en.Wedoany.com Reported - Japan's SoftBank, in collaboration with Yaskawa Electric and US-based NVIDIA, has demonstrated a deformable object manipulation system based on the "AI Data Center GPU Cloud" currently under development by SoftBank. SoftBank has also developed physical AI development support tools running on this GPU cloud, covering robot motion data collection, synthetic data generation, AI model training, simulation evaluation, and physical robot deployment, aiming to integrate the traditionally fragmented robot development processes into a unified platform.
This platform connects to NVIDIA's Omniverse library, the Cosmos open-world model, and the physical AI factory blueprint. Robots first connect to SoftBank's "AI Data Center GPU Cloud" to collect motion data, then use Cosmos to generate synthetic data to supplement the training set. Engineers can leverage GPU cloud computing resources to train models and conduct performance evaluations in a simulation environment integrated with the Omniverse library. Validated models are subsequently deployed to physical robots. SoftBank plans to use this platform to support enterprises in continuously improving models using field data and extending development outcomes to other operational tasks and robotic equipment.
Yaskawa Electric is responsible for developing a deformable object manipulation system based on vision-language-action technology, enabling robots to generate motion plans based on visual information and task instructions to perform recognition, grasping, and manipulation. This demonstration selected wire harnesses with constantly changing shapes and arrangements as test objects. The robot assesses the wire harness's state through its vision system, and the AI generates real-time action commands to stably place the wire harness into a box. Related validations also cover objects such as cables and fabrics, whose shapes change with tasks, primarily targeting operational scenarios that are difficult to handle stably with traditional pre-programmed robot control.
