Figure 03 tidies living room via Helix 02

This is the most convincing household-robotics demo to date — not because of what the robot did, but because of how it did it: one model, zero task-specific algorithms, emergent generalization from data alone.

• –Helix 02 introduces a three-tier architecture: System 0 (1 kHz balance/contact control), System 1 (200 Hz full-body joint targets from perception), and System 2 (slow reasoning for scene understanding and goal sequencing) — a clean separation of timescales that mirrors how humans move
• –The "pixels to torque" framing is significant: it collapses the traditional perception-planning-control pipeline into a single learned model, removing a major source of brittleness
• –No new code was written for this demo — the robot generalized to living room tidying purely from additional training data, validating Figure's scaling hypothesis
• –Figure 03's redesigned hands (sensitive to a few grams of pressure) and wider-FOV cameras are what make this level of dexterity physically possible
• –With BotQ manufacturing targeting 12,000 units/year, Figure is signaling this isn't a research demo — it's a product roadmap