Figure just turned 4 years old
The ride so far has been surreal, and I wanted to share my perspective on what’s happened over the last four years
When I started Figure, the core technologies were still extremely nascent. Humanoid hardware was heavy, hydraulic, unsafe, expensive, and unreliable. Deep learning also wasn’t there yet - there was no real AI precedent and no clear path to building a truly general-purpose robot
My expectation was always that humanoid robotics would become the largest industry in the world, but that it would take a very long time to put all the hardware and AI pieces together
Honestly, I assumed it would take 20 years just to have a real shot at solving general robotics - decades before we could seriously attempt to build iRobot in real life
Instead, everything has moved much faster than I ever expected. Figure has gone through 4 major breakthroughs that probably accelerated our timeline by a decade:
> Cheap electric humanoid robots are now possible. There are many reasons for this but some include benefits from actuator torque density, sensor technologies, battery specific energy, high flop/memory onboard compute, and high-rate manufacturing techniques all helped significantly. At this point, we’ve largely de-risked the hardware side and it’s becoming a straightforward engineering problem (highly difficult engineering but tractable)
> Deep learning from camera pixels to torques actually works. The dimensionality of a humanoid robot is simply too high for hand-written code. You have 40+ motors that can rotate continuously, creating more possible robot body states than atoms in the universe. There is no path to building a truly general-purpose robot with heuristics and manually programmed C++ logic. This has to be AI-first, and Helix has now demonstrated that repeatedly
> Whole-body RL control changed everything. It’s what keeps the robot balanced, allows it to use stairs, move its arms, and coordinate its entire body through the world. We train these controllers entirely in GPU-accelerated physics simulation using reinforcement learning. The robustness is far beyond anything we ever achieved with hand-written heuristics. This fundamentally breaks the old thesis that walking robots are inherently too hard or too unstable to scale
> These robots can now perform useful human-like work at human-level speeds. This is an insanely hard problem because humans perform millions of different tasks in the real world - and Figure has to do this with 1 hardware platform in a general purpose architecture. btw, yesterday Figure completed a logistics use case running continuously for 200 hours without a failure - all with scalable hardware and internal AI models
The future is starting to feel very real
We have a real chance to build iRobot in real life - the good version
Thank you for everyone's support. Pedal to the metal
