Watch: Brainy humanoid robot masters package sorting at lightning 4-second pace

According to the accompanying update on the company’s website, the underlying Helix visuomotor system now processes each package in about 4.05 seconds, a 20 percent speed gain despite a jump in task difficulty.

How Helix closed the gap

Figure credits the leap to both more data and a deeper model. Training demonstrations grew six-fold, from ten to sixty hours, giving Helix far more examples of real-world fumbles, awkward shapes, and lighting quirks. On top of that volume boost, engineers added modules for short-term visual memory and force feedback. 

The memory lets the network remember partial glimpses of a barcode it saw a moment earlier and plan a rotation to expose it, while the force signal acts as a coarse sense of touch, guiding gentler grips and quick releases.

These adjustments pay off most clearly with non-rigid parcels. Alongside ordinary cardboard boxes, the latest run mixes in deformable poly bags and flat padded envelopes that flex, wrinkle, or fold as the gripper closes. 

Helix lets Figure humanoid alter its strategy on the fly, pinching thin mailers, flicking soft bags to flip them, and patting down bubbled plastic so a wrinkled label will scan cleanly. The result is a jump in barcode-ready orientation from about 70 percent to roughly 95 percent, achieved while throughput rises instead of stalls.

A glimpse of near-term autonomy

Small-parcel logistics offers the kind of fast-changing scene that favors end-to-end learning, and Figure’s update suggests the method matures quickly. In the controlled evaluation cited by the company, Helix’s combined speed and accuracy improvements track closely with the extra demonstrations and the new architectural blocks, pointing to a straightforward path for further gains. Collect more data, refine the memory, sharpen the touch.

Perhaps more intriguing is the evidence that those same policy weights can generalize. With only a handful of added examples, the robot reportedly learned to recognize a human worker’s outstretched hand and treat it as a signal to hand over a parcel instead of placing it on the belt, no separate handover script required. 

That kind of minimal-friction retasking hints at a future in which a single learning pipeline can cover dozens of warehouse micro-jobs, from kitting bins to palletizing.

For now, the demo remains a milestone rather than a mass deployment. Yet it makes the trajectory clearer. Figure, along with rivals such as Tesla’s Optimus, Agility’s Digit, Apptronik’s Apollo, and Unitree’s H1/G1, is betting that steady increments in data and network design will close the last gap between cautious prototype and reliable shift worker.

If an hour of uninterrupted sorting already approaches human dexterity, the next iterations could stretch the window to an entire shift, and then to a fleet. Each uptick in speed, touch, and memory will chip away at one more repetitive human task, until humanoid robots are no longer visitors on the warehouse floor but part of the crew.