From 2 hours 40 minutes to 50 minutes, the leap of one year is only the beginning.
This morning, at the finish line in Nanhaizi Park, Yizhuang, Beijing, a &#34;Lightning&#34; Honor humanoid robot sprinted across, clocking a net time of 50 minutes and 26 seconds. It completed the 21.0975 kilometers, exactly the same distance as the human competitors on the field today. When the first human runner crossed the finish line, &#34;Lightning&#34; had already been waiting there for nearly 17 minutes.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/e9046ae9-cd38-4898-a874-8abf3c27ec4f?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/e9046ae9-cd38-4898-a874-8abf3c27ec4f"/>
Just one year ago, on this very same track, the best-performing robot—Tiangong Ultra—took 2 hours, 40 minutes, and 42 seconds, with a navigator guiding it throughout the entire race.
One year, the same distance, shortened by nearly two hours—and this time, no one was following along.
<h2>What Happened Today</h2>
Today's champion actually faced some twists before securing the title. The first to cross the finish line was not &#34;Lightning,&#34; but the remote-controlled team &#34;Jueying Chitu,&#34; with a net time of 48 minutes and 19 seconds, nearly two minutes faster than &#34;Lightning.&#34; However, the rules of this competition stipulate that teams participating via remote control must have their completion times multiplied by a weighting factor of 1.2—a simple rationale: remote operation and full autonomy represent two different levels of technical difficulty. After applying the weight, &#34;Jueying Chitu's&#34; effective score was approximately 57 minutes, placing it behind Honor's &#34;Lightning.&#34;
This rule design itself deserves special mention. By using a coefficient rather than separate tracks to handle the two types of robots, the organizing committee ensured they would compete on the same leaderboard. Remote-controlled teams must clear a higher threshold to surpass the autonomous group. In other words, the awards in this race were essentially designed specifically for robots capable of navigating and making decisions independently.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d88f61dc-aa62-4846-b3dc-6831c21b6768?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d88f61dc-aa62-4846-b3dc-6831c21b6768"/>
Last year, almost all participating teams relied on remote or semi-autonomous solutions. This year, about 40% of teams chose fully autonomous navigation. This shift in proportion speaks even more volumes than the champion's time itself.
Changes in participation scale are also worth noting: the number of robots grew from 21 to over 300, teams increased from 20 to more than 100, and 26 brands covered the mainstream humanoid robot manufacturers currently operating in China. Additionally, four international teams from Germany, Brazil, Portugal, and France made their debut on this track.
<img src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTYwNTE0Nzkw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="" original-src="https://imageproxy.pbkrs.com/https://bang.xuangubao.cn/LTYwNTE0Nzkw"/>
Over two editions of the race, the distance remained the same, but this year added an ecological section through Nanhaizi Park, introducing more complex terrain. The difficulty was no less than last year's. Under these conditions, the champion time improved from 2 hours 40 minutes to under 50 minutes—all achieved without any human intervention. This is the industry's report card delivered in just one year.
However, one point needs clarification: the racetrack is a carefully designed known environment. Routes were planned in advance, terrain tested repeatedly, and participating teams had ample time to fine-tune for specific scenarios. Breaking into the 50-minute range here is a genuine achievement, but it measures capability under a specific set of conditions—not the same challenge as another type of test.
<h2>Robots Inside Factories</h2>
On the very same week as today's race, another batch of humanoid robots was operating in a completely different setting: assembly lines.
Over the past six months, several domestic vehicle manufacturers, battery producers, and component suppliers have gradually introduced humanoid robots into their production processes. According to public reports, CATL's Luoyang plant has begun trial runs of humanoid robots to handle material handling and assembly assistance at specific workstations. BYD and Agibot Robotics are advancing joint testing efforts. SAIC and GAC have also made strategic layouts at their respective manufacturing bases.
What are these factory robots doing? Roughly this scenario: pick up a bolt, align it with the hole, tighten it, set it down, then grab the next one. Repeating thousands of times daily. Workers walk around nearby; the floor may occasionally have oil stains or scattered parts; batches of components may vary slightly in shape; occasionally an anomalous part appears, requiring the robot to decide how to handle it.
This presents a fundamentally different kind of challenge compared to running a marathon. A marathon route is fixed, obstacles are known, rules are pre-defined, and the worst-case scenario is simply falling down and getting back up to continue. Factories are different—they require robots to operate reliably throughout an entire shift in a partially unpredictable environment, whether workers change their movement paths, a component shifts by two centimeters, or the floor becomes slightly slicker than yesterday.
Currently, most trial-run robots focus on roles with relatively relaxed precision requirements but high stability demands: chemical handling, high-intensity repetitive transfer tasks, and auxiliary operations near high-temperature environments. These are typically jobs humans avoid or find detrimental to health after prolonged exposure. The industry categorizes such scenarios as &#34;3D&#34;: dangerous, dirty, and dull—precisely where robots are most likely to establish a foothold first.
The prevailing industry view is that 2025 marks the proof-of-concept year for humanoid robots entering factories, 2026 will be the small-scale trial run year, and true large-scale deployment won't begin until 2027 to 2028 at the earliest. The 50-minute performance on the racetrack is a report card, but there remains a considerable distance between that achievement and full-scale factory operations.
<h2>Why This Is China's Home Ground</h2>
Of the over 300 robots participating today, almost all represent a condensed snapshot of China's humanoid robot industry: Unitree, Agibot, Tiangong, Fourier Intelligence, Charbot, Honor... 26 brands from across the country. The sheer scale of this gathering alone tells a story.
From components to complete units, China's presence in this industry far exceeds typical external perceptions. Key components forming the robotic motion system—reducers, servo motors, force sensors, linear actuators—account for approximately 63% of global supply chain participation. Industry estimates suggest that building a similarly scaled supply chain in the United States would cost roughly 2.2 times more than in China. In 2025, global humanoid robot shipments totaled approximately 17,000 units, with Chinese companies contributing about 14,000, representing over 80% of the total.
This density in supply chains stems less directly from humanoid robots themselves and more from decades of accumulated capabilities in lithium batteries, consumer electronics, and new energy vehicles. Most core components used in humanoid robots already had mature supplier ecosystems in smartphones and electric vehicles; they're now being integrated into a new product form. China holds another advantage: the world's largest manufacturing base means the greatest number of potential deployment scenarios and the densest real-world operational data sources—which are crucial for training robot brains.
Of course, this doesn't mean there's no competition. Boston Dynamics, Figure AI, and 1X Technologies still maintain significant technological advantages in high-end models. Tesla's Optimus mass production roadmap remains one of the most closely watched variables in the industry. In the latter stages of this competition, hardware gaps may gradually narrow, but software—the reasoning and generalization capabilities of embodied AI models—will ultimately determine the final landscape.
China has walked this road once before.
<h2>The Script from Ten Years Ago</h2>
In 2012, global electric vehicle annual sales were approximately 120,000 units, mostly Tesla Model S. At that time, few believed EVs could become mainstream within a decade. Range anxiety, lack of charging infrastructure, and excessively high costs seemed like unsolvable problems.
By 2022, global EV sales surpassed 10 million units, with China contributing around 6 million. BYD's annual sales exceeded Tesla's.
Looking back at those seemingly &#34;unsolvable&#34; problems from 2012, you'll find they were eventually resolved—not through single-point technological breakthroughs, but via an interlocking flywheel: mass production drove down costs, lower costs stimulated demand, expanded demand accelerated charging network development, and improved infrastructure further reduced consumer concerns. Once this flywheel started spinning, it became difficult to stop externally, and its acceleration often outpaced anyone's early predictions.
Humanoid robots' cost curves are now traversing a remarkably similar path. In 2024, top-tier models cost approximately $100,000 to $150,000 per unit. By 2025, some Chinese manufacturers have already driven this down to the $30,000 to $50,000 range. Based on current scaling expectations, prices could fall below $10,000 by 2027 to 2028. The slope of this curve nearly mirrors the price decline trajectory of lithium batteries.
Rest of World describes this industry's current position as the &#34;Pre-iPhone moment&#34;—the period just before smartphones emerged: hardware was functional, software was catching up, costs weren't yet accessible to the masses, but all critical elements were moving simultaneously.
Humanoid robots possess one additional accelerator absent during the EV era: large language models. Before 2022, robot logic relied on decision trees and hard-coded programming; changing task scenarios essentially meant redeveloping everything. Today, a robot's brain can be a general-purpose reasoning model capable of understanding natural language instructions, rapidly adapting to new scenarios, dramatically reducing &#34;onboarding costs&#34; and rapidly expanding the boundaries of generality.
Yet this analogy has a boundary that cannot be ignored. Electric vehicles replace internal combustion engines—another machine—not directly threatening anyone's livelihood. While power systems changed, driving itself didn't disappear. Humanoid robots, however, replace humans: workers earning salaries in factories. This resistance never appeared in the EV history, but in the future of humanoid robots, it will inevitably become a serious issue requiring careful handling, not merely a footnote to overlook.
<h2>Several Hurdles Yet to Be Cleared</h2>
Today's results look impressive, but several realities remain hidden beneath the glitz.
First is adaptation cost for deployment. Even a robot that broke the 50-minute barrier today would need to rebuild environmental models, recalibrate safety boundaries, and replan motion paths if moved to a new factory or production line. The cost and timeline for this process currently represent one of the biggest frictions hindering large-scale promotion. How to enable robots to get up to speed within days of entering a factory, like a new employee, rather than requiring months of dedicated tuning—is the core challenge each manufacturer is tackling. Whoever solves it first secures entry into the scaling phase.
Second is duration of stable operation. A marathon is a one-time event; once finished, it ends. Factories run three shifts daily, 300 days a year. Robots cannot stop due to fatigue. Current pilot data shows some factory robots operate continuously for four to six hours. Whether they can reliably sustain a full eight-hour shift without failure remains insufficiently validated.
Third is the legal and regulatory vacuum. If an autonomously deciding humanoid robot causes workplace injury or property damage in a factory, who bears responsibility? The robot manufacturer, the factory management, or the software team writing the algorithms? Current industrial safety standards are designed for mechanical arms, not dual-legged walking robots with autonomous judgment. Establishing this standard framework lags significantly behind technology itself, and this gap will remain a concern for enterprises entering the market until actual accidents occur.
A rarely discussed issue involves the humans in factories. Some pilot feedback indicates workers initially struggle to adapt to robot behavior—their movements don't resemble human motions, rhythms differ, and path choices sometimes leave humans uncertain about next steps. This unpredictability affects collaboration efficiency and can even impact worker morale. Establishing operating norms and psychological adaptation for human-robot coexistence is a gradual process requiring time accumulation, not something automatically resolved upon installing robots.
None of these issues are deadlocks, nor can they all be eliminated in the short term.
<h2>Conclusion</h2>
Today, &#34;Lightning&#34; completed 21 kilometers, finishing nearly 17 minutes ahead of the leading human competitor on the same day, with absolutely no one instructing it where to go along the entire route.
Just one year ago, accomplishing this required someone trailing alongside to provide guidance.
This speed improvement is real. But the racetrack is a deliberately engineered environment: the 21-kilometer route was pre-planned, terrain thoroughly tested, and the entire system optimized for this single competition. Factory floors are different—every day brings new conditions, component positions shift, ground conditions change, and worker movement patterns follow no fixed rules. Robots must operate reliably throughout an entire shift in such open uncertainty, at costs low enough for enterprises to genuinely pay, with clear rules so that when incidents occur, everyone knows whom to contact.
This industry has delivered a beautifully impressive report card today.
But between this polished performance record and the destination the industry truly aims to reach, there remains a considerable distance yet to travel.

00370

300024

06600

In just one year, humanoid robot half-marathon times plummeted from 2 hours 40 minutes to 50 minutes, with no human guide leading the way; Chinese companies now account for 80% of global shipments, and the cost curve is replicating the price decline path of lithium batteries. But a factory floor is not a racetrack—the true challenges have only just begun behind this impressive performance record

- The humanoid robot "Lightning" completed a half marathon in 50 minutes 26 seconds, a significant improvement from a previous record of over 2 hours 40 minutes just a year ago.  
- In contrast, the top remote-controlled team finished in 48 minutes 19 seconds, but a penalty caused their score to fall behind "Lightning," highlighting the competition's focus on autonomous navigation.  
- Despite the impressive achievements, challenges remain in deploying robots in unpredictable factory environments, underscoring that while the technology has advanced rapidly, many hurdles must still be addressed for widespread use.

Wallstreetcn

HK ROBOTICS

SIASUN

ONEROBOTICS

CATL

BYD COMPANY

SAIC Motor

GAC Group

GAC GROUP

Tesla

CATL HK SDR 30to1

BYD HK SDR 10to1

BYD Company

BYD Company Limited

BYD COMPANY-R

Honor Takes the Crown, Humanoid Robots Beat Humans in Half-Marathon, But the Most Critical Race Isn't in Yizhuang