AI, Physical AI, World Models, VLA, VLM, and Other Terms We Should Stop Mixing Together
The robotics AI stack is becoming crowded with terms that are often used together, and sometimes used incorrectly.
Physical AI. World Models. World Foundation Models. VLMs. VLAs. Robot Foundation Models. Digital Twins. These terms are related, but they are not interchangeable.
For engineers, researchers, product managers, and founders working around robotics and autonomous systems, the distinction matters. It affects architecture, data collection, evaluation, safety claims, deployment risk, and product positioning.
This post is a practical glossary for the new Physical AI stack, an attempt to make the terminology sharper.
The short map
| Term | Practical meaning |
|---|---|
| AI | General computational intelligence for perception, prediction, generation, reasoning, planning, or optimization. |
| Physical AI | AI systems whose outputs are coupled to physical environments through robots, vehicles, machines, sensors, and actuators. |
| VLM | A Vision-Language Model that connects visual inputs with language understanding and reasoning. |
| VLA | A Vision-Language-Action model that maps visual observations and language instructions into robot actions. |
| World Model | A learned model of how an environment evolves over time. |
| World Foundation Model | A reusable large-scale world model intended to generalize across many physical scenarios. |
| Robot Foundation Model | A general-purpose model intended to support robot behavior across tasks, environments, and sometimes embodiments. |
| Digital Twin | A virtual representation of a specific real-world asset, system, process, or environment. |
1. Physical AI
Physical AI refers to AI systems that perceive, reason, decide, and act in physical environments. The important difference is that the output of the AI system may affect the state of the real world. In robotics, autonomous driving, drones, industrial automation, humanoids, warehouse automation, medical robotics, and defense autonomy, model outputs may become motion commands, control inputs, trajectories, manipulation actions, navigation decisions, or intervention decisions.
This changes the engineering problem.
In a Physical AI system, a failure may produce an unsafe motion, a collision, a failed grasp, a navigation error, or an unstable control response. That is why Physical AI requires system-level reasoning. The model is only one component. The full system also includes sensors, actuators, latency, dynamics, constraints, uncertainty, state estimation, control loops, and recovery logic.
Key point: Physical AI is AI whose outputs are coupled to the physical world.
2. World Model
A world model is a learned representation of how an environment evolves over time. In practical terms, it tries to answer:
Given the current state and a possible action, what is likely to happen next?
In robotics and autonomous systems, a world model may represent spatial structure, objects, motion, contact, scene dynamics, uncertainty, affordances, and future states. This makes world models relevant for planning, policy learning, simulation, scenario generation, evaluation, and synthetic data.
A world model is not the same as a simulator.
| Simulator | World Model |
|---|---|
| Usually engineered explicitly. | Usually learned from data. |
| Encodes rules, physics, rendering, maps, and scenarios. | Learns statistical and physical regularities from observed data. |
| Useful for controlled training and evaluation. | Useful for prediction, generation, planning, and learned simulation. |
| Can fail because reality is more complex than the simulator. | Can fail because learned predictions may look plausible but be physically wrong. |
NVIDIA Cosmos is one of the most visible examples of this direction. It is positioned as an open platform for Physical AI with world foundation models, video data processing, evaluation, post-training frameworks, and guardrails.
The important distinction is this: A simulator encodes a world. A world model learns a world. Both can be useful, Both can fail.
Key point: A world model predicts how the world may evolve, but visual realism is not the same as physical correctness.
3. World Foundation Model
A World Foundation Model is a large reusable world model intended to generalize across many environments, tasks, and physical scenarios. The ambition is to train a model that can represent, generate, or reason about many physical environments, not only one specific robot, warehouse, vehicle, or task.
In this framing, world foundation models may become infrastructure for robotics, autonomous driving, simulation, policy learning, and synthetic data generation.
NVIDIA Cosmos is currently one of the clearest examples of this category. Waymo has also introduced world-model-based simulation for autonomous driving. World Labs is working on spatial intelligence and 3D world understanding. Yann LeCun’s AMI direction is also explicitly focused on world models rather than language-only AI.
But a world foundation model should not be confused with a complete autonomy stack. It may help generate scenarios, may support policy training, may improve simulation, may provide predictive priors...But it does not automatically solve control, safety, state estimation, verification, or runtime failure detection.
Key point: A World Foundation Model aims to generalize predictive world understanding across many physical scenarios.
4. VLM - Vision-Language Model
A Vision-Language Model (VLM) connects visual information with language. A VLM can look at an image or video and answer questions, describe the scene, identify objects, interpret visual relationships, or reason about visible context.
For example, a VLM may answer:
| Visual input | Possible VLM output |
|---|---|
| A robot camera sees a pallet in the aisle. | “There is a pallet blocking part of the aisle.” |
| A camera sees a tool near a table edge. | “The tool is close to the edge and may fall.” |
| A warehouse camera sees a forklift and a human worker. | “A forklift is operating near a person.” |
VLMs are important for Physical AI as they connect raw visual inputs with human concepts, task descriptions, and semantic reasoning. But a VLM usually does not directly control the robot.
Key point: A VLM understands and describes visual information, but it does not necessarily act.
5. VLA - Vision-Language-Action Model
A Vision-Language-Action model connects visual perception, language instruction, and robot action. A VLA does not only describe the scene. It can generate actions.
| Input | Output |
|---|---|
| Camera images, robot state, and the instruction “pick up the red cup.” | A sequence of robot actions intended to pick up the red cup. |
Google DeepMind’s RT-2 was an important milestone in this direction. It showed how VLMs trained on web-scale data could be adapted into VLA models that output robot actions. Google DeepMind’s Gemini Robotics continues this direction by focusing on perception, spatial reasoning, task planning, and real-world robotic behavior.
Figure AI’s Helix is another important example - a generalist humanoid VLA model that connects perception, language understanding, and learned control.
The architectural shift is important: A VLM can say what it sees. A VLA can decide what to do. But this also changes the evaluation problem. Once a model generates actions, the question is no longer only whether the output is semantically correct. The question is whether the action is physically feasible, safe, recoverable, and valid under the current state of the world.
Key point: A VLA moves from understanding the world to acting in the world.
6. Robot Foundation Model
A Robot Foundation Model is a general-purpose model intended to support robot behavior across tasks, environments, and sometimes embodiments. This is still an emerging category, and different companies define it differently.
NVIDIA Isaac GR00T is one example, focused on humanoid robot reasoning and skills. Physical Intelligence’s π0 is another example, described as a generalist robot policy trained across diverse robot data. Skild AI frames its system as an embodied AI or robot brain that can generalize across different robotic bodies.
The analogy to language foundation models is useful, but limited. In language, the model produces tokens. In robotics, the model’s output may become motion.
That means embodiment matters. Hardware matters. Control frequency matters. Contact matters. Safety margins matter. Evaluation is also harder because the same policy may behave differently across robot bodies, sensors, environments, and tasks.
Key point: A Robot Foundation Model aims to provide reusable behavioral intelligence for robots, but it is not a complete robotic system.
7. LeRobot
LeRobot is Hugging Face’s open-source robotics library for real-world robot learning in PyTorch.
Different robots.
Different control stacks.
Different datasets.
Different hardware assumptions.
Different evaluation protocols.
LeRobot helps reduce this friction by providing tools for collecting data, training policies, sharing datasets, fine-tuning models, and running learned policies on real hardware.
This is an ecosystem contribution, not only a software contribution.
Open-source LLMs changed the pace of language AI because they made models, datasets, evaluation, and fine-tuning more accessible. LeRobot may play a similar role for robotics and Physical AI by making robot learning more reproducible, inspectable, and shareable.
SmolVLA is a concrete example from this ecosystem. Hugging Face describes SmolVLA as a lightweight foundation model for robotics, designed for efficient fine-tuning on LeRobot datasets. Architecturally, SmolVLA takes multiple camera views, the robot’s current sensorimotor state, and a natural language instruction, then conditions an action expert to generate action chunks.
The point is that accessible models, shared datasets, and reproducible training pipelines change the speed at which the field can learn.
Key point: LeRobot helps move robotics AI from isolated lab demos toward shared, reproducible, open development.
8. Digital Twin
A digital twin is a virtual representation of a specific real-world system, asset, process, or environment. For example:
| Real-world system | Possible digital twin |
|---|---|
| Factory | Virtual model of a production line. |
| Warehouse | Virtual representation of aisles, robots, inventory, and workflows. |
| City | Virtual model of roads, traffic, infrastructure, and mobility patterns. |
| Robot cell | Virtual representation of a specific robot, workspace, tools, and constraints. |
A digital twin is usually tied to monitoring, operations, planning, maintenance, and engineering analysis. It may include real-time data feeds, historical data, simulation components, and dashboards. A digital twin can use AI, simulation, and world models, but it is not identical to any of them.
Key point: A digital twin represents a specific real-world system, while a world model learns a more general predictive structure.
Why these distinctions matter
The Physical AI stack is becoming layered.
A practical architecture may include perception models, VLMs, VLAs, world models, simulators, digital twins, robot foundation models, synthetic data pipelines, and runtime integrity layers.
Each layer solves a different problem.
| Layer | Primary role |
|---|---|
| Perception model | Identifies what is observed. |
| VLM | Connects visual inputs to language and reasoning. |
| VLA | Maps perception and instructions into actions. |
| World model | Predicts possible futures. |
| Simulator | Provides controlled training and evaluation environments. |
| Digital twin | Represents a specific real-world system. |
| Synthetic data pipeline | Expands training and evaluation coverage. |
| Robot foundation model | Provides reusable behavior priors. |
| Runtime guardrail | Evaluates whether the system should trust and execute a proposed action. |
The field is moving quickly, but the basic engineering questions remain simple:
- What does the system observe?
- What does it believe about the world?
- What does it predict will happen next?
- What action does it propose?
- What physical constraints apply?
- What uncertainty exists?
- What can go wrong?
- Should this action be executed now?
That last question is becoming one of the most important questions in Physical AI.
Final thought
The next decade of AI will not only be about models that talk, see, or generate. It will be about models that act. Once AI begins to act in the physical world, the central question changes. It is no longer only:
Can the model perform the task?
It becomes:
Can the system know when its own action should not be trusted?
That is the boundary between AI demos and deployable Physical AI.
References
- Hugging Face LeRobot: https://github.com/huggingface/lerobot
- Hugging Face SmolVLA documentation: https://huggingface.co/docs/lerobot/en/smolvla
- Hugging Face SmolVLA blog: https://huggingface.co/blog/smolvla
- NVIDIA Cosmos: https://www.nvidia.com/en-eu/ai/cosmos/
- NVIDIA Cosmos research: https://research.nvidia.com/publication/2025-01_cosmos-world-foundation-model-platform-physical-ai
- Google DeepMind RT-2: https://deepmind.google/blog/rt-2-new-model-translates-vision-and-language-into-action/
- Figure Helix: https://www.figure.ai/news/helix
Author note
I am Dr. Barak Or, working at the intersection of AI, control systems, and autonomous systems.
This glossary reflects my current work on runtime integrity for Physical AI: how autonomous systems can evaluate whether their perception, prediction, and proposed action are trustworthy enough to execute.
