Key Takeaways:
The new benchmark in humanoid robotics is no longer a backflip — it's a webcam pointed at a factory floor, running for days without failure.
Humanoid robot makers are abandoning choreographed demos for multi-day factory livestreams, forcing the industry to prove its machines can work real production lines without human intervention. The shift marks a turning point for a sector that has spent years promising utility but only recently began delivering it in public view.
"This closes the gap between demo and deployment," said Peng Zhihui, founder of Agibot, which began a six-day livestream of its G2 humanoids working a tablet production line at Longcheer Technology's factory in Nanchang. The robots, equipped with customized grippers, operate in the quality-inspection section of a mass-production line, picking up, moving and placing tablets for testing alongside human workers.
The livestream trend follows Figure AI's 200-hour continuous run in which its Figure 03 processed roughly 249,560 packages — about 1,248 boxes an hour, or one every 2.88 seconds. In an earlier 10-hour challenge, a human intern named Aime beat the robot by a margin of just 192 packages, sorting 12,924 to the robot's 12,732, but quit with sore hands while the machine kept working. "Humanity's last win, perhaps," Figure's team posted.
The push toward public validation signals a maturation that could accelerate commercial adoption and investment. Chinese manufacturers alone shipped roughly 13,000 humanoid robots in 2025, according to research firm Omdia, and Morgan Stanley has since doubled its 2026 forecast for China to 28,000 units, projecting the market could reach 54 million units annually by 2050.
The Credibility Gap That Livestreams Are Closing
For years, humanoid makers competed on spectacle: backflips, dance routines, and tightly edited demo reels. The problem was that none of those proved a robot could survive a factory shift. The tedium of a livestream — watching a robot pick up the same object for hours — is precisely the point.
Figure AI's earlier Figure 02 went into BMW's plant in Spartanburg, South Carolina, placing sheet-metal parts into fixtures. Agility Robotics' Digit became the first humanoid with a real job in late 2024 and has since signed a Robots-as-a-Service agreement with Toyota Motor Manufacturing Canada, deploying seven units at the Woodstock, Ontario plant that produces the RAV4 and RAV4 Hybrid. Tesla's Optimus has been demonstrating battery sorting and parts handling inside Tesla's own factories.
In China, the list of real-world deployments is growing fast. Ubtech's Walker S has trained inside auto plants for NIO, Geely's Zeekr, BYD and others, and expanded to Foxconn and SF Express. Xiaomi hired a humanoid for an "internship" in its own car factory, reporting a 90.2% success rate on bilateral nut-fastening over a three-hour autonomous run. Agibot's broader bet is a full-stack approach combining hardware, developer tools, and motion datasets that let robots learn from physical environments rather than rely purely on pre-programmed routines.
None of these machines is yet the optimal tool for every task. For fixed, repetitive motions, a purpose-built robotic arm remains faster and more reliable, and skilled humans are still cheaper and far better at handling unexpected events like a jammed box or a dropped part. But the gap is narrowing fast, and the burden of proof has shifted: the credible flex is no longer a demo video but a livestream that runs for days and shows whether the robot breaks.
The Economics of Embodied AI
The financial stakes are becoming clearer. UBTech has begun mass deliveries of its Walker S2, targeting 500 robots delivered this year on the way to an annual production capacity of 5,000 units by 2026 and 10,000 by 2027. Orders for the Walker series have topped 800 million yuan (about $112 million) since early 2025, with customers including BYD, Dongfeng Liuzhou Motor, Geely, FAW-Volkswagen, and Foxconn. Enterprise pricing runs from roughly $145,000 to $180,000 per unit.
At the lower end, Unitree's G1 humanoid starts at about $13,500, with 23 degrees of freedom, while EngineAI's PM01 is priced around $12,000 to $13,700 for research and early commercial development. The wide price range reflects a market still searching for product-market fit, but the direction is clear: costs are falling as deployment scales.
The demographic imperative driving adoption is most acute in Asia-Pacific. Japan, South Korea, and China are all aging faster than almost any other major economies, each short hundreds of thousands of workers in logistics, construction, eldercare, and retail. World Bank research on East Asia and the Pacific found that between 2018 and 2022, robot adoption helped create an estimated 2 million jobs for skilled workers across five ASEAN countries while displacing roughly 1.4 million lower-skilled positions.
For investors, the question is which companies will capture the value. The humanoid supply chain spans actuators, sensors, batteries, AI chips, and software platforms — a web that touches Nvidia, Tesla, and a growing roster of Chinese manufacturers. Nvidia's physical AI pipeline, combining Isaac Sim for physics rendering and Cosmos world foundation models for synthetic training data, represents the current state of the art in closing the sim-to-real gap that remains the central engineering constraint.
This article is for informational purposes only and does not constitute investment advice.
