Figure's new humanoid model controls its entire body from pixels

Humanoid robots can backflip and dance, but ask one to walk across a kitchen, grab a mug, and load the dishwasher, and it usually freezes.

That's because mobile manipulation has always been a plumbing nightmare in robotics. Engineers typically stitch together two completely separate systems: one software stack drives the legs, another moves the arms. The robot walks, stops, plants its feet, picks up an object, and walks again. If it bumps a counter or drops something mid-stride, the whole state machine breaks.

Figure just launched Helix 02 to kill that handoff problem for good. While the original Helix only controlled the upper body from pixels, Helix 02 turns the entire robot into one continuous, learned neural system.

Throwing away 100,000 lines of C++

To make a 100-pound machine move naturally without falling over, Figure completely ditched 109,504 lines of hand-crafted C++ balance code. In its place sits System 0, a compact 10M-parameter neural model trained on over 1,000 hours of human motion data and 200,000 parallel simulation environments.

The entire brain operates in a clean three-tier stack:

System 2 acts as the planner. It handles language and scene reasoning, passing high-level intents like "carry the bowls to the counter."

System 1 is the visuomotor core running at 200 Hz. It pulls in head cameras, palm cameras, fingertip tactile data, and joint proprioception, outputting full-body joint targets.

System 0 executes at a blistering 1 kHz, taking those targets and translating them into torques while maintaining balance and contact stability.

Figure showed this off in a four-minute, end-to-end autonomous kitchen demo. The robot walks across the room, unloads dishes, and reloads them over 61 distinct actions without a single human reset. When both hands are full of bowls, it casually uses its hip to nudge a drawer shut and lifts the dishwasher door with its foot.

Feeling a paperclip

Locomotion is only half the battle; fine motor skills usually fail when arms occlude the head camera. Helix 02 takes advantage of Figure 03's new hardware upgrades to solve this.

Each fingertip now houses tactile sensors capable of detecting forces as light as three grams—about the weight of a paperclip. Combined with cameras embedded directly in the palms, the robot can now manipulate objects it can't see from above.

Figure demonstrated four autonomous dexterous tasks: unscrewing a bottle cap with force-modulated grip, plucking a single tiny pill out of an organizer, depressing a syringe plunger to dispense exactly 5 ml of fluid, and sorting tangled metal parts straight out of bins in its BotQ factory.

Why it matters

For years, humanoids have suffered from the demo curse: impressive short clips that fall apart in the real world because their software stacks are brittle patchwork quilts.

We think Helix 02 represents a massive architectural shift. Replacing hand-coded balance controllers with an end-to-end neural prior gives the machine something resembling genuine muscle memory. Moving, grasping, and recovering from balance disturbances are finally treated as parts of the same physical problem.

For enterprise customers and warehouse operators, this is what actual utility looks like. A robot that can carry objects while walking and use its hips or feet when its hands are full is orders of magnitude more useful than a static arm on wheels.

The caveat, of course, is reliability at scale. A continuous 4-minute autonomous run in a staged kitchen is a world-first benchmark, but production floors and real homes throw thousands of unpredictable edge cases an hour. Figure just proved the architecture works; now it has to prove it doesn't break when the real world gets messy.