SoftBank Corp. and Yaskawa Electric Corporation Demonstrate a Deformable Object Manipulation System Using "AI Data Center GPU Cloud" as
a Physical AI Development Platform

Background

To implement Physical AI in society, it is important to establish a development platform that streamlines complex development processes, including the collection of robot motion data, AI model training, simulation-based evaluation, and deployment on physical robots, enabling not only robotics experts but also a wide range of engineers to quickly and easily introduce Physical AI into real-world operations.

To this end, SoftBank has developed a Physical AI development support tool that runs on the "AI Data Center GPU Cloud." By utilizing NVIDIA Omniverse libraries, which enable the creation of robot simulation environments, and NVIDIA Cosmos open world models and the NVIDIA Physical AI Factory Blueprint, this tool can streamline the series of development processes required for the introduction of Physical AI.

Conventional robot control assumes that the shape and position of the target object are predetermined, making stable handling of deformable objects such as strings, cloth, and bags, which vary from task to task, one of the more challenging areas. To address this challenge, Yaskawa Electric has developed a deformable object manipulation system that utilizes VLA (Vision-Language-Action), which generates robot motions based on visual information and task instructions, to recognize, grasp, and manipulate such deformable objects.

With the aim of putting the development support tool and the deformable object manipulation system into practical use, SoftBank and Yaskawa Electric conducted a demonstration of wire harness handling, a task that requires the advanced decision-making capabilities of Physical AI.

Physical AI Development Support Tool Running on the "AI Data Center GPU Cloud"

The Physical AI development support tool developed by SoftBank enables development processes to be conducted in an integrated manner on the "AI Data Center GPU Cloud," from the collection of robot motion data and data augmentation (synthetic data generation) to AI model training, simulation-based evaluation, and deployment on physical robots.

By using this tool, companies can efficiently establish a Physical AI development environment and advance the learning cycle on the "AI Data Center GPU Cloud." In addition, by making it easier to introduce Physical AI into worksites, companies can utilize data obtained from physical robots, AI model execution data, and synthetic data collected across multiple operations and sites as assets in accordance with their own management policies, and expand their use to other operations.

Deformable Object Manipulation System

The "deformable object manipulation System" developed by Yaskawa Electric is a system that uses VLA to recognize the state of target objects based on visual information obtained from cameras and other sources, as well as task instructions, and enables robots to grasp and manipulate those objects.

Deformable objects such as strings, cloth, and bags are objects whose shapes and positions change with each task, and are prone to sagging, tangling, bending, and variations in grasping positions. As a result, stable handling of such objects is difficult using conventional rule-based control alone, which assumes that shape and position of the target object are predetermined.

Accordingly, the deformable object handling system is configured to separate the motions that can be handled reliably with conventional robot control from the motions for which Physical AI is effective, such as recognizing the state of deformable objects and making grasping position decisions, and to incorporate Physical AI into existing robot systems as a functional module.

Demonstration Overview

In this demonstration, as a use case for manipulating deformable objects, SoftBank's Physical AI development support tool and Yaskawa Electric's deformable object manipulation system were applied to the task of placing wire harnesses, whose shapes and arrangements change with each task, into a box. Through this demonstration, the companies confirmed not only that Physical AI could be introduced quickly, but also that deformable objects could be handled stably, with the robot recognizing the state of wire harnesses based on visual information and AI learning in real time to provide accurate instructions to the robot.

Future Outlook

SoftBank and Yaskawa Electric will continue to promote the development of robot systems utilizing Physical AI and verify their application to more advanced tasks. SoftBank will also expand its verification of Physical AI using the "GPU Cloud," work to advance robot development by utilizing various types of data, and further enhance its development platform.