Figure AI, the humanoid robotics startup currently boasting more high-profile backers than a Glastonbury headline slot, has finally revealed the ace up its sleeve: Index. In an announcement today, the firm unveiled its ambitious plan to construct what it’s calling “the most diverse robot training dataset ever built” by capturing a “global sampling of physics” from the real world. Put simply, they’re building a library of reality to teach their robots how to actually be useful.
The core premise is refreshingly straightforward: the internet, for all its cat videos and dodgy life hacks, is a rubbish place to learn how to navigate the physical world. To create a truly general-purpose robot, you need data that simply doesn’t exist online. Figure’s solution is to create it from scratch by recording humans performing everyday activities. The “Index Collect” system, showcased in recent demonstrations, captures first-person video and granular motion data from a human operator, simultaneously mapping those actions onto a digital twin of its Figure 01 robot.
This isn’t just about teaching a robot to do the washing up. Figure is aiming for a massive, foundational dataset that allows its machines to understand and interact with the physical world in a generalised way. The company stated that the data required to scale a general-purpose robot “has to come from the real world”—a direct shot across the bows for those relying heavily on purely simulated environments.
Why does this matter?
The biggest bottleneck for truly useful, autonomous humanoids has always been data, not just the nuts and bolts of the hardware. A robot’s clumsy fumbling usually stems from a lack of the billions of data points we humans pick up throughout a lifetime of physical experience. Figure’s Index is a wildly ambitious attempt to brute-force that learning curve.
By building its own proprietary “ImageNet for robotics,” Figure isn’t just training its bots; it’s creating an incredibly valuable and difficult-to-replicate asset. While rivals like Google aggregate existing research and Tesla leans on its vehicle fleet, Figure is betting that a bespoke, human-centric dataset is the secret sauce for embodied AI. The only question remains: can watching humans do the chores truly capture the nuance of physical intelligence, or are they just building the world’s most overqualified dishwasher?
You can read the full, albeit brief, announcement on their official site: Introducing Index.