Hardware-Coupled AI for Humanoids
Physical Intelligence
Figure is betting that the winning humanoid robot will be the one whose brain is built around its body, not bolted on from a general purpose model provider. That matters because a humanoid has to turn camera feeds, speech, joint positions, battery limits, and actuator behavior into split second movement. Figure now runs Helix as its own onboard vision, language, and action stack, while also making batteries, actuators, and robots in house. The goal is better task reliability and lower cost per deployed robot as data from each job feeds the next model update.
-
This is the clearest split between two robotics business models. Physical Intelligence sells a hardware agnostic robot brain for many robot types at about $300 per robot per month. Figure instead sells the full machine as a service at about $1,000 per robot per month, so it can capture both the software margin and the hardware learning loop.
-
The tradeoff is focus versus reach. Figure can tune Helix around its own 25 actuators, hand design, battery pack, and factory tasks like parts handling at BMW. Physical Intelligence can spread across many robot bodies and tasks, but it must keep adapting to fragmented hardware and cannot control the full deployment stack.
-
Agility shows the middle path. It is vertically integrated on hardware and fleet software, with RoboFab sized for 10,000 robots a year and live tote handling deployments, but it still centers on narrower warehouse workflows and conventional planning rather than a general foundation model. That makes it easier to commercialize now, but less likely to own a broad embodied AI platform.
The next phase of humanoids will be decided by who compounds real world data into lower cost, more reliable work cells fastest. If Figure can keep turning tightly coupled hardware and model updates into better uptime and cheaper units, vertical integration will look less like a robotics preference and more like the default playbook for category leaders.