire harnesses of varying shape and position and packing them into boxes.
Wire harnesses are the canonical failure case for industrial robotics. Conventional teach-and-playback control assumes an object is where it was last time and shaped the way it was last time; a harness is neither. That is why harness assembly remains one of the most stubbornly manual operations in automotive and electronics manufacturing.
Physical AI as a module, not a replacement
The notable design choice is the split. Yaskawa divided the task between conventional control and physical AI, using the model for state recognition and grasp-point selection while leaving the rest of the motion stack alone. Physical AI was integrated as a module into the existing robot system rather than replacing it.
That is the opposite of the end-to-end framing most VLA work adopts, and it is the version a factory can actually buy — existing safety cases, existing cycle times, existing controllers, with learning applied only where deterministic control fails.
The system ran on SoftBank's AI Data Center GPU Cloud, announced May 25, 2026 and powered by the Infrinia AI Cloud OS, used here as a physical AI development platform ahead of its commercial launch in October 2026. Both companies report sharply improved efficiency across the data collection, training, evaluation and redeployment loop, with office environments named as the first target use case.
Key Facts
- Announced July 13, 2026 by Yaskawa Electric (Kitakyushu) and SoftBank Corp.
- Task: picking and packing wire harnesses — deformable objects that defeat teach-and-playback control
- Runs on Yaskawa's MOTOMAN NEXT AI robot with a vision-language-action model
- Physical AI integrated as a module for state recognition and grasp-point selection, not as a replacement for conventional control
- Executed on SoftBank's AI Data Center GPU Cloud (announced May 25, 2026; Infrinia AI Cloud OS; commercial launch October 2026)
Why it matters
Deformable-object manipulation is the honest frontier of industrial robotics. Rigid parts were solved decades ago; anything that bends, folds or tangles has stayed with human hands. A credible harness result from an incumbent industrial robot maker is a bigger commercial signal than a better benchmark score from a lab.
The modular architecture is the transferable idea. Most factories will not replace a working motion stack to adopt learning, but they will drop a perception module into one. That is a far shorter path to deployment than the end-to-end pitch, and it lets the model be wrong in a bounded way.
Frequently Asked
What was the task?
Picking and boxing wire harnesses — deformable objects conventional control cannot handle.
What is architecturally interesting?
Physical AI was added as a perception module inside an existing industrial stack, not as an end-to-end replacement.
Where did it run?
SoftBank's AI Data Center GPU Cloud, which launches commercially in October 2026.