Figure AI’s Helix 02 Gives the Figure 03 Full-Body Autonomy — and It Just Did Your Dishes

Imagine walking into your kitchen and watching a humanoid robot — completely on its own — walk to the dishwasher, unload every dish, carry them across the room, stack them neatly in the cabinets, then return to reload the dishwasher and start it. No reset. No human hand-holding. Four uninterrupted minutes of real-world autonomy.

That’s exactly what Figure AI demonstrated on January 27, 2026, when it unveiled Helix 02 — the most capable AI model yet to control a humanoid robot’s entire body.

One Brain, Whole Body

The original Helix was already impressive: a single neural network controlling a humanoid’s upper body directly from camera pixels. Helix 02 blows that wide open. It extends unified neural control to the entire robot — walking, reaching, grasping, balancing — all as one seamless, continuous system.

What makes this remarkable is that it solves one of robotics’ longest-standing hard problems: loco-manipulation. Moving and manipulating objects at the same time has resisted clean engineering solutions for decades. The reason is deceptively simple — lift something and your balance shifts; step forward and your reach changes. Arms and legs constrain each other constantly.

Traditional robotics worked around this with clunky state machines: walk → stop → stabilize → reach → grasp → walk again. These hand-offs are slow, brittle, and deeply unnatural. Helix 02 ditches all of that.

The Architecture Behind the Magic: Systems 0, 1, and 2

Helix 02 runs on a three-layer hierarchy, each operating at its own timescale:

• System 2 thinks slowly — interpreting scenes, understanding language, and sequencing multi-step goals.
• System 1 thinks fast — translating perception into full-body joint targets at 200 Hz.
• System 0 (new to Helix 02) executes at 1,000 Hz, handling balance, contact forces, and whole-body coordination simultaneously.

System 0 is the star. It’s a 10-million-parameter neural network trained on over 1,000 hours of human motion data and reinforcement learning across more than 200,000 simulated parallel environments. Rather than engineering separate reward functions for walking, turning, or reaching, System 0 learned to move the way humans move — stable, natural, and adaptive.

The result? System 0 replaced 109,504 lines of hand-engineered C++ code with a single learned prior. That’s not an incremental improvement — it’s a paradigm shift.

Figure 03: The Body That Makes It Possible

Helix 02 runs on the Figure 03 hardware platform, equipped with embedded tactile sensors and palm-mounted cameras. These unlock a new class of dexterity. With its “all sensors in, all actuators out” visuomotor policy, Figure 03 can now:

• Extract individual pills from blister packs
• Dispense precise syringe volumes in healthcare scenarios
• Singulate small, irregular objects from cluttered piles — even when its own hands block the camera view

That last capability — handling self-occlusion — is a major milestone. Most robotic systems fail when they can’t see what they’re touching. Figure 03 feels its way through.

Why This Matters Beyond the Kitchen

The dishwasher demo is a compelling headline, but the implications stretch far beyond household chores. The same autonomy and dexterity that lets Figure 03 stack plates applies directly to manufacturing (handling irregular parts without pre-programmed fixtures), healthcare (precise medication handling with tactile feedback), and logistics (picking items of any shape in dynamic warehouse environments).

Figure AI has stated plans for 100,000 units of Figure 03, signaling this isn’t a research prototype — it’s a commercial play. As Helix continues to evolve, each software update makes every deployed robot smarter overnight, much like a smartphone receiving a new OS.

The Bigger Picture

What Figure AI is building with Helix 02 is something the robotics field has been chasing for decades: a robot that thinks and moves as a unified whole. Not a collection of subsystems duct-taped together, but a single learning system that perceives, reasons, and acts — continuously, in real time, in the real world.

The dishwasher task ran for four minutes without a single human intervention. That might not sound long, but in the timeline of humanoid robotics, it’s an eternity — and a preview of what’s coming next.