Physical AI systems need something language models don't: massive amounts of real-world robot interaction data to learn how to act safely in physical environments. Unlike large language models (LLMs) trained on text scraped from the internet, embodied AI (AI systems that perceive and act in real environments through sensors and motors) requires action-labeled data from actual robot operations, teleoperation logs, and fleet records. This fundamental difference is reshaping how researchers, educators, and companies approach AI development.

The challenge is becoming urgent as physical AI moves from research labs into real-world applications. Major technology companies including Google DeepMind, Nvidia, and Tencent are investing heavily in embodied intelligence, while universities and startups are racing to build platforms that can support this emerging field. Yet the bottleneck remains: generating enough high-quality training data to teach robots to perform complex tasks reliably and safely.

What Exactly Is Embodied AI, and Why Does It Need Different Training Data?

Embodied AI represents a fundamental shift from virtual AI systems. While a language model predicts the next word in a sentence, embodied AI systems perceive the world through cameras and sensors, reason about what they observe, and then act through motors and actuators. A robot clearing a kitchen, for example, can now draw on an LLM to infer that a mug is fragile, should be grasped by the handle, and probably belongs in a cupboard, reasoning that once required hand-coding for every object and room.

"Most AI today is virtual, you write a prompt and it gives back text. Embodied, or physical, AI is different, systems perceive the world through sensors such as cameras, and act on it through motors and actuators," explained Dr. Stephen Smith, a professor in the Department of Electrical and Computer Engineering and Director of Waterloo.AI at the University of Waterloo.

Dr. Stephen Smith, Professor of Electrical and Computer Engineering and Director of Waterloo.AI, University of Waterloo

This capability depends on a type of data that cannot be scraped from the internet. Robot trajectories, teleoperation records, and fleet operation logs represent the foundation of embodied AI training. Unlike text data, which is abundant and freely available online, this physical interaction data must be generated by actually operating robots in real-world or carefully controlled environments.

Why Is the Data Bottleneck Creating a Concentration Problem?

The requirement for real-world robot interaction data creates enormous barriers to entry for smaller competitors, startups, universities, and public institutions. Only well-capitalized firms with the resources to deploy physical systems at scale can generate the massive datasets needed to train competitive embodied AI systems. This concentration of capability raises concerns about market dominance and public access to critical infrastructure.

"World models depend on a type of data that's fundamentally different from what trained language models, action-labeled interaction data. Robot trajectories, fleet logs, teleoperation records. This data can't be scraped from the internet. You have to generate it by operating physical machines in real-world environments. That creates enormous barriers to entry and gives advantages to a small number of well-capitalized firms that can deploy systems at scale," noted Fei-Fei Li, Founding Director of Stanford's Human-Centered AI Institute.

Fei-Fei Li, Founding Director, Stanford Institute for Human-Centered AI

The implications extend beyond commercial competition. If governments depend on proprietary simulations for training autonomous systems or planning infrastructure but cannot inspect or replace those systems, critical public capabilities become dependent on private infrastructure controlled by a handful of firms. This dependency creates vulnerabilities in national security, emergency response, and infrastructure planning.

How Are Universities and Philanthropic Organizations Addressing the Gap?

Recognizing the importance of embodied AI research, educational institutions and philanthropic organizations are beginning to invest directly in graduate-level research and platform development. The University of Waterloo received a $500,000 investment from the Tang Family Foundation to support graduate research in embodied AI through student scholarships. Eight inaugural scholarship recipients from the faculties of Engineering and Math each received $15,000 to focus on embodied intelligence research.

These investments are enabling students to work on foundational problems in embodied AI across multiple disciplines. Graduate researchers are exploring exoskeletons for mobility assistance, robot manipulation systems, autonomous navigation, and the integration of AI reasoning with safe physical action. The interdisciplinary nature of embodied AI requires expertise spanning foundational AI, control systems, mechatronics, and the social and ethical questions of how such systems operate around people.

Steps Researchers Are Taking to Advance Embodied AI Development

• Building Open Platforms: Companies like Elephant Robotics are developing accessible robotic platforms featuring open hardware, intelligent interaction capabilities, and flexible development ecosystems that enable researchers and educators to explore embodied AI applications without massive capital investment.
• Integrating Large Language Models with Robotics: Researchers are combining LLMs and Vision Language Models (VLMs) with robotic systems to enable natural human-robot interaction and more sophisticated reasoning about physical tasks and environments.
• Creating Compound Robot Systems: Advanced mobile manipulation platforms that combine mobility and manipulation capabilities, equipped with LiDAR sensors and camera systems, enable researchers to explore autonomous navigation, perception-driven manipulation, and AI-powered robotic applications at scale.

At the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) scheduled for September 28 through 30 in Pittsburgh, companies and researchers will showcase how open, accessible robotics technologies can empower innovation in embodied AI. Elephant Robotics plans to present platforms including the Mercury B1 semi-humanoid robot with 17 degrees of freedom, the myArm M750 6-degree-of-freedom robotic arm, and compound mobile manipulation systems designed for research and commercial applications.

What Policy Challenges Does Embodied AI Create?

As embodied AI systems move into real-world deployment, policymakers face governance challenges that are fundamentally different from regulating language models. While language models primarily create information risks like misinformation and bias, embodied AI systems create physical risks. When an AI system's understanding of the physical world guides a robot, drives a vehicle, or informs emergency response decisions, errors don't just misinform, they can injure people, destroy property, or cost lives.

Stanford researchers have identified three functional categories of world models, AI systems that build working representations of physical environments to predict how they change in response to action. Renderers generate realistic images or video for design and training. Simulators model underlying physics and dynamics for engineering applications. Planners determine what action an agent should take in autonomous vehicles and disaster response systems. Each category requires different governance approaches, with safety-critical planners demanding the most rigorous evaluation.

The governance gap is particularly acute because embodied AI development is accelerating rapidly. Major technology companies are pouring resources into this space, and these systems are moving from research prototypes toward real-world deployment in robotics, autonomous vehicles, infrastructure planning, and national security applications. Getting policy frameworks in place before these systems are deeply embedded in critical infrastructure remains a pressing challenge for regulators worldwide.