The next battleground in humanoid robotics may not be how well a robot performs a chore, but how much it already knows before entering the room. Figure AI says its latest model completed household tasks across 30 unfamiliar homes, suggesting broad human-behavior data could become as important to robots as hardware itself.

Figure Raises The Bar

Figure's Helix 2.5 was tested in 30 Bay Area homes the robots had never seen, with no data collected in those homes and no environment-specific fine-tuning. The humanoids were asked to tidy living rooms, fold towels and make beds, tasks requiring navigation, perception, manipulation and whole-body coordination.

The striking result came from Figure's Index pretraining. According to the company, an otherwise identical model trained from scratch succeeded on just 9% of blind zero-shot trials, while the Index-pretrained version reached 56%. Figure says the only experimental difference was Index pretraining, with architecture, optimization, downstream data and evaluation held constant.

"Zero-shot" does not mean the robot learned these tasks without training. Rather, the homes and objects used for testing were new to the model. The result suggests a robot can carry knowledge from human experience into environments it has never encountered.

Tesla Has A Different Data Moat

That puts an interesting spotlight on Tesla, Inc. (TSLA)'s Optimus strategy.

Tesla has said it is using its strengths in real-world AI data to advance Optimus, its general-purpose autonomous humanoid robot. Its broader AI strategy is built around training neural networks on billions of examples of real-world data.

Tesla therefore has a potentially significant data advantage from its vehicle fleet and AI infrastructure. But Figure is attacking the same generalization problem from a different direction: collecting human behavior specifically to teach robots how people navigate and manipulate the physical world.

Figure says its Index dataset is now generating roughly 35 minutes of new human experience every second. It has also committed $3.5 billion of compute to training Helix through a partnership with Nscale.

That makes the competitive question less about which humanoid looks most capable in a controlled demonstration and more about which company can build the better physical-intelligence training loop.

The Next Test Is Scale

Figure's 56% result is notable, but it does not mean humanoid robotics is solved. More than four in 10 trials still failed, and the evaluation covered three household behaviors.

For Tesla investors, the more important question is whether the company can turn its real-world AI data advantage into the kind of generalization Figure is now demonstrating.

That is the metric worth watching as both companies scale: not just whether Optimus or Figure can complete a task, but whether each new piece of training data makes the robot capable of handling environments it has never seen before. The company that solves that problem at scale could move humanoid robotics closer to mass deployment.