AgiBot put 110 full-size Expedition A3 humanoid robots through a synchronized routine at the Yellow River International Convention and Exhibition Center in Jinan on October 24, earning a Guinness World Record for the most humanoid robots dancing at the same time1. The stunt was a stunt. But the infrastructure underneath it, a swarm-coordination stack called LinkBots that computed formation changes from live robot state rather than replayed stored motions, is the part that matters for the industry.
ANALYSIS If multi-robot scheduling can hold 110 production-grade humanoids in real-time coordination without a single error over nearly six minutes, the demo proves that fleet-scale scheduling works under controlled conditions—fixed UWB infrastructure, known start and end states, no unpredictable obstacles. It does not prove the software layer is ready for messier environments, but it does establish a baseline that no other humanoid company has publicly matched.
AgiBot's demo and DEEP Robotics' same-day publication of five embodied-AI research advances both center on software rather than hardware. The competitive claims these companies are now making publicly are about coordination stacks, world models, and policy optimization, not actuator torque or joint count. Whether that represents a durable shift in the sector's competitive axis or simply reflects what is novel enough to generate headlines on a given week is an open question, but the emphasis is notable.
Start with what AgiBot actually demonstrated. A Guinness adjudicator confirmed the 110-unit count on site. The robots completed several sets of complex moves over 5 minutes and 43 seconds, switching between synchronized motion and changing formations. Wang Chuang, an AgiBot partner and senior vice president, said coordinating 100 robots is not simple motion copying but a full scheduling stack covering real-time position sensing, distributed decision-making, and dynamic obstacle avoidance. AgiBot says all 110 were standard production units. The robots walked out of their shipping boxes, calibrated their own positions, and ran the routine without engineers steering them one by one by remote control.
Three details in the LinkBots architecture deserve attention, with the caveat that all derive from AgiBot's own statements and IT Home's reporting of those claims. No independent engineer has verified that motions were computed from live state rather than replayed with local correction, nor that the system degrades gracefully rather than simply freezing failed units. The claims are plausible given the observable complexity of the formation changes, but they remain single-source. First, formation changes are computed from each robot's live state rather than replayed from stored motions. Second, AgiBot said ultra-wideband positioning accurate to under 10 centimeters lets robots navigate, walk, and run in coordination while the system tracks their relative positions. Third, AgiBot said separate groups can receive different instructions, with orchestration designed to stop robots interfering with one another. IT Home reported that the system can compensate locally if a single unit drops out.
ANALYSIS If those capabilities perform as described, they represent minimum requirements for fleet deployment in logistics, hospitality, or public services. But the demo's conditions—fixed UWB beacons, a flat convention-hall floor, choreographed routines with known endpoints—are far more constrained than any of those target environments. The gap between scheduling 110 robots through pre-planned formations and coordinating even 10 robots around unpredictable human foot traffic is large and unquantified. AgiBot is already testing that proposition: the company said the robots are being trialed at Jinan subway stations including Olympic Sports Center and Zhuanshan West Road, offering directions, crowd-flow assistance, safety patrols, and cultural guidance. AgiBot said the record-attempt robots will be deployed in tourism performances, public services, exhibition halls, and education.
This is not AgiBot's first fleet deployment. Vector Wire reported in September that AgiBot put 300-plus robots into continuous guest service at a Zhuhai theme park.
ANALYSIS The jump from theme-park service to subway-station trials with crowd-flow assistance represents a step toward less controlled environments, though a subway concourse is still far from an unstructured industrial floor.
The same day AgiBot set its record, DEEP Robotics published five research advances supporting DEEPRobotics-Brain, its embodied intelligence architecture for real-world robot operation2. The five research areas comprise the DR-WM-1.0 world model, the DR-VLA-2.0 embodied foundation model, the DR-APO policy optimization framework, the DR-Agent embodied agent, and the DR-Benchmark data and evaluation platform. The research involves collaborations with researchers from MIT, UC San Diego, the University of Illinois Urbana-Champaign, Tsinghua University, Zhejiang University, the Hong Kong University of Science and Technology, and the Beijing Institute of Technology.
ANALYSIS Where AgiBot's demo is about coordination across many bodies, DEEP Robotics' research package is about intelligence within a single body: predicting action outcomes, translating instructions into motion, improving learned policies, and revising outdated spatial beliefs. The two efforts address complementary layers of the fleet-autonomy problem, though they are separate companies with separate stacks and no integration between LinkBots and DEEPRobotics-Brain has been demonstrated or claimed. The pairing here is analytical, not operational: any company that eventually ships autonomous humanoid fleets will need to solve both layers, whether in-house or through integration. A fleet of humanoids that can hold formation is useful only if each unit can also handle novel objects and changing terrain autonomously.
Several of DEEP Robotics' benchmark results illustrate how far single-agent embodied AI still has to travel. In real-world navigation tests on the LYNX M20 wheel-legged robot, DR-APO raised success from 25% to 65%.
ANALYSIS A jump from 25% to 65% is large in relative terms but still means more than a third of real-world navigation attempts fail, a rate incompatible with unsupervised commercial deployment. In embodied visual navigation post-training, DR-APO raised a pretrained model's simulated success rate from 70.33% to 89.55%. In real bimanual tasks with unseen layouts, the Atomic Motion Coordinate system's normalized task progress reached 87.8%. Adding force feedback raised task progress to 78.5% for a drawer-and-plug task and 75.2% for vase wiping.
These numbers matter because they quantify the gap between demonstration and deployment. AgiBot's 110-robot dance achieved a zero-error standard over a choreographed routine. DEEP Robotics' benchmarks show that once robots face open-ended manipulation or navigation, success rates drop well below the thresholds commercial operators would require. A 65% navigation success rate means fleet coordination is moot: there is no business case for scheduling 110 robots if each one fails a third of its tasks. The industry's near-term constraint is not how many robots you can field but how reliably each one performs when the task is not scripted.
The broader context is a Chinese humanoid sector that is scaling fast. Vector Wire reported that China held 78% of global humanoid robot shipments as volume jumped 432% in H1 2026. AgiBot itself filed for IPO in July 2026 [ANNEX-D1]. XPeng Robotics closed a $900 million round at a $6.3 billion valuation in August. Robotera has weighed a Hong Kong IPO that could raise up to $1 billion.
ANALYSIS Capital is abundant; what separates winners will be the software stack that lets fleets of humanoids operate with minimal human oversight. AgiBot's LinkBots and DEEP Robotics' Brain architecture represent two bets on that layer, though neither has yet demonstrated the reliability required for unsupervised commercial operation.
Several open questions would change this read. AgiBot has described the Jinan subway trials only through its own statements; independent assessments of reliability or passenger interactions would materially alter the picture. DEEP Robotics' benchmark results are self-reported; peer review or independent replication would change their weight. Whether AgiBot publishes operational data from the subway deployments, particularly error rates and human-intervention frequency, is the first concrete marker. Whether DEEP Robotics' DR-APO real-world navigation success rate, reported at 65%, improves substantially in a subsequent publication is the second. The third is whether any Chinese humanoid company demonstrates multi-robot coordination in an unstructured industrial environment rather than a convention center or theme park.