While technological advances are unlocking previously unachievable applications, the true shift is not merely in what is now technically feasible, but in what is economically viable. As highlighted in Figure 2, the future of intelligent robotics is defined by simplified deployment and more intuitive human–machine interaction, enabling reductions in implementation time and greater scalability. As physical AI supports a wider variety of operations and becomes easier to deploy – requiring fewer specialized skills and less task- specific customization – automation becomes viable across a much broader range of operations. This evolution does not just enable new use cases, it redefines the overall economics of automation. 1.3 Limitations yet to be resolved Even as rapid advances continue, data scarcity, 3D spatial intelligence and dexterity remain challenges to be solved in the coming years: –Data scarcity: Large language models (LLMs) thrived because of the ability to scrape large amounts of data from the internet. These LLMs have now read many websites and ingested many books. Physical AI also thrives on high- quality data, yet curated robotics datasets remain limited and costly, because they have to be collected in the real world. This challenge is rapidly being overcome through advances in synthetic data generation. Photorealistic rendering and domain randomization help to simulate varying lighting, textures and object shapes in virtual environments to teach robots how to grasp items in diverse real-world conditions. Combined with open-source efforts and the accelerated deployment of real-world robotic fleets, these developments are set to close the data gap and dramatically enhance learning efficiency. For example, as a first step, robotic companies such as Sanctuary AI10 started with teleoperation – having an operator control one or more robots – while collecting data. The goal is to use this data to train the robots to operate autonomously at a later point. Developers can use technologies provided by Nvidia and other companies to produce numerous plausible futures or variations based on real or synthetic data, serving as an automatic data multiplier grounded in real physics. –3D spatial intelligence: Data scarcity is chief among the reasons that perceiving, reasoning within and interacting with complex 3D environments remains a demanding task. However, progress is accelerating. Start-ups such as World Labs and Covariant, as well as academic leaders, are using simulation, real- world data and multimodal AI architectures to enable robust spatial understanding. Vision– language–action (VLA) models are rapidly evolving as a promising path, with foundation models poised to unlock generalizable spatial reasoning capabilities. –Generalizable dexterity with high degrees of freedom: Achieving human-level dexterity remains a frontier challenge due to the mechanical, sensory and computational constraints. For example, robotic hands must operate with high degrees of freedom (DoF) – often more than 20 joints – which makes real- time motion planning, force control and collision avoidance very complex. Crucially, progress in 3D spatial intelligence is an enabler of dexterity: fine manipulation depends on precise perception of object geometry, pose and occlusions. Foundation models that integrate 3D scene understanding with manipulation planning will enable robots to better select stable grasps, adapt to object variability and execute corrective strategies when conditions change. While advances in robotics continue and technological breakthroughs keep pushing boundaries, addressing emerging challenges is crucial for sustainable adoption. Key areas requiring attention include cybersecurity. Vulnerability to cyberthreats increases as factories and warehouses become increasingly software-defined and robotic systems become more interconnected. Ensuring robust cybersecurity measures is essential to protect against potential disruptions and data breaches. Vision–language– action (VLA) models are rapidly evolving as a promising path, with foundation models poised to unlock generalizable spatial reasoning capabilities. Physical AI: Powering the New Age of Industrial Operations 10

Physical AI Powering the New Age of Industrial Operations 2025