Helix Runs on Embedded GPUs

Running the model on the robot itself is what makes Figure look less like a lab demo and more like a factory product. In a BMW plant, a humanoid cannot wait on a network round trip before deciding how to grip a part, shift its feet, or pull a cart. Figure is pairing that low latency control loop with industrial deployments that feed new training data back into Helix, which is the core reason its stack looks more operationally mature than most humanoid peers.1 2 3 4 5

• Helix is designed as one neural control stack that takes onboard sensor input and drives the full body, not just the arms. That matters because factory work like sequencing parts or moving carts mixes seeing, grasping, balancing, and stepping in one continuous motion.4 5
• Figure has already used that stack in production settings. Figure 02 ran regular shifts at BMW Spartanburg and contributed to 30,000 vehicles, then Figure 03 moved into a more complex logistics workflow there. Onboard inference is much easier to trust in that setting than a cloud dependent control system.2 3
• The strategic tradeoff is cost and vertical integration. Figure is spending heavily on GPU infrastructure, training data, and robot hardware together, and has raised capital at a $39B valuation to do it. That creates a very different model from a software only robotics company or a teleoperation first approach like 1X's earlier deployments.1 6 7

The next phase is a race to turn embedded VLA control into a data flywheel. Companies that can keep robots working in real customer environments, learn from those runs, and push better policies back onto onboard compute will widen the gap between impressive demos and repeatable commercial deployments.1 3 6