SoftBank and Yaskawa Advance Physical AI Robotics on Cloud Infrastructure
Industrial robotics is achieving a major capability breakthrough as cloud-based artificial intelligence connects directly with real-time physical manipulation. Telecommunications pioneer SoftBank Corp. and industrial robotics manufacturer Yaskawa Electric Corporation, collaborating alongside NVIDIA, have successfully demonstrated a specialized robotic system capable of handling highly deformable, flexible objects. Powered by SoftBank’s "AI Data Center GPU Cloud" platform, the project creates a streamlined pipeline connecting real-world movement data capture, cloud-based model training, synthetic simulation, and live execution on shop-floor industrial manipulators.
Historically, conventional industrial motion control has relied on deterministic trajectory planning, assuming that target workpieces maintain static geometries and fixed locations. Handling flexible materials such as wiring harnesses, fabric panels, or industrial cables presents severe control difficulties due to unpredictable sagging, bending, and tangling. To bridge this gap, Yaskawa developed a specialized manipulation setup leveraging Vision-Language-Action (VLA) architecture. By decoupling deterministic mechanical movements from real-time spatial evaluation, the system isolates high-level perception and grasp-point selection into a modular AI unit, ensuring stable execution without re-engineering core motion routines.
SoftBank’s underlying development environment streamlines this loop by integrating NVIDIA Omniverse simulation libraries alongside NVIDIA Cosmos open world frameworks. Motion data and vision streams collected from physical robots are uploaded to the GPU cloud, where synthetic data augmentation generates millions of plausible edge-case scenarios. AI models train directly within this cloud infrastructure, allowing controls engineers to monitor optimization through a centralized graphical interface before evaluating trajectory safety inside virtual environments. Once validated, perfected models deploy straight to field hardware. During live validation tests, the setup successfully recognized, routed, and boxed shifting industrial wire harnesses in real time, proving that cloud-connected Physical AI can drastically shorten deployment cycles for complex assembly operations across global manufacturing networks.