Figure AI: 64 Hours, 80,000 Packages and a Roadmap Ahead

Figure AI wrapped a 64-hour uninterrupted autonomous humanoid livestream at its logistics facility. Five Figure 03 units sorted over 80,000 packages while CEO Brett Adcock publicly addressed teleoperation skepticism and outlined a series of strategic plans for the coming years.

Key takeaways

The Figure 03 fleet operated autonomously for 64+ hours without a single mechanical failure, maintaining a 2.9-second throughput per package
The company plans to fully eliminate dependence on Chinese component suppliers within the coming quarter
Figure 04 has entered design lock — Adcock called it robotics' "iPhone 1 moment"
The new humanoid hand in Figure 04 packs more actuators than the rest of the robot's body combined
Figure has begun training Helix-3 on a full NVIDIA Blackwell B200 datacenter cluster — the largest model in Figure's history

How live autonomy works

The livestream started as a direct response to a public endurance challenge from RoboStrategy's Scott Walter, who questioned whether humanoids could sustain real industrial utility. Adcock responded with a live broadcast commitment. The planned 8 hours turned into a multi-day marathon.

Each robot — Bob, Frank, Gary, Rose, and Jim — ran entirely locally, with no network connection during task execution. Helix-02, Figure AI's proprietary vision-language-action (VLA) model, processes camera input and computes motor torques directly on an onboard computer inside the robot's torso. No remote control, no cloud in the loop.

When a unit's battery runs low — typically after around four hours — it autonomously signals an idle fleet member to take its place. The handoff takes under 30 seconds. The departing unit docks at an inductive charging pad built into the floor. Across the full broadcast, zero mechanical robot failures were recorded.

Addressing the teleoperation argument

The most pointed critique of the livestream centered on a visible gesture: robots frequently raised their left arm toward the head mid-task. For skeptics, this pattern signaled human-in-the-loop teleoperation.

Adcock addressed this directly in a Bloomberg Technology interview. When the robot turns left to grab a package from the conveyor, the whole-body controller automatically lifts the left hand to clear the right arm's path. The movement isn't a human cue — it's an optimization feature of Helix-02 designed to prevent self-collision during compound motions.

Figure 04: a ground-up redesign

Alongside the livestream, Figure confirmed design lock for Figure 04 — full architectural closure ahead of production preparation. Adcock described the machine as "unrecognizable" compared to previous generations.

The centerpiece change is the hand. Prior Figure units used forearm-actuated, tendon-driven hands — a design Adcock acknowledged as an engineering mistake that capped the quality of training data. Figure 04's new hand was engineered with watchmaker-level precision and carries more actuators than the rest of the robot combined. Full finger range-of-motion matters because limited kinematics in previous hands "polluted" demonstration datasets — the robot couldn't mirror complex human movements, such as folding socks, accurately.

Figure 04 will run on Helix-3, whose training run started last week. The company is using a full NVIDIA Blackwell B200 datacenter cluster deployed in the weeks prior.

Onshoring and Chinese supply chain elimination

Beyond the technical milestone, the broadcast served as a platform for strategic announcements. Figure holds over $1 billion in cash and is deploying that position to de-risk its supply chain. Over the past year, the company has relocated or diversified production of motors, gearboxes, sensors, and PCBs. Adcock stated that by the end of next quarter, Figure will have zero supply chain exposure to China.

This is a deliberate response to geopolitical risk: tariffs, export restrictions, diplomatic pressure. Figure is not alone in making such claims, but it's one of few robotics companies naming a specific deadline.

In parallel, production throughput is scaling. Figure's BotQ factory on its campus is producing 60–70 robots per week, putting the company on an annual run rate of several thousand units.

Why it matters

The Figure AI livestream was more than a tech demo. It was a direct challenge to the structural skepticism that has followed humanoids for years: that videos are cut from long teleoperation sessions, that robots don't work without constant supervision, that robotics is hype without real-world substance.

The broadcast ran for days without cuts. Sorting data was publicly verifiable in real time. Adcock answered questions live, including those targeting specific movements that raised doubts. This sets a new bar for transparency in the sector — and increases pressure on competitors.

At the same time, Figure disclosed a roadmap that extends well beyond package sorting. The Figure 04 hand, the Helix-3 architecture, the B200 cluster, the onshoring — these are building blocks of a company positioning itself for generalization: the ability to drop a robot into any unknown environment and have it succeed based solely on verbal instruction. That goal remains distant, but the next steps are concrete and committed.

What's next?

Figure plans OTA deployment of Helix-3 across the full fleet once the model passes internal validation
Figure 04 production starts per company roadmap — Adcock has committed to generalization demonstrations in new environments later in 2026
Full Chinese supply chain elimination is targeted for Q3 2026