The 7 Best-Performing Robots at Beijing’s 2026 World Humanoid Robot Games, So Far
From Tiangong Ultra running 100 metres in 8.86 seconds to AgiBot G2 winning contests built around libraries and emergency response, Beijing’s 2026 World Humanoid Robot Games are revealing two very different versions of robotic progress: machines engineered to break physical limits, and machines beginning to demonstrate that they can actually work.
Beijing’s second World Humanoid Robot Games were always designed to produce spectacle. More than 2,000 robots have descended on the National Speed Skating Oval, better known as the Ice Ribbon, where the programme ranges from sprinting and long jump to martial arts, gymnastics, hotel service and industrial work. The scale has expanded dramatically in a year: the 2026 competition runs from 22 to 26 August, with 2,056 robots from 666 teams representing 16 countries across 51 events.
The eye-catching results have come quickly. Humanoid robots are now producing running times numerically faster than some of athletics’ most famous human world records, although those comparisons should not be mistaken for equivalent World Athletics performances: the machines compete under different rules, starting procedures and mechanical constraints. More revealingly, however, this year’s Games have expanded their emphasis on scenario competitions that attempt to reproduce the mundane physical complexity of actual work. A robot that can run extraordinarily quickly on a prepared track is demonstrating something important about motors, control and dynamic balance; a robot that can locate objects, manipulate them accurately and complete a sequence of unfamiliar tasks is demonstrating a rather different kind of intelligence.
That distinction matters when deciding which machines are actually performing best. This is therefore not an official medal table. It is Nuvastra’s ranking of the standout identifiable robot platforms based on results verified through 25 August 2026, weighing competitive performance alongside autonomy, technical difficulty, repeatability and what the result tells us about the machine beyond the arena. Where published results identify only a team rather than the robot model it used, we have avoided quietly filling in the gap. The Games are still in progress, and the rankings will be updated after the remaining finals on 26 August.
1. Tiangong Ultra — the machine that has changed the meaning of fast
There is little ambiguity about the athletic star of Beijing 2026. On 25 August, Tiangong Ultra ran 100 metres in 8.86 seconds during the large-robot semi-final, according to Reuters, cutting more than half a second from the 9.39 seconds it had recorded earlier in the competition. The number is extraordinary enough without pretending it constitutes an official human athletics record. More significant for robotics is the rate of development: at the inaugural World Humanoid Robot Games in 2025, Tiangong Ultra won the 100 metres in 21.50 seconds. In twelve months, its competition pace has more than doubled.
That improvement sits within a broader display of locomotion performance from the Beijing Humanoid Robot Innovation Center’s Tiangong platform. Teams using the company’s machines have produced a 38.15-second 400 metres and a 2:21.64 1,500 metres, while the centre has also reported a 2.88-metre standing high jump. The numbers make for convenient human comparisons, but the engineering achievement is more interesting: sustained high-speed bipedal locomotion requires the machine to manage balance, joint loading, heat, impact and control repeatedly rather than execute one spectacular movement. Reuters reports that the centre is backed by companies including Xiaomi and Baidu, giving the project industrial as well as sporting significance.
Tiangong Ultra was already intended for precisely this kind of extreme locomotion. The Beijing centre describes it as the performance-oriented member of its current Tiangong platform, designed around high-speed autonomous running, long-distance navigation and a lightweight, rigid body with improved thermal management. Those are manufacturer descriptions rather than independent measurements of commercial capability, but the competition results provide unusually public evidence that its locomotion system works outside a laboratory demo.
The remaining caveat is almost comically physical: running fast and stopping elegantly are not the same problem. Early races produced images of high-speed humanoids continuing into the padded end barriers after crossing the line. It is a useful reminder that a spectacular benchmark can expose the next limitation almost immediately. Tiangong Ultra is nevertheless the clear number one on this list because Beijing 2026 has not merely shown it winning races; it has shown how quickly a specialised humanoid platform can improve once a difficult physical capability becomes an optimisation target.
Visit the official Tiangong Ultra platform page at X-Humanoid
2. AgiBot G2 — the robot whose victories look most like actual work
If Tiangong Ultra is the robot most likely to generate a viral clip, AgiBot G2 may be the machine procurement teams should watch more carefully. AgiBot says its G2 took first place in both the library and emergency scenario challenges on 23 August, while reporting around the Games attributed two golds, three silvers and two bronzes in the early scenario programme to the G2 platform. These are less visually dramatic achievements than a sub-nine-second sprint, but the underlying tasks come considerably closer to the problems commercial humanoids are supposedly being built to solve.
The distinction becomes clearer when looking at the rules. In scenario events, organisers give fully autonomous operation a scoring coefficient of 1.0 while teleoperated completion receives only 0.5. The library contest required robots to perform organised physical work, while the emergency event tested behaviours around a simulated hazardous environment. Beijing organisers have deliberately pushed the programme towards these scenarios because manipulation, perception and long chains of actions remain harder to generalise than running on an unobstructed track.
AgiBot positions G2 as a “universal embodied intelligent robot” built to industrial standards, highlighting force-controlled manipulation, precision assembly and rapid reinforcement-learning deployment. Those specifications are the company’s own claims, and a medal does not prove that G2 can transfer its competition performance unchanged into an uncontrolled warehouse or emergency site. Yet two victories involving useful physical tasks give the platform something many humanoid demonstrations still lack: evidence generated by a competitive environment in which other machines were attempting the same job.
That is why G2 ranks ahead of several visually more athletic robots here. The emerging humanoid industry does not ultimately need machines that can impress an audience for thirty seconds; it needs machines capable of completing work repeatedly when the environment is less cooperative than the demonstration. Beijing’s scenario competitions remain an imperfect proxy for that future, but G2’s performance makes them considerably more interesting.
Visit the official AgiBot G2 product page
3. Galaxea Kengo — gymnastics becomes a test of whole-body control
A humanoid completing a backflip is no longer sufficiently unusual to tell us much by itself. What moved Galaxea Dynamics’ Kengo higher up this ranking was not a single trick but the way its freestyle gymnastics routine connected difficult movements into something closer to a controlled sequence. Competing as the Galaxea Star Fleet team, Kengo won the final with 269 points, narrowly ahead of the Ningxia AgiBot team on 268.25 and Beijing AgiBot on 266.25.
According to Beijing News, Kengo completed ten choreographed actions in less than three minutes, including a 360-degree butterfly kick and a kip-up. Galaxea’s team said professional gymnasts were motion-captured and their movements used as imitation-learning material, after which the robot’s execution was refined. Crucially, once Kengo entered the routine, the machine ran the sequence autonomously rather than being manually steered through each movement.
That makes the routine more technically instructive than a collection of isolated stunts. Connecting one dynamic movement to another forces a robot to manage its changing centre of mass, absorb landing forces, recover its posture and prepare for the next action without returning to a perfectly neutral starting condition. Organisers noted that this continuity was one of the clearest advances over the inaugural Games, where machines more commonly performed individual actions one at a time.
Kengo itself is a roughly 1.4-metre, 40-kilogram biped with at least 23 degrees of freedom, according to Galaxea’s current product page. The company markets it as a commercially oriented embodied-AI platform rather than a dedicated gymnastics machine. Winning a gymnastics competition does not establish its productivity elsewhere, but the result is meaningful precisely because full-body coordination is a foundation that many future humanoid applications will depend upon.
Visit the official Galaxea Kengo product page
4. AgiBot X2 — winning when the track stops being predictable
Straight-line speed gives engineers a relatively clean optimisation problem. The 100-metre obstacle race, by contrast, introduces the kind of disruptions that quickly expose a robot whose locomotion policy works only when the world behaves itself. On the evening of 24 August, AgiBot’s Lingxi X2 won the event, giving AgiBot what was reported at the time as its eighth gold medal of the Games.
The course included narrow turns, uneven surfaces, restricted passages and multiple obstacles that required the robot to change its posture and path while maintaining speed. Organiser briefings described features including an L-shaped restricted section, crawling passage, S-bend bridge, spiral steps, slopes, slalom elements and hurdles. This transforms the exercise from a simple question of actuator power into one involving perception, terrain adaptation, dynamic balance and rapid control decisions.
AgiBot told Chinese media that X2 competed using the production version of its body without physical modification for the race, and attributed its performance to the company’s AGILE locomotion framework and visual perception system. The “unmodified production version” description is a company claim rather than something independently disassembled and verified at the Games, so it should be treated accordingly. Even with that qualification, winning a course designed to upset a robot’s rhythm is a more revealing locomotion test than a pristine sprint lane.
This is why X2 sits just below Kengo. Both machines are demonstrating that modern humanoid control is moving beyond isolated motions towards continuous adaptation. Kengo did it through a highly choreographed routine; X2 did it while the physical environment was actively trying to knock it off balance.
Visit the official AgiBot X2 product page
5. UniX AI Panther — a hotel-service gold with more commercial relevance than spectacle
The hotel guest-service final will not generate the same international headlines as a robot beating a famous sprint time, but UniX AI’s Panther belongs in this ranking because it is competing in a category closely aligned with the company’s stated commercial strategy. Beijing News reported that the UniX AI “Black Panther” team won the hotel guest-service final, while the company says its Panther series also placed third in a fire-response scenario.
The result is more interesting in context. UniX AI says Panther is already being deployed in hotel environments for duties including reception, delivery and cleaning, and the company has deliberately chosen a wheeled dual-arm configuration rather than pursuing a fully bipedal design for every application. Its current Panther platform combines an omnidirectional four-wheel base with vertically adjustable dual arms, a design that sacrifices the anthropomorphic purity of walking on two legs in favour of stability and working reach. Those statements about deployments and product capability come from UniX AI itself and should be read as company claims, not independent audits.
That architectural choice is worth paying attention to. The humanoid industry has become fascinated with machines that reproduce the human silhouette as closely as possible, but many workplaces were designed around human reach and manipulation rather than an absolute requirement for human legs. A wheeled humanoid that reliably moves through a hotel and handles objects could create more immediate economic value than a beautiful biped that spends much of its energy solving the mechanics of walking.
Panther’s gold is therefore interesting for what it does not prove as much as for what it does. A controlled Games scenario cannot establish hotel-wide reliability over thousands of hours, and UniX AI’s claims about mass production and deployment require separate scrutiny. What the result does provide is a competitive test consistent with the job the machine is being sold to perform. At this stage of the humanoid market, that alignment is valuable.
Visit the official UniX AI Panther product page
6. AgiBot A3 — Tai Chi turns balance into a serious benchmark
Calling a robot a “Tai Chi master” is an irresistible headline and a poor technical description. What makes AgiBot A3’s gold medal in the Tai Chi competition interesting is that the routine demanded repeated transitions through configurations that are particularly unforgiving for a humanoid: single-leg support, low centres of gravity, rapid rotation, airborne movement and stable landings. The Shanghai AgiBot team won the event, and subsequent reporting identified the machine as the company’s A3.
The routine included combinations reported as an airborne kick followed by a 450-degree outside crescent movement, a 360-degree rotation into horse stance and a 540-degree rotation into another horse stance. These are demonstrations of full-body motion control rather than evidence that A3 possesses a human understanding of martial arts, but maintaining stability across those transitions is precisely the kind of challenge that reveals the quality of a robot’s joint control and Sim2Real training.
AgiBot says A3 uses a lighter 55-kilogram body, high-power proprietary joints and a dual-battery architecture, while marketing the platform primarily around performance, interaction and expressive movement. Those specifications are manufacturer-supplied. The gold medal itself, however, gives A3 an independently observable competitive result that corresponds neatly to those design priorities.
A3 ranks below Panther because its victory demonstrates physical capability more clearly than economically useful autonomy. Yet the result still matters. Humanoid robots cannot become genuinely useful in spaces built for people if they constantly require generous safety margins around balance, turning and posture. Tai Chi may look like entertainment, but the underlying control problem is serious.
Visit the official AgiBot A3 product page
7. HONOR Robotics D1 “Lightning” — the challenger forcing Tiangong to keep accelerating
HONOR Robotics D1, better known as Lightning, has not displaced Tiangong Ultra at the top of this ranking, but it has helped make the sprint competition genuinely competitive. Reuters reported a 9.47-second 100-metre competition run from Lightning during the early stages of the Games, while HONOR had recorded a 9.32-second run during pre-competition testing. The distinction is important: the faster number was a test, not the equivalent of Tiangong Ultra’s subsequent 8.86-second competition semi-final.
Lightning is not a new arrival. Earlier in 2026, the robot — formally identified in Chinese reporting as HONOR Robotics D1 — completed the Beijing E-Town humanoid half-marathon in 50 minutes 26 seconds, while HONOR’s own robotics page presents D1 as its speed-and-stability platform. Competition conditions and weighting mean robot-versus-human record claims need qualification, but the combination of long-distance endurance and extreme short-distance pace makes D1 one of the more intriguing locomotion projects in the field.
There is also a strategic curiosity here. HONOR is primarily a consumer-electronics company, not an established industrial robotics manufacturer. Its move into humanoids suggests that the technologies required for mobile devices — batteries, sensing, computing, thermal design and increasingly on-device AI — are beginning to overlap with the engineering stack required for embodied machines. Whether HONOR turns that overlap into a durable robotics business remains unanswered, but Lightning has already ensured that Beijing’s running events are not a Tiangong demonstration disguised as a competition.
Its seventh-place position reflects the cutoff date rather than a lack of capability. The decisive 100-metre stages are still being completed, and a strong final result on 26 August could move Lightning considerably higher.
Visit HONOR’s official Robotics D1 page
What Beijing’s best robots are actually telling us
The temptation after Beijing is to construct a single technological hierarchy from the medal table: one robot is first, another second, and therefore one company must be ahead. The Games themselves make that interpretation increasingly difficult. Tiangong Ultra is operating at a level of locomotion performance that would have looked implausible at the inaugural competition only a year ago, while G2’s strongest evidence concerns manipulation and task completion. Kengo excels at chained whole-body movement, X2 at adapting locomotion to obstacles, Panther at service work and A3 at controlled high-dynamic movement. They are competing under the umbrella term “humanoid robot” while solving substantially different engineering problems.
That is also why the scenario competitions may ultimately matter more to the industry than the record-breaking races. Running on a prepared track rewards optimisation around a relatively narrow physical objective. Workplaces are hostile in more ordinary ways: objects move, cables bend, shelves differ, paper slips, humans interrupt and a task that succeeds 90 per cent of the time may still be commercially useless. Organisers have recognised this by giving autonomous completion a significant scoring advantage in scenario events and by expanding the programme into factories, hotels, homes, offices and emergency environments.
For readers tracking the companies behind embodied AI rather than only the competition, Nuvastra’s Global AI Company Directory includes a dedicated Robotics & Embodied AI category covering businesses building physical AI systems around the world.
Beijing 2026 is consequently producing two races at once. The visible one is measured in seconds, metres and medals, and Tiangong Ultra is currently setting its pace. The more consequential race is towards a machine that can enter an ordinary human environment, understand what needs doing and finish the job without an engineer standing nearby. No robot at the Ice Ribbon has demonstrated that general capability yet. What has changed is that a growing number of them are beginning to demonstrate pieces of it under competitive pressure.
That makes the 2026 Games more revealing than the usual robotics showcase. The falls, failed grasps and awkward recoveries are not distractions from the technology; they are part of the evidence. The best-performing robots in Beijing are impressive not because they make humanoid robotics look finished, but because their radically different strengths show more clearly what still has to be solved.
FAQ
What is the fastest robot at the 2026 World Humanoid Robot Games?
As of results verified on 25 August 2026, Tiangong Ultra has produced the fastest clearly reported 100-metre competition time, running 8.86 seconds in the large-robot semi-final. The final stages of the Games were still in progress when this article was prepared, so that figure should be rechecked after competition concludes on 26 August.
Did Tiangong Ultra really beat Usain Bolt’s world record?
Numerically, yes: 8.86 seconds is faster than Usain Bolt’s 9.58-second human 100-metre world record. It should not, however, be described as replacing Bolt’s World Athletics record because humanoid robots compete under different rules and conditions. The useful comparison is how rapidly robot locomotion is improving, particularly given Tiangong Ultra’s 21.50-second winning time at the 2025 Games.
Which robot has performed best at real-world tasks?
AgiBot G2 has one of the strongest cases among identifiable platforms so far, with golds reported in both library and emergency scenarios. Panther’s hotel-service victory is another notable result because it aligns closely with the commercial service environments UniX AI says it is targeting. These events are particularly informative because the organisers reward autonomous rather than teleoperated completion.
How many robots are competing in Beijing in 2026?
The organisers say 2,056 robots from 666 teams and 16 countries are participating in 51 events. The second Games run at Beijing’s National Speed Skating Oval from 22 to 26 August 2026.
Are the World Humanoid Robot Games only about sport?
No. The 2026 programme deliberately combines athletic competition with scenario-based tests covering industrial, hotel, household, logistics, office and emergency work. The latter are intended to assess practical functions including perception, manipulation, reliability and autonomous task execution rather than speed or spectacle alone.
Is this the final ranking for the 2026 Games?
No. This article ranks the strongest robot-level performances that could be verified through 25 August. Competition runs through 26 August, including remaining finals, so results should be updated once the Games have officially concluded.