A live broadcast on social media X that has continued since on the 14th. Gary, the Humanoid Robot "Figure 03," standing in front of the conveyor, picked up parcels one by one. It adjusted the orientation so the barcode faced down, set it on the conveyor, and immediately picked up the next parcel. When the items on the conveyor belt ran out, it bent toward the stack of delivery packages, picked up more parcels, and continued sorting. After Gary worked for about 3 to 4 hours and the battery level dropped, it autonomously requested a shift change, and other robots such as "Frank" and "Bob" took over. Multiple robots have been taking turns on their own, continuing to sort parcels for more than 50 hours. As of the live broadcast, they processed 65,300 parcels in 52 hours, or about 2.9 seconds per parcel.
Figure 03 is a next-generation humanoid developed by the U.S. robot startup Figure AI. Figure AI said this operation is fully autonomous based on the in-robot proprietary neural network AI model "Helix-02." Brett Adcock, Figure AI founder and chief executive, said, "The robot determines the position and orientation of parcels using only the pixel information from the cameras, without relying on external signals," adding, "AI inference happens on-device." He added, "When a person works, it takes about 3 seconds on average per package, and Figure 03 is now at a similar level."
Adcock said multiple humanoids are connected by a network to communicate with each other and are designed to maximize conveyor uptime. A robot that detects a low battery or an anomaly moves to the maintenance area on its own, and a replacement robot takes over the task. Adcock said, "The original goal was to run for 8 hours, but since nothing broke, we decided to keep going," noting, "Running continuously and autonomously for more than 24 hours without failure is uncharted territory."
◇ From demo videos to validating long shifts
Figure AI chose a live broadcast instead of a tightly edited promo video because the evaluation criteria for Humanoid Robots are changing. Until now, humanoid competition has centered on demo scenes such as walking, running, dancing, and grasping objects. But to be used in real factories and logistics warehouses, it is more important to steadily perform the same task for long periods, like a person, than to succeed in a move once or twice.
A robotics industry official said, "In long-duration work, a robot's battery life, joint durability, hand precision, and error response capabilities are laid bare," adding, "In this live broadcast by Figure AI, the robot showed limits in using fingers with fine precision, but its working speed improved and it performed autonomous shift changes, meaning its commercialization operations capability has climbed several steps."
Cases of deployment to actual production and logistics sites are also increasing. At BMW's Spartanburg plant in the United States, Figure AI's previous model, "Figure 02," was put into the vehicle body process last year. According to Figure AI, Figure 02 worked 10-hour shifts and moved more than 90,000 parts, and it was operated for more than 1,250 cumulative hours, being utilized in the production of about 30,000 BMW X3 vehicles. BMW also began a Humanoid Robot pilot operation at its Leipzig plant in Germany this year.
The trend is to assign highly repetitive tasks, such as moving and sorting boxes, to Humanoid Robots first. Global logistics company GXO Logistics is testing robots from several companies, including Agility Robotics, Apptronik, and Reflex Robotics, for warehouse work. Agility Robotics' bipedal robot "Digit" was deployed for tote box movement at GXO logistics facilities to build validation data, and in Feb. this year it also signed a robotics-as-a-service contract with Toyota Canada, expanding its applications to manufacturing sites.
◇ "Securing large-scale operational data will determine competitiveness"
In the industry, there is a view that Chinese corporations are a step ahead in the field of commercial validation for humanoids. They are rapidly increasing real deployments and accumulating on-site data based on a solid hardware supply chain and price competitiveness. China's UBTECH expanded cooperation with Airbus in the aerospace manufacturing field in Jan. this year and is testing the industrial Humanoid Robot "Walker S2" in aircraft manufacturing processes. Walker S2 touts the ability to replace its battery on its own within 3 minutes for 24-hour continuous operation. Last year, UBTECH's Humanoid Robot order value exceeded 1.4 billion yuan (about 309 billion won), and this year it is expanding to mass production with a goal of producing more than 10,000 units.
Another Chinese robot corporation, AgiBot, is also emphasizing long-duration operational performance. AgiBot's "A2 Ultra" avoided obstacles and determined its route without remote control during a 24-hour outdoor walking live broadcast in Aug. last year. In the process, it highlighted reducing downtime by replacing the battery without turning off the power. In Nov. of the same year, it walked 106.286 kilometers from Suzhou to Shanghai to set a Guinness record for the longest-distance walk by a Humanoid Robot. AgiBot said it has deployed more than 1,000 A2 Ultra units in actual operating environments.
An industry official said, "Chinese robot corporations are building operational data through low prices and rapid on-site deployment, while U.S. companies are chasing with AI models and major clients like BMW and Mercedes-Benz," adding, "Ultimately, whoever secures high-quality large-scale operational data first will determine the pace of commercialization."