Figure AI's Humanoid Robots Sort Packages Nonstop: What the Livestream Proves

Mon, 18 May 2026 17:58:10 +0800

Figure AI has pushed humanoid robots back into the center of the conversation.

Starting on May 14, 2026, Figure AI placed three F.03 humanoid robots in a logistics sorting scene and streamed the process continuously. Viewers nicknamed the robots Bob, Frank, and Gary. Beside a conveyor belt, they identify packages, pick them up, rotate them, scan barcodes, and place them back on the belt as required.

At first glance, the livestream looked like a public response to skepticism. If humanoid robots want to prove real utility, edited short clips are not enough. They need to survive full shifts, repetitive tasks, and long-running operation.

By the time The Paper reported on the stream, Figure AI had been broadcasting for five days and claimed that the robots had sorted more than 100,000 packages. The livestream can still be viewed on YouTube: F.03 Livestream.

Why this livestream matters

The humanoid robot industry has long had one recurring problem: demo videos are too short.

A few minutes of footage can show that a robot “can do” something, but it rarely proves that it can keep doing it. In real logistics, manufacturing, and warehousing, the key question is not whether one grasp succeeds. It is whether the system stays stable over long periods, handles exceptions, follows a maintainable rhythm, and makes economic sense per unit of work.

By choosing a livestream, Figure AI put the hard questions on the table:

Can the robots work continuously for hours or even days?
Do they require remote human control?
Can they handle battery, handoff, and maintenance needs?
Is the error rate acceptable in repetitive work?
Can they stay stable with soft parcels, rigid boxes, and packages of different sizes?

Compared with an edited clip, a long livestream exposes problems more easily. Dropped packages, failed grasps, short pauses, and changes in conveyor rhythm are all visible to viewers.

That is also its value. It does not prove the robots are perfect. It gives outsiders a more direct view of how far humanoid robots still are from reliable industrial use.

What is Figure F.03 doing?

The task is not complex, but it is typical.

The robot needs to observe packages on a conveyor belt, identify the barcode position, pick up the package, adjust its orientation, and place it back with the barcode facing down. It looks like a simple “pick up and put down” routine, but for a robot it involves several hard problems:

Recognizing packages with different shapes, materials, and sizes.
Estimating grasp points and weight changes.
Avoiding deforming soft parcels or pushing boxes off the belt.
Moving arms within limited space.
Maintaining rhythm without slowing the conveyor.
Recovering after a failure instead of freezing.

Figure AI founder Brett Adcock said the robots average about three seconds per package, close to human speed. He also stressed that the system is not scripted, but reasons and controls directly from camera pixels.

That point matters. The claim is not that the robot can repeat a preset motion, but that it can adjust grasping and placement strategies from real-time visual input.

Helix-02 is the core story

Figure AI emphasized that F.03 runs on its in-house Helix-02 system.

According to public descriptions, Helix-02 is not a traditional industrial robotics pipeline with neatly separated perception, planning, and control layers. It is closer to an end-to-end full-body autonomy system. It integrates vision, touch, proprioception, and whole-body control into one model framework so the robot can adjust its actions in real time.

You can think of it as three layers of capability:

Low-level control: keeping balance and executing joint movements.
Visuomotor policy: turning camera and tactile input into grasping, moving, and placing actions.
Semantic reasoning: understanding task goals, scenes, and abnormal states.

This is also where humanoid robots differ from traditional automation equipment.

Traditional sorting systems are usually optimized for fixed processes. They can be highly efficient, but changing the scene often means redesigning the line. Humanoid robots try to enter existing environments with human-like form factors and perform multiple tasks without rebuilding too much equipment.

The direction is tempting, but difficult. A robot’s hands, eyes, body, and “brain” must work together. If any part is unstable, the final result suffers.

The livestream also exposed problems

The stream was not flawless.

Based on reports from The Paper and other observers, the robots sometimes made short mistakes: inaccurate grasp judgments, shifted package positions, or even pushing packages off the conveyor.

These issues may be edited out of a demo video, but they cannot be ignored in real work.

Logistics environments care deeply about accuracy. One dropped package may be a small mistake. If the same pattern happens frequently in a large warehouse, it creates manual review, delays, damage, and responsibility issues.

U.S. robotics expert Ayanna Howard has raised a similar concern: the demonstration looks more like a science project than a mature commercial service. Speed matters, but in real deployment, accuracy, exception handling, and supervision cost matter just as much.

Are sorting workers about to lose their jobs?

In the short term, this livestream should not be read as “sorting workers are about to be replaced.”

Figure AI demonstrated a relatively controlled, repetitive, clearly bounded task. It shows that humanoid robots are approaching usability for some logistics motions, but it does not prove they can seamlessly take over a full warehouse workflow.

Real logistics sites face many more complications:

Damaged packages, liquid leaks, and unusual shapes.
Dirty barcodes or barcodes that are not visible.
Stacked, blocked, or jammed packages.
Temporary human intervention.
Equipment alarms and conveyor pauses.
Safety rules and liability boundaries.

Human workers are good at these non-standard exceptions. For commercial deployment, robots need to prove not only that they can approach human speed on standard actions, but also that they can handle long-tail problems reliably.

The more realistic change may not be full replacement. Robots may first take over part of the repetitive, boring, night-shift, or high-intensity work, while humans move toward supervision, maintenance, exception handling, and process optimization.

What it means for the industry

The livestream matters because it shifts the benchmark for humanoid robots from “can it perform an action?” to “can it keep working?”

In the past, the industry often compared isolated abilities: walking, moving boxes, folding clothes, cooking, washing dishes. Now Figure AI is trying to prove that humanoid robots can run for long periods in a real task while letting the public watch the process.

That creates pressure for competitors.

If other companies continue to release only edited clips, observers will naturally ask: Why not livestream it? Why not run it for eight hours? Why not disclose the error rate? Why not let the robot work at something closer to an industrial rhythm?

Of course, livestreaming is not the final answer. Commercialization still depends on:

Robot sale price and rental cost.
Maintenance frequency and battery life.
Deployment and tuning cost.
Throughput per unit time.
Error rate and accident rate.
Integration with existing warehouse systems.
Whether customers are willing to pay for a humanoid form factor.

If these numbers do not work, even a popular livestream is still just an impressive technology demonstration.

Summary

Figure AI’s F.03 package-sorting livestream is an important signal on the road to commercial humanoid robots.

It shows that humanoid robots are no longer limited to lab prototypes performing a few isolated motions. They are beginning to attempt long-running, repetitive, industrial tasks. The end-to-end full-body autonomy approach represented by Helix-02 also moves robots from “fixed-motion machines” toward “labor tools that understand scenes.”

But it still does not prove that humanoid robots are ready to replace warehouse workers at scale.

Speed, accuracy, exception handling, cost, safety, and maintenance remain open questions. The real thing to watch is not how exciting a livestream looks, but whether these robots can work for months at real customer sites with controllable costs.

If they can, the next stage of logistics automation may really be arriving.

Livestream Link

Figure AI F.03 Livestream - YouTube

Logistics Automation on KnightLi Blog