Most artificial intelligence models trained on video simply show how the world looks. The new PAN model, created at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), takes a different approach — it tries to understand how the world actually works.
When PAN was first revealed, much of the public reaction focused on its ability to generate realistic videos. But that isn’t its true purpose. PAN was not designed to become another creative tool — it was built to develop reasoning skills for robots and autonomous systems.
Jon Carvill, Vice President of Marketing and Communications at MBZUAI, explained that difference clearly: unlike traditional video models that focus on generating visuals from text instructions, PAN is engineered to learn and simulate real-world logic. He put it simply: “Video models copy appearance — PAN understands the forces and dynamics that make that appearance possible.”
While models like OpenAI’s Sora or Google’s Veo excel at cinematic imagery, PAN aims to equip machines with an ability humans take for granted: thinking logically about actions, outcomes, and physical interactions.
Solving a Major Problem in Robotics
Training physical robots in real environments is expensive, time-consuming, and risky. Humanoid robots like Tesla’s Optimus or Figure AI’s Helix require enormous human supervision, and every physical mistake during training could damage or destroy hardware worth hundreds of thousands of dollars.
Today’s robotics teams often rely on hundreds of human operators demonstrating tasks thousands of times — all just to train robots on a handful of basic skills. Development cycles can stretch into years.
PAN is designed to break that barrier.
Rather than learning through real-world trial and error, robots can learn inside PAN’s advanced world simulation model, where they can test actions repeatedly — infinitely — without physical consequences.
A robotic hand inside PAN, for example, can practice hundreds of different ways to pick up a cup before ever interacting with one in reality.
How PAN Works
Inside this model, robots can simulate countless real-life interactions — from autonomous cars navigating traffic to home robots loading a washing machine — refining decisions before executing anything physically.
PAN’s architecture enables this by generating scenes step-by-step rather than all at once. It keeps track of what objects exist, how they move, and what physical rules apply to them with each new frame it predicts.
Carvill explained that this blend of systems is key: visual quality comes from diffusion modeling, while a large language model maintains logical understanding, memory, and consequences over longer periods.
The potential is enormous — physics-based simulation can accelerate robotic learning by over 430,000 times compared to real-world training. Tasks that normally require decades of trial and error could be mastered within hours of computation.
Toward a New Era of Embodied Intelligence
PAN represents an emerging category known as embodied AI, where artificial intelligence must understand not only data and language — but real-world consequences.
Modern language models are powerful, but they lack awareness of the physical world. PAN bridges that gap.
Carvill summed it up: “To build intelligent agents, AI must move beyond text. It must understand how actions shape the world.”
PAN’s development was made possible through coordinated research teams operating across Abu Dhabi and Silicon Valley — combining expertise, time zones, and large-scale infrastructure.
A Vision for the Future
PAN aligns with MBZUAI’s larger strategy in building foundational AI systems — including recent breakthroughs like the reasoning-focused system K2 Think.
Looking ahead, MBZUAI sees a future where systems like PAN become the core framework for intelligent robotics, autonomous vehicles, future smart cities, virtual simulations, and next-generation digital environments.
By 2030, PAN-level world modeling could become the standard infrastructure for robots capable of safely navigating and interacting with the real world.
For Abu Dhabi, this achievement strengthens its role at the intersection of advanced AI and robotics — supported by innovation programs, global talent, and ambitious national technology strategies.
As Carvill noted, the UAE’s ecosystem provides a unique advantage: a place where research, government, and industry align in a way rarely seen elsewhere — creating fertile ground for breakthroughs that could define the future of intelligent machines.
Saad Farooq brings a blend of engineering prowess and a passion for writing to technewsme.com. As a tech engineer based in the vibrant tech landscape of the UAE, Saad has immersed himself in the latest technological advancements, making him an invaluable asset to our team. His deep understanding of tech intricacies comes from hands-on experience in the field, coupled with a genuine love for all things tech.
In his role at technewsme.com, Saad specializes in demystifying complex tech subjects, making them accessible and engaging for our readers. His articles are a reflection of his life in the UAE, a hub of technological innovation, offering a unique perspective on global tech trends and local developments. Whether it’s exploring new tech gadgets, delving into software innovations, or analyzing tech market trends, Saad’s writing is fueled by his extensive engineering background and his enthusiasm for sharing knowledge. Join him on a journey of discovery and insight into the world of technology, right from the heart of the UAE.
