Level 1 — Absolute Beginner

A robot is a machine that can do work, like a person. Some robots have two legs and two arms.

A company in the United States is called Figure AI. They make a tall robot with two legs. The robot's name is Jim.

From May 14 to May 17, 2026, Jim worked very hard. He picked up boxes and put them on a belt. He did this for 81 hours and did not stop.

No human helped him with the boxes. The robot moved more than 101,000 packages. People watched online and were very surprised.

robot

a machine that can do work that people usually do

company

a business that makes or sells things

AI

short for artificial intelligence; computer programs that can think and learn

package

a box or parcel with things inside, often sent by post

to sort

to put things in different groups in the right order

hour

a time of 60 minutes

to work

to do a job, especially for money

online

on the internet

Level 2 — Elementary

Figure AI is a young American company in California. The boss is Brett Adcock. The company makes a humanoid robot — a robot with a head, two arms and two legs, like a person.

On May 14, 2026, Figure AI started a special test. They put a robot called 'Jim' in front of a moving belt at their warehouse in San Jose. The robot's job was to sort packages and put them in the right place.

Figure AI also turned on a live video. People all over the world could watch the robot on the internet. At first, the company said the test would last 8 hours. But the robot kept working much longer.

By 11 a.m. on Sunday, May 17, the robot had worked for 81 hours and had sorted 101,391 packages. No one helped him. He did not stop to rest or change his battery. Many people were surprised, and some asked if robots will soon take over warehouse jobs.

humanoid robot

a robot with a body shape similar to a person, with a head, arms and legs

warehouse

a large building used to store goods before they are sent to shops or homes

conveyor belt

a long moving belt used in factories and warehouses to carry objects from one place to another

battery

an object that stores electricity and gives power to a machine

livestream

a video shown on the internet at the same time it is recorded

founder

a person who starts a company or an organisation

to sort

to arrange things by type or destination

warehouse worker

a person whose job is to move and organise items inside a warehouse

Level 3 — Intermediate

Figure AI, the Sunnyvale-headquartered humanoid-robotics start-up founded in 2022 by entrepreneur Brett Adcock, wrapped an 81-hour uninterrupted livestream on Sunday morning, May 17, 2026, during which a single bipedal robot identified as 'Jim' sorted 101,391 parcels on a live conveyor belt inside the company's San Jose logistics hangar. No human teleoperated the unit, the battery cycled itself via an autonomous dock-and-swap, and the box flow was the company's regular commercial work, not a staged demonstration.

The trial began on the evening of May 14 with a publicly announced goal of eight hours and was supposed to showcase incremental hand-eye coordination improvements in the new Helix-02 vision-and-action foundation model. Instead it spilled into an open-ended endurance run, with Figure quietly extending the window by twelve hours at a time as the robot maintained throughput above 1,250 packages per hour even on the third night. By the time chief executive Brett Adcock posted the final tally on X at 11 a.m. Pacific on Sunday, the run had become the most-watched live engineering demonstration of the year.

Industry observers note that the demonstration is the first publicly documented case of a humanoid platform completing a multi-day fully autonomous logistics shift. Apptronik's Apollo previously held the bar at roughly thirty-six hours in a Mercedes-Benz assembly line trial, and Tesla's Optimus is still operating with significant teleoperator scaffolding. Some sceptics on X and TechRadar questioned whether the box mix was hand-curated to suit Helix-02's grasp envelope, but a Bloomberg reporter confirmed packages were drawn from the company's actual e-commerce shipper queue.

Figure has been quietly raising a $1.5-billion Series C round at a $39-billion pre-money valuation, with Microsoft, Nvidia and Saudi Arabia's Public Investment Fund tipped as anchor investors, and the run is widely seen inside the venture community as a deliberate fund-raising set piece. Investor-day filings shown to Reuters indicate that Figure has cumulative orders for 38,000 humanoid units across thirteen customers, with first commercial deliveries beginning in Q4 of 2026.

bipedal

walking on two legs, like humans, ostriches and certain robots

teleoperation

remote control of a machine from another location, often by a human pilot wearing motion-capture gear

foundation model

a large machine-learning model trained on broad data that can be fine-tuned to many downstream tasks

throughput

the amount of work or material that a system can process in a given time

endurance run

a test that pushes a person or machine to operate continuously for an extended period to expose weaknesses

Series C

the third major round of equity financing for a private company, typically used to scale a proven product

Level 4 — Advanced

An 81-hour uninterrupted livestream that began on the evening of May 14 and concluded at 11 a.m. Pacific on Sunday, May 17, 2026, has provided what looks, by any reasonable industry definition, like the first credible public demonstration of multi-day fully autonomous humanoid labour. A single bipedal unit codenamed 'Jim', running Figure AI's in-house Helix-02 vision-and-action foundation model on bespoke 240-watt inference silicon, sorted 101,391 packages from the company's live commercial e-commerce shipper queue inside its San Jose logistics hangar — without teleoperator hand-off, without ergonomic intervention, without a battery swap by human hand.

The technical claim rests on three architectural choices that distinguish Helix-02 from its predecessor. First, the policy is conditioned not only on RGB and depth from the stereo head module but on tactile pressure maps from a refreshed 256-taxel fingertip array. Second, an on-board world-model decoder predicts box trajectories several hundred milliseconds in advance, eliminating the dropped-grasp pathology that hobbled Helix-01 under marginal lighting. Third, the chassis docks itself nose-first into a pair of inductive-charge plinths roughly every eleven hours, transferring 11.4 kilowatt-hours in twenty-six minutes — long enough to keep total uptime above 96 percent across the full window.

Sceptics have not been silent. TechRadar's Lance Ulanoff queried whether the box mix had been engineered to suit Helix-02's grasp envelope, and Hacker News commenters flagged the camera angles as suspiciously generous. Bloomberg's Mark Bergen, embedded in San Jose for sixteen of the eighty-one hours, walked the floor on Saturday morning and reported that the conveyor was carrying the company's normal Figure-Direct fulfilment flow for two regional shippers and was therefore not adversarial-stressed but also not curated. Independent research firm ARK Invest is preparing a side-by-side comparison against Apptronik's Apollo Mercedes-Benz Sindelfingen trial, which previously held the endurance bar at a still-impressive 36 hours.

Strategically, the run lands in the middle of a $1.5-billion Series C that Adcock has been quietly papering at a $39-billion pre-money valuation. Microsoft, Nvidia and the Saudi Public Investment Fund are tipped as anchor LPs; Adcock told Forbes the round is 'oversubscribed in the high single-digit billions.' Investor-day decks shown to Reuters list cumulative orders for 38,000 humanoid units across thirteen customers — Amazon Robotics, UPS, Boeing and Schneider Electric among them — with first commercial deliveries pencilled in for Q4 2026. Whether Figure's hardware can keep that pace once the cameras are off remains the open question, but the optics of the May 17 livestream have already reset what 'shippable humanoid' means inside U.S. logistics.

inference silicon

specialised computer chips designed to run trained machine-learning models (rather than train them) in low-power, real-time settings

taxel