Written by |
Edited by | Wen Shuqi
At 8 AM sharp on June 25, eight humanoid robots clocked in again, marching back to the assembly line like seasoned veterans ready for a shift. They were there for quality control duties on the 3C production line.
This scene unfolded at the Longcheer Technology mass-production facility in Nanchang, Jiangxi. Since June 23, these eight Genie G2 units from NUPIAO have been running a six-day live stream operation that keeps going until June 28.
The Genie G2s are sticking to the standard human work schedule: “start at 8, finish at 7.” They’ve taken over the loading, unloading, and testing phases for the entire tablet mass-production QC section, syncing up perfectly with human workers for continuous collaboration.
This marks NUPIAO’s second time putting robots under the public eye for non-stop work. After an eight-hour live stream in April , this time the Genie G2 covered the full spectrum of tablet QC processes. It handled high-precision tasks independently—multimedia testing, audio checks, radiation spurious emission tests, and coupling tests—showing off a mature commercial capability for end-to-end autonomous operations.
Ai Wen, Project Director of the Genie Business Unit at NUPIAO, told reporters that the robots’ work pace has basically caught up to human levels, hitting about 80% to 90% of a human worker’s efficiency.
Data from NUPIAO shows that by the second day of the livestream, the eight robots had worked for over 22 hours total. They processed 6,335 tablets on the line, executed 23,384 robot actions, and maintained an impressive success rate of 99.97%.
When asked about that tiny gap to 100% success, Ai Wen was honest: “Strictly speaking, it’s not really the robot messing up. It’s the friction caused when two completely different industrial systems try to merge.” The drop happened around 3 PM on day one. A robot acting as the communication hub requested a move on the conveyor belt, but the belt didn’t respond in time. Because the previous tablet wasn’t cleared away, the robot ran out of space while trying to place the next one according to its program, leading to a stack-up error.
This highlights the biggest pain point of getting humanoid robots into factories: the fundamental logic clash between brand-new intelligent systems and traditional industrial automation.
Ai Wen explained that bringing robots into a factory is more like the rollout logic of self-driving cars versus conventional driving. However, factories still run on old-school industrial automation thinking. The definitions of faults, how they’re handled, and even the requirements for recovery are worlds apart.
In traditional factory equipment standards, allowing 4% to 5% downtime per day is normal. You just clear the error or reboot, and you’re back online. But for embodied AI, the core metric is “continuous operational uptime”—meaning the robot needs to run flawlessly for months without a hitch.
Ai Wen believes the goal is for the system to keep running for at least a month after a single-point failure, allowing only one small, quickly recoverable glitch.
Before this, the Genie G2 had already survived over 3,000 hours of extreme long-duration testing. He said, “From this perspective, we basically passed a passing grade exam.”
When talking about specific deployment details, Ai Wen pointed out that during long stretches of continuous work, the robot’s rhythm must match the entire production line perfectly. Any disconnect causes bottlenecks. To put a number on it, current robot efficiency sits at about 80% to 90% of human performance.
Notably, since different tests take varying amounts of time, NUPIAO adopted a more flexible strategy this time. One robot often has to manage two to four testing devices simultaneously to ensure everything stays perfectly synced with the line’s overall rhythm.
ROI (Return on Investment) is a hard metric you can’t dodge in the business world. Currently, NUPIAO calculates the return on deploying these robots based on roughly two years of labor costs for an equivalent human position.
Ai Wen revealed that due to the currently high cost of the full machine, NUPIAO isn’t prioritizing maximum profit right now. Instead, the focus is on getting robots into real industrial settings to gather the most valuable, real-world production line data.
A Goldman Sachs survey from January showed that clients are willing to invest in humanoid robots once they hit about 50% of human productivity, which corresponds to a payback period of about two years. Even a three-year payback period is considered acceptable.
Evidently, the ratio of human-machine efficiency and the commercial conversion cycle have become the key thresholds for the industry to move from “lab demos” to real-world application.
NUPIAO has defined 2026 as the “Year of Scale Deployment.” According to Ai Wen, the annual roadmap has two phases: the first half focuses on small-scale deployments and pilot validations across multiple scenarios, where seven solutions—including tablet testing and material handling in the 3C sector—are already proven. The true multi-scenario scale deployment will roll out gradually in the second half of this year.
In terms of volume, NUPIAO expects this year to reach the thousands level, aiming for tens of thousands next year.
This year, publicly broadcasting robots working continuously via livestreams is becoming the industry’s go-to method for stress-testing capabilities.
Just over a month ago in May, Figure AI pulled off a 200-hour fully automated livestream. Four Figure 03 robots worked in shifts, sorting nearly 250,000 packages without a single hardware glitch.
During the 10-hour head-to-head comparison, the Figure 03 sorted one package every 2.83 seconds on average—just 0.04 seconds slower than a human intern. But the livestream also exposed weaknesses: the robots struggled with unexpected issues like dropped items, sometimes stopping for no reason, failing to grab packages, and not adjusting their strategies proactively.
Facing this global competition, Ai Wen noted that many overseas vendors are currently stuck in lab environments, endlessly repeating single tasks. NUPIAO’s core strategy is to truly root itself in actual production lines, filling the industrial value gaps created by labor shortages.
However, stepping back from the application-level bragging rights, the industry is still far from true maturity when looking at the underlying capabilities of embodied intelligence.
“If we have to compare it to the history of large models, today’s embodied AI hasn’t even reached the GPT-1 level yet,” Ai Wen admitted bluntly. “The so-called ‘intelligence emergence’ is nowhere near happening.” The real bottleneck remains data accumulation. Currently, the total amount of real-world scenario data across the industry is only in the millions of hours. To truly trigger the emergence of embodied models, we need at least hundreds of millions of hours.
It’s not just about needing massive data; diving deep into real production lines also exposes other practical bottlenecks for large-scale robot deployment. Ai Wen mentioned that limited by hardware pain points like dexterous hands, high-precision precision assembly remains a blind spot for robots.
Furthermore, sending robots into factories isn’t a simple “machine replaces human” swap. Factories must retrofit workspaces and production line equipment to enable seamless communication. For now, smooth robot operations still rely heavily on human maintenance technicians standing by to handle anomalies or swap out batteries.
