Robotics News: Figure's F.03 humanoids sort over 100,000 packages across 81 hours of — May 21, 2026

Figure ran a humanoid nonstop on camera for 81 hours — the flex landed, but a livestream still isn't a deployment.


This was the week robotics stopped demoing and started clocking in. Figure livestreamed a humanoid sorting packages for more than 81 straight hours, Korea's RLWRLD open-sourced a foundation model that nearly doubled a pouring task's success rate, and Waymo's robotaxi footprint grew larger than the state of Rhode Island. The physical world is moving fast — so this issue separates what actually shipped from what was merely streamed.


Watch & Listen First

The Robot Report Podcast, Ep. 244 — Ajay Agarwal on why great robotics founders are systems thinkers · May 16, 2026 · The Robot Report
The Bain Capital partner who backed Kiva Systems before Amazon bought it explains why winning robotics companies sell an integrated system, not a robot.

Figure CEO Brett Adcock: "no teleoperation" in the package-sorting run · May 15, 2026 · YouTube
Adcock answers the livestream skeptics directly — worth watching to judge the autonomy claim for yourself.


Key Takeaways

  • Endurance is the new demo. Figure's multi-day run reframes the humanoid pitch from "can it move" to "can it hold a shift" — rivals will now be judged on uptime, not parkour.
  • A livestream is not an audit. No third party verified the run, and researchers flagged visible misgrasps; buyers still need independent uptime, maintenance and cost data.
  • Dexterity is outpacing locomotion. RLWRLD's open-weight RLDX-1 roughly doubled a pouring task's success rate — manipulation, not walking, is where the model race is heating up.
  • Robotaxis scale while permits tighten. Waymo's footprint now exceeds Rhode Island even as China froze new licenses — capacity and regulation are diverging by geography.
  • Defense autonomy is reaching the runway. Northrop's uncrewed fighter cleared autonomous taxi tests, the last ground checkpoint before first flight.

The Big Story

Figure's F.03 humanoids sort over 100,000 packages across 81 hours of nonstop livestream · May 17, 2026 · Interesting Engineering
Figure's F.03 robots ran Helix-02 fully on-board — no teleoperation, no cloud — clearing a barcoded package roughly every three seconds and recharging in staggered shifts so the line never stopped. Earlier humanoid demos were measured in minutes and heavily edited; an unbroken multi-day run is a genuinely different claim, shifting the benchmark from spectacle to stamina. But no independent party verified the stream, robotics researchers flagged visible misgrasps and packages nudged off the belt, and the "no teleop" assurance remains unfalsifiable — endurance was proven, deployment-readiness was not.


Also This Week

Uber commits over $10B to robotaxi rivals as Waymo's map outgrows Rhode Island · May 15, 2026 · Electrek
Waymo's coverage now tops Rhode Island ahead of the World Cup while Uber pours $10B into rival fleets — the robotaxi fight has shifted from who drives best to who can fund the most metro coverage.

Northrop's YFQ-48A uncrewed fighter clears autonomous taxi tests ahead of first flight · May 18, 2026 · Army Recognition
Independent steering, braking and throttle on the ground is the last milestone before an uncrewed fighter flies itself, putting the Collaborative Combat Aircraft program on a near-term flight timeline.

Romark Logistics deploys Dexory's autonomous inventory robots across its Pennsylvania warehouse · May 19, 2026 · Robotics & Automation News
A quiet, unglamorous AMR deployment for real-time inventory scanning is exactly the kind of boring-but-paying robotics that scales while humanoids chase headlines.


From the Lab

RLWRLD releases RLDX-1, an open-weight dexterity-first foundation model for robot hands · May 14, 2026 · The Robot Report
RLDX-1 is an 8.1-billion-parameter model whose Multi-Stream Action Transformer fuses vision, language, action, tactile and memory into a single policy. On real hardware — WIRobotics' ALLEX humanoid — it hit a 70.8% success rate on pot-to-cup pouring, roughly double comparable models stuck in the high-30s, and it runs across the Franka Research 3 and OpenArm too. RLWRLD published the weights, training code and checkpoints openly on Hugging Face and GitHub — a useful reminder that manipulation, not locomotion, is the hard frontier: a robot that walks all day still can't reliably pour your coffee.


Worth Reading


Robots pulled an 81-hour shift on live TV this week — building the worker turned out to be easier than trusting the timesheet.