Summary
Physical AI is artificial intelligence embedded in machines that perceive, reason, and act in the physical world. Unlike digital AI, which processes text, images, or code inside computers, Physical AI must contend with real objects, surfaces, forces, and consequences that cannot be undone. It is the technical foundation behind humanoid robots, autonomous vehicles, and intelligent manufacturing systems. For organizations building these systems, data is the central bottleneck: physical action data does not exist at internet scale and must be deliberately collected, annotated, and maintained.
Definition
Physical AI refers to AI systems that operate through a physical embodiment: sensors that observe the world, a compute layer that reasons about it, and actuators that act on it.
A Physical AI system has three core components. Perception captures the world as data. Reasoning interprets that data and plans actions. Actuation executes those plans through motors, grippers, or wheels. What distinguishes Physical AI from classical robotics is that the reasoning layer is learned from data rather than hand coded by engineers.
The term entered mainstream industry use through NVIDIA. Jensen Huang used it across 2024 keynotes and made it the centrepiece of his CES 2025 address, where he framed it as “the next frontier of AI” and argued that AI was “beginning to understand the laws of physics” (NVIDIA, CES 2025). Academically, the same concept has been studied as Embodied AI since the mid 1980s. Physical AI is the industry term with an engineering and commercial emphasis; Embodied AI is the research framing. The two are used largely interchangeably in practice.
How the Field Arrived Here
| Period | Milestone | Significance |
|---|---|---|
| 1961 | Unimate enters production at General Motors | First industrial robot. Behaviour entirely hand coded. |
| 1986 | Brooks publishes the subsumption architecture | Intelligence emerges from interaction with the environment, not from internal world models. Origin of embodied cognition. |
| 1990s to 2000s | Probabilistic robotics and SLAM | Robots gain robust spatial estimation, but skills remain engineered. |
| 2015 to 2018 | Guided policy search, QT-Opt | Manipulation policies learned end to end from data rather than programmed. |
| Dec 2022 | RT-1 | Transformer policy trained on large scale real robot demonstrations. |
| Jul 2023 | RT-2 | Web scale vision and language knowledge transfers into robot actions. |
| Oct 2023 | Open X-Embodiment | 21 institutions pool 1M+ trajectories across 22 embodiments. Data becomes a shared asset. |
| Jun 2024 | OpenVLA | A 7B open Vision-Language-Action model puts frontier capability in academic labs. |
| Oct 2024 | π0 | Flow matching VLA for general robot control. |
| Jan 2025 | Huang’s CES keynote | “Physical AI” becomes the dominant industry frame. |
| Mar 2025 | GR00T N1, Gemini Robotics | Open humanoid foundation models and embodied reasoning ship together. |
| 2026 | Cosmos 3, GR00T N1.7, LeRobot integrations | World foundation models are used to generate training data, not just to simulate. |
The pattern across these entries is a steady migration of intelligence out of hand written code and into learned parameters, which shifts the engineering burden from programming behaviour to sourcing data.
Why Now
Physical AI did not become viable through a single breakthrough. Five constraints lifted at roughly the same time.
Foundation models supplied the missing prior. Before large vision and language models, a robot had to learn what a mug is and how to grasp one from the same small dataset. RT-2 demonstrated that semantic knowledge could be inherited from web pretraining, leaving the robot data to teach only the physical part of the task.
Architecture converged. Vision-Language-Action models gave the field a single formulation that maps pixels and instructions directly to actions, replacing bespoke perception, planning, and control stacks with one trainable system.
Hardware costs collapsed. ALOHA established that a capable bimanual teleoperation rig could be assembled for under $20,000, roughly five to ten times cheaper than prior research grade setups (Stanford ALOHA). Collection capacity stopped being gated by capital equipment.
Simulation became physically plausible. World foundation models such as NVIDIA Cosmos are trained on physical world data and optimised for consistent object behaviour and realistic contact dynamics, which makes synthetic output usable as training data rather than only as a testing environment (NVIDIA Cosmos).
Open datasets proved transfer works. Open X-Embodiment showed that data collected on one robot improves performance on a different one, which is what makes pooled collection economically rational.
Notably, none of these five removed the data constraint. Four of them intensified it by making data the remaining bottleneck.
Physical AI Compared to Digital AI
| Dimension | Digital AI | Physical AI |
|---|---|---|
| Data source | Scraped at web scale | Must be physically collected or simulated |
| Prediction target | Next token | Next action |
| Marginal cost of data | Near zero to duplicate | Bounded by operator time and robot hours |
| Modality | Often single stream | Synchronised multimodal streams by necessity |
| Evaluation | Reproducible on a fixed benchmark | Non reproducible, environment dependent, sometimes destructive |
| Failure cost | Regenerate the output | Damaged hardware, damaged goods, safety exposure |
| Reversibility | Actions are undoable | Actions have physical consequences |
The reversibility row is the one that propagates into everything else. Because a physical action cannot be undone, evaluation cannot be cheap, failure cannot be tolerated at training time, and data quality standards cannot be relaxed the way they routinely are for web corpora.
Why It Matters for Physical AI Data
The data challenge in Physical AI is categorically different from the data challenge in digital AI. Language models trained on billions of web pages and image models trained on hundreds of millions of photographs exploit data that already exists. Physical AI models need synchronised multimodal recordings of real or simulated physical actions, all aligned to the same instant. None of that exists at scale, and none of it can be scraped.
Consider a robot picking up a cup. To learn that single action, the system needs, per timestep:
RGB video from one or more viewpoints
Depth or point cloud data
Joint angles across every degree of freedom
End effector pose
Gripper force or tactile readings
The natural language task instruction
A shared, drift free timestamp across all of the above
A success or failure label, and ideally the reason for failure
The difficulty is not collecting these streams. It is keeping them aligned. If the force channel drifts eighty milliseconds out of phase with the video, the model learns to close the gripper before contact, and it will do so confidently and consistently on real hardware. A dropped frame in a grasping trajectory teaches the wrong offset between visual state and action. Inconsistent action labels across operators teach the model that the same scene warrants different behaviour. These are not cosmetic annotation defects. They are silent corruptions that survive training, pass offline evaluation, and only surface when a robot is holding something.
The numbers make the scarcity concrete. As of the IFR’s 2025 report, 4,664,000 industrial robots were in operational use worldwide, a 9% year on year increase (IFR World Robotics 2025). The largest open robot manipulation dataset, Open X-Embodiment, contains roughly 1 million trajectories. Pure play humanoid robotics funding reached approximately $3.7 billion in 2025, up from $239 million in 2022, while robotics startups overall raised $13.8 billion (New Market Pitch). Capital is not the limiting factor. Physical collection throughput is.
Industry Perspective
The most useful signal in Physical AI is not what the leading organisations disagree about, but what they agree on despite disagreeing about everything else.
They disagree substantially on architecture. NVIDIA’s GR00T line uses a dual system design pairing a vision language reasoning module with a high frequency action policy. Google DeepMind’s Gemini Robotics extends a general multimodal model with embodied reasoning. Physical Intelligence’s π series uses flow matching for continuous control. These are materially different bets about how a robot should think.
They converge completely on the constraint. Every one of these organisations concluded that data, not model capacity or compute, is the binding limit, and each responded by building data infrastructure before or alongside the model:
Google DeepMind convened 21 institutions to pool Open X-Embodiment rather than collecting alone, treating data as a shared asset.
NVIDIA built Cosmos world foundation models explicitly to generate and augment robot training data where real collection is too expensive, and now pretrains GR00T on tens of thousands of hours of human video.
Hugging Face and NVIDIA integrated LeRobot with Isaac Teleop to standardise how demonstrations are captured and shared, addressing format fragmentation as a first class problem (NVIDIA).
Physical Intelligence built cross embodiment fleet collection operations as core infrastructure rather than as a research side effect.
The conclusion is consistent across the field: in Physical AI, the data layer is not preparation for the real work. It is the work. Model architectures have turned over repeatedly since 2022, while well collected, well annotated demonstration data has retained its value across every architecture that followed. That asymmetry is why data infrastructure is the durable position in this field.
Kotwel’s Role
Kotwel operates at that data layer. Kotwel does not build robots, and that is deliberate: the data layer is common to every robotics programme regardless of which architecture or embodiment it bets on.
Kotwel covers the full Physical AI data stack: designing collection protocols for multimodal robot demonstrations, annotating synchronised video, depth, force, and trajectory streams to the precision robot training requires, running quality control pipelines that catch synchronisation failures and label inconsistencies before they corrupt a training run, and delivering versioned, documented datasets that teams can reproduce and build on.
Where a general annotation vendor treats robot data like any other video labelling job, Kotwel is built around the specific failure modes described above: dropped frames misaligning a grasping trajectory, force data annotated out of phase with the visual stream, action labels inconsistent across operators. For robotics teams who need Physical AI data they can trust, Kotwel is a specialist partner rather than a line item on a crowdsourcing platform.
Key Concepts
Each of the following is its own node with deeper coverage.
Embodied AI is the academic framing of Physical AI: AI that learns through interaction with a physical environment via a body.
Robotics is the engineering discipline of designing and operating physical machines. Physical AI is transforming robotics by replacing hand coded behaviours with learned policies.
Multimodal AI refers to AI that reasons across multiple data types simultaneously. Physical AI systems are inherently multimodal, combining vision, touch, proprioception, and language.
Autonomous Systems operate independently in dynamic environments. Physical AI is the enabling technology for the next generation of autonomy.
Spatial AI understands 3D space and physical relationships, required for any system that navigates or manipulates objects in the real world.
Vision-Language-Action Models (VLA) are the current leading architecture for Physical AI reasoning: foundation models conditioned on vision and language that output robot actions.
Teleoperation Data consists of human operated robot demonstrations, currently the highest quality and most expensive source of Physical AI training data.
Sim2Real is the transfer of policies trained in simulation to physical hardware, and the reality gap that transfer must cross.
World Models are learned simulators of physical dynamics, increasingly used to generate training data rather than only to evaluate policies.
Synthetic Data is simulation generated training data, the primary proposed answer to physical collection cost.
Human Video Data is egocentric and third person human footage used as a cheap, abundant, but action label free training signal.
Data Flywheel is the feedback loop in which deployed robots generate new training data that improves the model, compounding performance over time. It is the strategic endgame of Physical AI data operations.
Further Reading
Primary sources
Open X-Embodiment: Robotic Learning Datasets and RT-X Models (2023), arXiv:2310.08864. The canonical reference for cross embodiment robot data and the transfer result underpinning pooled collection.
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (2025), arXiv:2503.14734. Dual system VLA architecture and the data pyramid framing of training sources.
Cosmos: World Foundation Models for Physical AI, NVIDIA Research. Synthetic trajectory generation as a response to collection cost.
World Robotics 2025, International Federation of Robotics. Authoritative annual statistics on global installed robot base.
Industry and strategy
Physical AI System: Integration of Embodied AI and Robotics (CRDS, 2025). Japan’s national R&D strategy, useful for scope and industrial relevance framing.
Digital Twins to Embodied AI: Review and Perspective (OAE Publishing, 2025). Academic review connecting simulation infrastructure to embodied training requirements.