Sergey Diakov
⬤ Figure just dropped Helix 02, a major upgrade to how their humanoid robots move and work. Instead of using about 109,000 lines of manually written C++ code, they're now using a single learned neural system that handles walking, grabbing objects, and keeping balance all at once. It's basically a shift from old-school programming rules to letting the robot learn how to move naturally.
⬤ The heart of Helix 02 is something they call System 0—a whole-body controller trained on over 1,000 hours of human movement data. They ran it through more than 200,000 simulated environments to teach it how to handle real-world situations. The model has around 10 million parameters and runs at 1 kHz, meaning it's sending commands to the robot's joints 1,000 times per second. That speed makes movements smoother and more responsive than older methods.
⬤ Figure showed off what Helix 02 can do by having their robot unload and reload a dishwasher completely on its own in about four minutes. They're calling it record-breaking performance for this type of complex task. The robot uses all its sensors—cameras, touch feedback, and joint position data—at the same time to adjust on the fly without needing special programming for each step.
⬤ Helix 02 fits into a bigger control system with three levels: System 2 handles high-level thinking and decision-making, System 1 manages vision and movement coordination, and System 0 takes care of the moment-to-moment control at high speed. The goal is to let the robot move autonomously through entire spaces instead of just doing single tasks. This release shows where humanoid robotics is headed—less hand-coding, more learning, and more capability in the real world.
Sergey Diakov
