Helix dual-GPU onboard autonomy

Running the full stack on the robot is what turns Helix from a lab demo into something a factory can actually buy. A warehouse or auto plant cannot depend on perfect Wi-Fi to keep a 61 kilogram machine balanced while it reaches, grips, and walks. Figure splits the work so one onboard GPU handles slower scene understanding and task planning, while a second chip keeps the fast motion loop running, which cuts latency and avoids a cloud round trip at the exact moment the robot needs to react.1 2 3 4

• This mirrors a broader two brain pattern in humanoid AI. High level reasoning can decide that a bottle should go on a shelf, but a separate action policy has to convert that into joint by joint arm, hand, and balance commands in real time. Figure formalizes that split in hardware, not just software.4 5
• The practical advantage is uptime in messy environments. If the robot is sorting packages, carrying parts, or loading a dishwasher, dropped connectivity does not stop perception or control. That matters more than headline model size because customers pay for completed tasks per hour, not for cloud dependence.1 2 6
• It also fits Figure's full stack strategy versus model only robotics companies. Figure trains Helix centrally on large GPU clusters, then deploys the distilled inference stack onto embedded NVIDIA GPUs in the robot. That keeps field data, training, and deployment inside one loop and makes every installed robot a data collection endpoint.1 3 7 8

The next battleground is shrinking more capability into the onboard stack while keeping control fast enough for real work. As more humanoid companies copy the same split between planning and control, the winners will be the ones that can improve dexterity and reliability through fleet learning without adding expensive remote supervision or fragile cloud dependencies.3 4 5 8