The artificial intelligence industry has spent the last few years perfecting language and code, but a quieter revolution is happening in parallel: AI systems that understand and interact with the physical world. Rather than learning from text and images on the internet, a new generation of models is being trained on video, robot trajectories, and even wearable sensor data from humans performing everyday tasks. This shift represents a fundamental change in how AI systems learn to perceive and act in three-dimensional space .

The dominant paradigm in AI today centers on large language models (LLMs), which are AI systems trained on massive amounts of text to predict and generate language. These models have clear scaling laws, meaning researchers understand how performance improves as models get larger and are trained on more data. But a set of adjacent fields is maturing rapidly alongside this progress, including vision-language-action models (VLAs), world action models (WAMs), and other approaches to robotics, autonomous science, and novel human-computer interfaces like brain-computer interfaces (BCIs) .

What Are the Technical Building Blocks Behind Physical AI?

Five core technical primitives are enabling AI to extend into the physical world. These aren't specific to any single application; rather, they're foundational technologies that multiple domains are building upon simultaneously .

• Compressed representations of physics: Models that learn how objects move, deform, collide, and respond to force, allowing systems to transfer knowledge across different physical tasks rather than learning from scratch each time.
• Action architectures: Systems that translate high-level intent into continuous motor commands, maintain coherence over long sequences of actions, and operate within real-time latency constraints.
• Spatial intelligence: Models that reconstruct and reason about the full three-dimensional structure of physical environments, including geometry, lighting, occlusion, and object relationships.
• Simulation and synthetic data infrastructure: Tools that generate training data through simulation, reducing the need for expensive real-world data collection.
• Closed-loop agentic orchestration: Systems that can perceive their environment, take action, observe the results, and improve iteratively.

The most fundamental of these is learning compressed, general-purpose representations of how the physical world behaves. Without this capability, every physical AI system must learn the physics of its domain from scratch, which is prohibitively expensive .

How Are Different AI Approaches Learning Physical Understanding?

Multiple architectural families are converging on the ability to understand physics from different starting points. Vision-language-action models (VLAs) take pretrained vision-language models, which already understand objects and spatial relationships from internet-scale image-text pretraining, and extend them with action decoders that output motor commands. Models like Pi-Zero from Physical Intelligence, Google DeepMind's Gemini Robotics, and NVIDIA's GR00T N1 have demonstrated this architecture at increasing scale .

World action models (WAMs) take a different path, building on video diffusion transformers pretrained on internet-scale video. These models inherit rich priors about physical dynamics, such as how objects fall and interact under force, and couple these priors with action generation. NVIDIA's DreamZero demonstrates zero-shot generalization to entirely new tasks and environments, achieving meaningful improvement in real-world generalization while enabling cross-embodiment transfer from human video demonstrations with only small amounts of adaptation data .

A third approach, which may be the most instructive for understanding where this field is heading, dispenses with both pretrained vision-language models and video diffusion backbones entirely. Generalist's GEN-1 is a native embodied foundation model trained from scratch on over half a million hours of real-world physical interaction data, collected primarily through low-cost wearable devices on humans performing everyday manipulation tasks. It is not a VLA in the standard sense, nor is it a WAM. Instead, it is a first-class foundation model for physical interaction, designed from the ground up to learn representations of dynamics from the statistics of human-object contact rather than from internet images, text, or video .

Why Does This Matter for AI's Future?

The convergence of these approaches reveals something important: whether representations are inherited from vision-language models, learned through video co-training, or built natively from physical interaction data, the underlying primitive is the same. Compressed, transferable models of how the physical world behaves can serve a robot learning to fold towels, a self-driving laboratory predicting reaction outcomes, and a neural decoder interpreting the motor cortex's plan for grasping .

The data flywheel for these representations is enormous and largely untapped. It encompasses not just internet video and robot trajectories, but the vast corpus of human physical experience that wearable devices are now beginning to capture at scale. This represents a qualitatively different data source than the text and images that have driven language model scaling .

Three domains specifically represent frontier opportunities for this technology: robot learning, autonomous science in materials and life sciences, and new human-machine interfaces including brain-computer interfaces, silent speech recognition, neural wearables, and novel sensory modalities like digitized olfaction. These areas are not entirely separate efforts; they share common technical substrates and are mutually reinforcing in ways that create compounding dynamics across domains .

The technical primitives for extending frontier AI into the physical world are maturing concurrently, and the pace of progress over the past eighteen months suggests that these fields could soon enter a scaling regime of their own. In technology paradigms, the areas with the greatest potential tend to be those that benefit from the same scaling dynamics driving the current frontier but sit one step removed from the incumbent paradigm. This distance creates a natural moat against fast-following and defines a problem space that is richer, less explored, and more likely to produce emergent capabilities precisely because the easy paths have not already been taken .

As these systems mature, they could unlock capabilities that language-only AI cannot achieve. A robot that understands physics can learn from human demonstration videos. An autonomous scientist that models physical dynamics can predict reaction outcomes. A brain-computer interface that understands motor intent can decode neural signals more accurately. The convergence of these capabilities suggests that the next major phase of AI advancement may not come from making language models larger, but from grounding AI systems in the physical world itself.