Figure AI's latest humanoid robots have achieved something that separates genuine products from expensive demos: they completed 67 consecutive hours of autonomous work with only one error, running entirely on neural networks with zero human intervention. This milestone represents a watershed moment in robotics, not because of flashy movements or viral videos, but because it demonstrates the kind of reliable, closed-loop autonomy that actually matters in the real world. What Changed: From Hand-Coded Instructions to AI Learning? When Peter Diamandis visited Figure AI headquarters in San Jose, he witnessed a fundamental transformation in how the company approaches robot control. A year ago, Figure's robots relied on several hundred thousand lines of C++ code, handwritten by engineers to anticipate every possible scenario. Today, that entire approach is gone. Figure deleted 109,000 lines of C++ code and replaced it with a single neural network called Helix 2 that controls the entire robot, from hands and arms to torso, legs, and feet. This neural network handles full-body coordination, real-time planning, and dynamic responses to unexpected situations. The shift matters because neural networks learn from experience rather than following explicit instructions. "If you can teleoperate the robot to do a task, you can train the neural net to learn it," said Brett Adcock, Figure AI's founder. Brett Adcock, Founder and CEO at Figure AI Here's the practical implication: when one Figure robot masters a task like folding laundry, every Figure robot on the planet instantly learns that skill. Humans don't work this way. Robots do. This creates an exponential learning curve that traditional, code-based robotics simply cannot match. How Is Figure Designing Robots Around AI Instead of Bolting AI Onto Hardware? Most robotics companies design the physical robot first, then figure out how to make AI work with it. Figure inverted this approach entirely. The company looked at the neural network architecture it wanted to run and asked a simple question: what hardware do we actually need to make this work? The result is Figure 3, a complete redesign built around Helix rather than an incremental upgrade. The improvements are substantial: - Manufacturing Cost: 90% reduction in manufacturing expenses compared to previous models - Weight: Approximately 20 pounds lighter at 135 pounds total, improving mobility and energy efficiency - Sensing Capabilities: Palm cameras for grasping objects in tight spaces, plus tactile sensors in every fingertip for precise manipulation - Mobility: Passive toe joint for better range of motion and natural walking patterns - Safety Design: Soft-wrapped body to eliminate pinch points that could injure humans or damage objects - Compute Independence: Onboard inference compute so the robot doesn't depend on cloud connectivity to function Critically, Figure designed Figure 3 as a data-gathering machine. Every sensor, camera, and interaction feeds back into the training loop, because betting on neural networks means betting on data quality and diversity. The more varied, high-quality data the company collects, the better the robot generalizes to new situations. Figure manufactures its own actuators, hands, battery systems, and embedded compute rather than relying on off-the-shelf components. This vertical integration exists because existing robotics components simply aren't reliable enough. If a vendor's motor fails in the field, you're stuck waiting for them to fix it. If you built it yourself, you iterate overnight. Why Is Figure Building Robots to Build More Robots? Figure's manufacturing facility in Baku has four production lines with capacity for 50,000 robots per year when fully ramped. But Brett Adcock isn't stopping there. He's already planning additional facilities capable of producing tens of thousands, then hundreds of thousands, then millions of units. The company is putting humanoids on its own production lines this year. Robots will assemble other robots, test other robots, and package other robots. This recursive approach isn't just about efficiency; it's about the only viable path to scaling to a billion units. Every improvement Figure makes to the robot's dexterity, speed, and reliability makes it better at building the next generation of robots, creating a flywheel effect that becomes nearly impossible to stop. "Brett estimates they could ship a billion robots today if the AI were fully general-purpose. The demand is there. The capital markets via leasing models can finance it. The constraint is solving general robotics," noted Diamandis. Peter Diamandis, Founder and Executive Chairman at Singularity University What Actually Counts as a Real Robot Product? Most people are impressed by teleoperation, open-loop behaviors, or 30-second demos. These are not impressive to serious roboticists. What matters is closed-loop, autonomous work in unseen environments over long time horizons. Closed-loop means the robot continuously senses its environment and adjusts in real-time, not replaying a pre-programmed sequence. Autonomous means no human in the loop and no remote operator controlling it. Unseen environments means you can drop the robot into a random home or factory floor it has never visited, and it figures out how to navigate and work there. Long time horizons means hours, days, or weeks of continuous operation, not edited clips. Figure's current benchmark is four to five hours of continuous neural network operation in logistics, kitchen work, and manufacturing tasks. The 2026 goal is more ambitious: drop a robot into an unseen home and have it do useful work for days with minimal human intervention. Once Figure achieves that milestone, the game fundamentally changes. If the robot can generalize to any home, it can generalize to any environment, from factories and warehouses to hospitals, senior care facilities, mining operations, and space stations. When Can You Actually Buy a Figure Robot for Your Home? Brett Adcock's answer is direct: not yet. He emphasized that he won't ship products that don't meet his standards. The timeline breaks down into clear phases. In 2026, Figure plans alpha testing in homes with a small number of robots doing long-horizon work like cleaning, organizing, laundry, and dishes in real households. The goal is measuring human interventions, tracking how often someone needs to step in and help. Industrial deployments currently see occasional errors; the target for home deployment is orders of magnitude better. From 2027 to 2028, Figure expects scaled home pilots with tens, then hundreds, then thousands of units. Each phase involves iterative design based on real-world feedback, safety validation, privacy validation, and reliability validation. This cautious approach reflects lessons Adcock learned from his previous company, Archer Aviation, where he discovered that shipping products before they're ready creates far more problems than it solves. The robotics industry has spent decades chasing the wrong metrics. Figure's 67-hour marathon isn't a marketing stunt; it's evidence that the company has solved the fundamental problem of building robots that actually work in the real world without constant human babysitting. That's the difference between a product and an expensive remote-controlled toy.
