Two sentences to start from
On 7 October 2026 Sequoia Capital published "Partnering with Mecka: Bringing AI to the Physical World" as it invested in Mecka, a company that builds training data for robots. Two sentences in it stand out.
The brain needs data that wasn't collected.
The second sets the condition for an answer: you need not one capability but all of them, working together and at scale. The first sentence names the problem; the second names what an answer has to look like.
The data a robot needs has to be made on site
Language models learned from text that had piled up on the internet. The data a field robot needs is different. Where a part sits right now, and how far a weld seam has drifted from the drawing, can only be known by reading the site itself.
On a heavy-industry floor the problem is sharper. Products are large and rarely identical, and the state of the work keeps changing. To do a task, a robot needs the 3D state of the site and the working coordinates as they are now. A drawing alone cannot tell it that.
The five capabilities Sequoia names
Research Anticipating the next shift in robotics and what it will take.
Hardware Designing the devices that capture data, and producing them quickly and economically.
Operations Collecting data at scale across many industries and jobs.
Software and AI Turning video, motion and sensor streams into data a model can learn from.
Deployment Placing robots on customer sites and running reinforcement learning and live operation.
Sequoia says the five must work together. LiOps goes one step further: on a heavy-industry or manufacturing floor they have to be one company's responsibility. There are two reasons.
First, error accumulates between stages
A sensor measures the site, software fits the measurement to the CAD, and the robot moves to those coordinates. Each stage introduces error, and each error is handed to the next stage. Checking the stages one at a time cannot account for the total that accumulates.
In millimetre-scale work that accumulated error decides the outcome. That is why LiOps chose to connect perception through control in one company and answer for the whole chain.
Second, site problems only show up on site
Site exceptions are hard to foresee in a lab. Only once the robot is actually on the floor does it become clear what is difficult, and the cause may be sensor placement, registration or the robot's path. LiOps keeps perception through control in one team so that a problem at any stage can be fixed directly.
How the data is handled matters too
Data from a heavy-industry site is design information and process know-how. Security is therefore a precondition for adoption, not an afterthought. When reviewing any technology that uses site data, ask first what is collected, where it is processed and who can access it. That holds even when one company runs the whole chain.
How LiOps does it
LiOps is a Physical AI company for shipbuilding and heavy industry. We supply Spatial MES, which reads the site, and the robot crews that work on top of it: the robot welding crew, autonomous logistics and robot machine tending. Both share the same 3D spatial model.
Research We rank task difficulty as inspection and measurement, logistics, painting and grinding, then welding, and we solved the hardest of them, free-form welding, first.
Hardware We build the multi-sensor system that reads the site continuously and our own autonomous AMRs. On the floor, sensors, robot and tool are tied into a single coordinate frame to form the robot work system.
Software and AI Our own 3D registration engine fits scan results to the CAD and measures progress. The robot crews work in the coordinates of that same spatial model.
Operations and deployment Robots are working in shipyard welding, equipment fabrication, factory logistics and machine tending for CNC lines.
Adoption runs in four steps.
Questions to ask before you buy
If your site is evaluating physical AI, we suggest asking any supplier these three questions.
In closing
LiOps defines Physical AI as bringing to a robot's physical work the level of generality ChatGPT showed in language work. The data a robot uses, meaning the state and coordinates of the site at the moment of work, has to be read on site and carried through to the robot's task, and LiOps believes one team has to answer for that whole chain.
The site cases continue in Robots already do the easy welds and We do not sell robots. We deliver welded volume.
Reference: Sequoia Capital, "Partnering with Mecka: Bringing AI to the Physical World" (Anas Biad, James Flynn, 7 October 2026). The original Korean post is on the LiOps blog (blog.liops.ai/physical-ai-field-data).