Physical AI Data and Tooling
Diving deeper into
Mike Xia, CEO of Anvil Robotics, on humanoid vs. non-humanoid robots
Inside big companies that had raised something like $150 million, the team working on manipulation and physical AI was still maybe five to six people
Analyzed 8 sources
Reviewing context
This says the bottleneck in physical AI is not headcount, it is data and tooling. Even well funded robotics groups often still look like skunkworks teams because the hard part is not hiring 50 more people, it is getting a small group enough robot time, sensors, simulation, and training data to make manipulation models improve in the real world. That is why infrastructure vendors can matter as much as robot makers.
-
The work itself is unusually concentrated. NVIDIA describes robotics as a stack spanning training, simulation, and on robot inference, and its research output is centered on manipulation, grasping, and sim to real transfer. That matches a market where a few researchers and engineers can drive core model progress if the tooling is strong.
-
Humanoid companies are raising huge sums, but the real race is still collecting real world task data. Figure, Apptronik, and Agility are funded at a scale far beyond typical startup robotics budgets, yet the category is still early, with deployment and learning loops mattering more than raw organization size.
-
That creates room for non humanoid suppliers like Anvil. If many teams are still tiny and all need arms, cameras, mounts, computers, and a fast path to first data collection, the winning product is the one that removes months of integration work before any model training starts.
Going forward, the winners in physical AI are likely to be the companies that turn a five person research pod into a repeatable deployment machine. That favors platforms that compress setup time, standardize hardware and software, and help small teams gather better manipulation data faster than larger but slower organizations.