Figure AI recently made waves with its Index initiative, a jaw-dropping billion-dollar commitment to collect mountains of real-world data for training robots. We’re talking 16 million videos from across the globe, all to teach future automatons how to navigate our complex human world. It’s an ambitious, costly endeavor built on the premise that you simply can’t get enough real human experience. But what if there’s a smarter, much more affordable way to get robots up to speed?
Enter Antioch, a New York-based startup that just snagged a cool $32 million in Series A funding, led by heavy hitters like Greylock. Their co-founder, Harry Mellsop, doesn’t think Figure’s approach is wrong, just that it’s out of reach for most companies. Antioch’s big idea? High-fidelity simulated data. Imagine being able to test every tweak and new skill for a robot at cloud scale, running thousands of parallel evaluations without ever needing to touch a piece of hardware. It’s a game-changer for physical AI development, dramatically shrinking the time and expense involved, and even Amazon’s Ring is on board, validating that Antioch’s simulations accurately mirror real-world results, even in tricky, “held out” scenarios.
So, is it simulated data *instead* of real data? Not quite. Mellsop argues that while real-world data is the gold standard, simulation helps companies become “sample efficient.” You still need some real data, but instead of spending a fortune gathering everything from scratch, you use simulation for the bulk of the training, then feed that precious, high-quality real data back into the simulator to refine its accuracy. This creates a powerful feedback loop, accelerating progress for everything from drone delivery and warehouse robots to smart security systems and construction automation, all without needing a Tesla-sized fleet to generate your training set.