Physical AI Isn't Only About Robots: Movement Data Starts at the Body
- Wave Company
- 23 hours ago
- 6 min read
MOVEMENT INTELLIGENCE · 01
Say "Physical AI" and most people picture a humanoid robot. The photographs that run with the articles are robots; the case studies that get presented to investors are robots. But the definition of the term is broader than that — and reading the definition carefully also reveals what the field is actually stuck on right now.
Updated: August 2026
What does Physical AI actually mean?
NVIDIA defines Physical AI as systems that process multimodal inputs — images, video, text, speech, and real-world sensor data — and convert them into insights or actions that an autonomous machine can execute. The scope includes robots and self-driving vehicles, but also fixed cameras, sensor-based systems, and physical spaces such as factories and warehouses.
The defining condition is not the form factor. It is where the data comes from. In Physical AI, data does not originate on a screen or inside a server; it enters through sensors from physical events in the real world. A system does not need legs to qualify.
Why is data the real bottleneck?
A language model trains on trillions of words that already exist in machine-readable form. A model that has to operate in the physical world is in a different position: it needs paired records of observation and action, and no corpus of that kind exists at internet scale.
Forbes framed the gap directly in June 2026. While language models consume trillions of tokens scraped from the web, robot foundation models are training on fewer than 5,000 hours of combined open-source real-world interaction data — because, as the piece puts it, every training example must be physically collected, one manipulation at a time.
Cost compounds the problem. Scaling a robotics dataset tenfold has been described as a millions-of-dollars exercise rather than a thousands-of-dollars one. Synthetic data generated in simulation helps with variety and scale, but the current consensus is that dexterity and failure recovery still have to be learned from real data.
The shortage in Physical AI, in other words, is not models. It is data coming off real bodies in the real world.
Where does human movement data come from today?
Human movement is one of the richest sources of physical-world data available. How far a joint bends, in what sequence load is applied, where balance breaks down — all of it is carried in movement. The difficulty is capturing it consistently outside a lab. Three approaches are in common use.
Camera-based motion capture is precise, but the data stops at the edge of the room.
Method | Where it measures | Main constraint | Everyday capture |
|---|---|---|---|
Camera motion capture | Outside the body, in a defined space | Bound to lighting and field of view; difficult in private settings | Hard |
Wrist-worn devices | At the wrist | Distance from the joint that is actually moving | Easy |
Sensor-integrated garments | On the joint itself | Wash durability, comfort, and sensor reliability must be solved | Feasible |
Cameras are precise but tied to a room. Step outside the lab or the studio and the data stops. Wrist devices can be worn all day, but there is a limit to what a wrist can tell you about how far a knee bent. That leaves a third option: the thing already on the body — clothing.
What changes when the sensor is the garment?
Clothing is the one interface people wear without thinking about it, and it sits directly over the joint. If the garment itself can sense, the point of measurement moves to where movement actually happens.
When the sensing layer is the garment itself, measurement happens at the joint rather than near it.
For a product team, that has a practical implication. The step where a user has to fetch a device, switch it on, and position themselves disappears, and data accumulates inside daily life rather than inside a designated space. Because no camera is required, measurement also becomes possible in settings — a home, a hospital room — where filming would not be appropriate.
There is a cost on the other side. Garments get washed, stretched, and worn differently by every body. Building a sensor that survives that is the real difficulty of this approach.
How does TracSil measure joint movement?
TracSil™ is a soft strain sensor printed onto fabric. When a joint moves, the fabric stretches, and the sensor reads that stretch and converts it into angle information. No camera and no separate measurement rig is involved.
A soft strain sensor printed onto knit fabric: the printed line deforms with the fabric as the joint moves.
Durability is a precondition rather than a feature. In internal testing, TracSil sensors withstood 10,000 cycles of continuous strain from 0% to 30%, stretched at a rate of 500mm per minute. Given that the strain typically experienced by the human body is up to around 15%, we anticipate that the sensors can withstand more than 100,000 cycles in real-world usage.
TracME™ is that sensor moved into a finished product. Sensors integrated into knee and elbow braces record joint movement, and the companion app uses that data to give feedback on posture and repetitions. TracME was named a CES 2024 Innovation Award Honoree. The sensor is not sold on its own — TracME is the product, and the movement data it produces is what other systems connect to.
What is possible today, and what is still being explored?
It is worth drawing the line clearly rather than overstating the position.
Available today
Recording joint angles and movement trajectories without a camera
Real-time posture checking during exercise and automatic repetition counting
Everyday-environment data capture at the knee and elbow
Still being explored
Extending coverage from individual joints toward full-body movement
Translating angle data into indicators that carry meaning in daily life, such as muscular endurance or injury risk
Clinical application, currently in a validation stage with partners
One clarification is worth making. TracSil is not EMG. It does not read electrical signals generated by muscle; it reads joint movement through the stretch of the fabric. The two methods measure different things and support different claims.
Why does this matter beyond fitness?
TracME is used for exercise today. The problem underneath it is not limited to exercise.
A way to record how a human body actually moves — outside the lab, over long periods, consistently — is a layer that rehabilitation, remote care, worker safety, and any system that has to understand human motion all need in common. If Physical AI is constrained by data, then widening the surface from which that data can come is the prior problem.
What Wave Company builds is the layer between body and data. We do not build robots. But whether the system is a robot or a coaching app, anything that has to understand real-world movement needs a place where that movement gets in. We are building that entry point into clothing, and delivering it as a finished product other systems can connect to.
Frequently asked questions
Does Physical AI include products that are not robots?
Yes. The working definition covers systems that process real sensor data and convert it into judgment or action in the physical world. Autonomous locomotion is not the criterion.
Is a garment sensor as accurate as camera motion capture?
They serve different purposes. Camera-based systems are strong at capturing the full body precisely within a controlled space. Garment-integrated sensors are strong at recording specific joints over long periods without spatial constraints. The two are better understood as covering different segments than as substitutes.
Can these garments be washed?
ElecSil™ has passed more than 100 tested washing cycles. For products using TracSil, please follow the care instructions supplied with the specific product.
Does TracME measure muscle signals?
No. TracME measures joint movement and angle. It is not an EMG-based system.
Can the TracSil sensor be bought on its own?
No. TracSil is the sensing technology inside the product rather than a component sold separately. TracME is the finished product, and the movement data it produces is what partners work with.
Curious about the product this movement data comes from?
Related
References
Forbes, Physical AI Hits A Data Labeling Wall That Only Cash Can Fix (29 June 2026)
Highways Today, The Real Bottleneck in Physical AI and Robotics is Data (19 August 2026)
This article is for general information and does not constitute medical advice.
