One Brain Across Robot Types

The key shift is that robot AI is starting to look less like firmware for one machine and more like a shared software layer that can move across many bodies. Skild AI is explicitly training one model across humanoids, quadrupeds, table top arms, and mobile manipulators, while Physical Intelligence has published cross embodiment models that transfer skills from mixed robot and human data into new tasks in homes. That makes the scarce asset the training loop, not the shell around it.1 2 3 4

• Skild frames its product as one brain for any robot, then uses data from different morphologies to improve the same base model. In practice, that means a warehouse arm, a walking humanoid, and a mobile base can all contribute examples to one training corpus instead of each needing its own stack.1 2
• Physical Intelligence is pushing the same idea from the research side. Its pi0 and pi0.5 work use diverse cross embodiment data, including human demonstrations and web data, to control different devices and handle longer home tasks like kitchen and bedroom cleanup in unseen houses.3 4
• This changes how hardware companies compete. Figure is collecting household video through Brookfield’s 100,000 residential units to improve Helix, while 1X treats teleoperation as a bootstrap for its world model. The winner is less likely to be the prettiest robot, and more likely to be the company that turns every deployment into reusable training data.5 6 7

Over time, this pushes robotics toward a split market. More hardware makers become bodies, sensors, and distribution, while standalone AI companies sell the control layer, fine tuning, and developer tooling that sit above them. As those models improve, new robot categories should come online faster because each new deployment trains the same shared brain instead of starting from zero.1 2 3 5 6