Physical Intelligence and Skild AI
General Intuition
This is a fight over who becomes the default software layer for robots that many different hardware makers can use. Physical Intelligence and Skild AI sit closest because both sell a general control model instead of a robot, but they push different wedges. Physical Intelligence leans into open weights and real robot fine tuning across tasks, while Skild leans into a cloud abstraction layer that lets enterprises map many robot types into one shared brain and improve it with field data.
-
Physical Intelligence looks closest on product shape. It takes camera feeds and language prompts, predicts action sequences in real time, and adapts across robot types. Its open sourced π0 weights and small data fine tuning model make it easier for developers and OEMs to test, customize, and feed deployment learnings back into the stack.
-
Skild AI looks closest on buyer pitch. A robot maker uploads joint and sensor specs into Skild Cloud, gets an auto generated control interface, then calls high level APIs across humanoids, quadrupeds, mobile robots, and arms. That maps neatly to enterprise buyers comparing hardware vendors but wanting one intelligence layer above them.
-
The three way difference is in data loops. General Intuition starts with Medal gameplay data and uses learned simulators to pretrain action models before light real world tuning. Physical Intelligence emphasizes direct robot control and partner deployments. Skild emphasizes cross embodiment fleet learning from many customer robots. Each is trying to solve the same data scarcity problem from a different starting asset.
The next phase is likely a land grab for integrations, deployment data, and developer mindshare. If modular buyers win, these independent labs can become robot middleware. If the best data keeps coming from fleets in the field, the edge will shift toward the company that turns partner deployments into the fastest model improvement loop.