Panasonic HD Develops a World Model for Mobile Work Assistance

Osaka, Japan, October 21, 2026 – Panasonic Holdings Corporation (Panasonic HD) and Panasonic R&D Company of America (PRDCA), in collaboration with researchers at the University of California, Los Angeles (UCLA), have developed a world model ⁽*¹⁾ that enables AI to anticipate and assist with tasks on mobile devices such as smartphones. The team also created MobileWorld, a training dataset containing approximately 1.4 million examples for training world models, and MobileWorldBench, a benchmark for evaluating their performance. Using these resources, the team successfully improved the success rate of AI agents operating mobile devices. This research will be presented at EMNLP 2026, a leading conference in natural language processing, to be held in Budapest, Hungary, from October 24 to 29, 2026.

With advances in generative AI, AI agents are expected to assist with or perform operations on smartphones and tablets on behalf of people at manufacturing, logistics, maintenance, and other worksites. However, conventional world-model approaches that predict screens as images must also reproduce information unrelated to the task, such as the time, notifications, and screen layout, which can make prediction difficult. This has created a need for a new kind of world model that predicts important state changes caused by an operation, rather than predicting the screen pixel by pixel, and uses those predictions to select the next action.

Rather than reproducing the post-operation screen as an image, this technology predicts operation-related changes in natural language, such as “pressing the button opens the settings screen” or “an input field appears.” As a result, without being affected by information unrelated to the operation—such as the time, notifications, or screen layout—the AI agent can evaluate predicted state changes for multiple candidate actions and select the action most likely to achieve the goal.

MobileWorld, developed as part of this research, is a dataset based on human interaction trajectories on Android devices. It includes operation details, screens before and after each operation, descriptions of screen changes, and questions and answers about subsequent states. The dataset contains approximately 942,000 examples describing screen changes and approximately 543,000 question-and-answer examples. A vision-language model was used to generate and automatically filter the data, with a subset also verified manually. Qwen3-VL-8B-Instruct ⁽*²⁾, further trained using MobileWorld, was evaluated on MobileWorldBench. When its predicted text was scored for accuracy, completeness, and relevance, it achieved the highest score among the methods compared. The world model was then integrated into an AI agent and evaluated on AndroidWorld ⁽*³⁾. In this evaluation, the world model predicts the future state for each candidate action, and the AI agent uses those predictions to select the action it will actually execute. The task success rate improved substantially from 46.9% for the baseline without the world model to 54.3% with the world model integrated.

By applying this technology to work that spans multiple applications—including equipment management, parts inventory, and work records—we aim to reduce workloads at worksites through AI agents that leverage the knowledge and judgment of experienced personnel. Going forward, in addition to smartphones, we will examine its applicability to websites and PC applications and advance research and development toward AI technology that supports a broad range of internal operations.

Title: MobileWorldBench: Towards Semantic World Modeling For Mobile Agents
arXiv: https://arxiv.org/abs/2512.14014
This research was conducted jointly by Konstantinos Kallidromitis and Yusuke Kato of PRDCA, Kazuki Kozuka of Panasonic HD, Shufan Li of UCLA, and their colleagues.

Notes

*1 A model that learns the state of an environment and the mechanisms by which it changes, and predicts subsequent states based on the current situation and actions.

*2 An open-source vision-language model with approximately 8 billion parameters that can understand images and text and generate text based on image content.

*3 A benchmark that has AI agents perform tasks on Android devices and evaluates their task success rate.

• Panasonic×AI Website
  https://tech-ai.panasonic.com/en/