Brett Adcock: Humanoid Run on Neural Net, Autonomous Manufacturing, $50T Market #229

What It Covers

Brett Adcock details Figure's progress building general-purpose humanoid robots powered entirely by neural networks. Figure removed 109,000 lines of C++ code, achieving full autonomy with Helix 2 running closed-loop control for hours. The company manufactures robots at 50,000 units annually, targets commercial deployment in 2026, and aims for home robots by 2027-2028 at $20,000 per unit.

Key Questions Answered

• •Neural Net Architecture Transition: Figure eliminated all 109,000 lines of C++ code from their control stack, moving to pure neural network control with Helix 2. This enables the robot to perform closed-loop autonomous tasks for 67+ consecutive hours without human intervention, compared to competitors showing only open-loop preprogrammed behaviors or teleoperation. The shift allows continuous learning and generalization across tasks rather than brittle coded responses requiring expensive maintenance.
• •Manufacturing Cost Reduction: Figure 3 achieved 90% cost reduction versus Figure 2 through vertical integration of actuators, motors, hands, and sensors designed specifically for neural network operation. The current facility produces one robot every 30 minutes targeting 50,000 units annually. Vertical integration became necessary because off-the-shelf components lack proper sensors, compute, thermals, and firmware integration required for autonomous operation at scale.
• •Federated Learning Advantage: Once one Figure robot learns a task through neural network training, every robot in the fleet instantly gains that capability through model updates. This creates exponential knowledge accumulation impossible with human workers who must be individually trained. The company accumulates diverse training data across logistics, kitchen tasks, and manufacturing environments, with positive transfer learning improving performance across all domains simultaneously.
• •Commercial Deployment Timeline: Figure deploys robots to multiple signed commercial customers in 2026 using Figure 3 hardware running Helix 2. The robots operate in warehouses performing logistics tasks like package sorting at human speed with one error per 67 hours. BMW partnership provided critical learnings about fleet operations, safety protocols, and repair maintenance requirements. Home deployment alpha testing begins 2026 with general availability estimated 2027-2028.
• •Hardware Capabilities Unlocked: Figure 3 actuators can operate three to five times faster than current neural network policies enable, providing significant performance headroom as AI models improve. The robot runs four to five hours per charge with one-hour wireless charging through foot-mounted inductive pads. Palm-mounted cameras supplement head cameras for occluded manipulation tasks. The design prioritizes matching human capabilities at lowest cost and weight rather than superhuman performance.
• •Validation Through Teleoperation: If a robot can be teleoperated to perform a task, the hardware proves capable and neural networks can learn that behavior through training data. Figure uses teleoperation as a hardware validation tool rather than a deployment strategy. Competitors shipping teleoperated robots or open-loop preprogrammed behaviors demonstrate no progress on the core challenge of closed-loop autonomous control in unseen environments.
• •Market Scale Projection: Adcock estimates tens of billions of humanoid robots will exist globally, with every human owning at least one robot plus five to ten billion deployed in commercial workforce applications. At $20,000 per unit, this represents a $50 trillion market replacing human labor. The limiting factors are solving general-purpose neural network control, achieving robot-built-robot manufacturing loops, and securing working capital for production scaling to millions then billions of units.