How Humanoid Robots Beat the Human Marathon Record – Inside the 2026 Beijing Race

Event Overview: Robots Beat the Human Marathon Record

On April 19, 2026, the Beijing Yizhuang humanoid‑robot half‑marathon launched in the Beijing Economic‑Technical Development Zone. More than 300 robots from 76 entities across 13 provinces competed. The Beijing Glory robot "Lightning" won with a net time of 50 min 26 s , a three‑fold improvement over the 2025 champion and faster than the human world record of roughly 58 minutes.

Approximately 38% of the teams used fully autonomous navigation—no human pilots or guide rails—demonstrating a shift from "remote‑controlled performance" to true autonomous operation.

Key Metrics (2025 vs 2026)
---------------------------
Teams          ~20 → 100+  (+400%+)
Robots         ~30 → 300+   (+900%+)
Autonomous %   ~5%  → 38%    (+660% )
Champion time  ~3h  → 50m26s (+3×+)
Finish rate    ~60% → 85%+   (significant rise)

Technical Deep Dive: Six Core Challenges

2.1 Motion Control System – Bionic Gait & Dynamic Balance

Long‑distance locomotion for a biped differs from industrial arms because the robot must constantly adjust its posture on uneven ground. The control loop consists of perception, planning, and joint execution.

Technical difficulty analysis:

┌─────────────────────────────────────────────┐
│          Biped Walking Control Architecture   │
├─────────────────────────────────────────────┤
│   ┌─────────────┐   ┌─────────────┐   ┌─────────────┐   │
│   │ Pose Sensor │──▶│ Central Ctrl│──▶│ Joint Actuator│   │
│   │ (IMU+Enc)   │   │ (Real‑time  │   │ (Motor+Gear) │   │
│   └─────────────┘   └─────────────┘   └─────────────┘   │
│            │            ▲            │            │
│            ▼            │            ▼            │
│   ┌───────────────────────────────────────┐   │
│   │ Whole‑Body Dynamics Model (WBC)       │   │
│   │ • ZMP trajectory planning            │   │
│   │ • COM trajectory optimization        │   │
│   │ • Adaptive foot‑placement             │   │
│   └───────────────────────────────────────┘   │
└─────────────────────────────────────────────┘

Key breakthroughs:

Integrated joint module: The champion "Lightning" uses a self‑developed joint that delivers a peak torque of 400 Nm —comparable to an adult’s full‑force soccer kick—by integrating motor, gearbox, encoder, and driver into a single high‑power‑density unit.

Liquid‑cooling system: A high‑capacity liquid pump provides over 4 L/min of heat exchange, keeping motors and electronics within safe temperature ranges during 21 km of continuous operation.

Gait‑optimization algorithm: The gait controller evolved from preset motions to adaptive learning; deep reinforcement learning now lets the robot adjust stride length, frequency, and leg lift height in real time based on terrain.

2.2 Autonomous Navigation – From "Seeing" to "Understanding"

If motion control is the robot’s "legs," navigation is its "brain" and "eyes." The perception‑decision‑execution loop integrates visual SLAM, LiDAR point clouds, IMU data, and BeiDou GNSS.

Perception‑Decision‑Execution loop:

┌──────────────────┐
               │   Perception Layer │
               │ • Visual SLAM      │
               │ • LiDAR point cloud│
               │ • IMU             │
               │ • GNSS (BeiDou)   │
               └───────┬──────────┘
                       ▼
   ┌───────────────────────┐   ┌───────────────┐   ┌─────────────────────┐
   │   Motion Execution    │◀─│ Decision Planning│─▶│   Task Management   │
   │ • Joint torque control│   │ • Path planning │   │ • Odometry          │
   │ • Pose stabilization │   │ • Obstacle avoidance│ │ • Energy management │
   │ • Gait switching      │   │ • Terrain classification│ │ • Fault recovery   │
   └───────────────────────┘   └─────────────────┘   └─────────────────────┘

BeiDou‑based positioning from Qianxun Location provides centimeter‑level accuracy for roughly two‑thirds of the autonomous teams, limiting cumulative position error over the 21 km course to less than a palm’s width.

China Unicom’s 5G‑A dedicated network supplies millisecond‑level latency, enabling real‑time command reception, status upload, and cloud‑edge compute off‑loading.

Terrain‑adaptation tests included a 9° slope, gravel, park ecology paths, and flat asphalt, each stressing different subsystems (balance, tactile perception, visual robustness).

2.3 Energy Management – 21 km Endurance Challenge

Human runners expend 2,000–2,500 kcal over a half‑marathon. Humanoid robots are more efficient but still require high‑capacity batteries for sustained high‑power output.

Robot Model   Battery Type      Theoretical Walk Range   Theoretical Run Range   Race Result
------------  ----------------  -----------------------  ----------------------  ------------
"Lightning"   High‑density Li‑ion   ~30 km                    ~15 km                  Champion (finished)
"Ultra"      Solid‑state          ~35 km                    ~18 km                  Finished
"H1"         Ternary Li‑ion       ~25 km                    ~12 km                  Finished

Energy‑optimization strategies:

Gait energy recovery: Motor back‑EMF recovers kinetic energy during downhill or deceleration phases.

Dynamic power allocation: High torque on climbs, power‑saving coasting on descents, and sprint bursts on flat sections.

Thermal management: Liquid‑cooling keeps batteries within optimal temperature windows.

Hot‑swap battery modules: Some teams replace batteries mid‑race to extend endurance.

2.4 Multi‑Sensor Fusion Perception

┌─────────────────────────────────────────────────────┐
│            Multi‑Sensor Fusion Architecture          │
├─────────────────────────────────────────────────────┤
│   ┌─────────────┐  ┌─────────────┐  ┌─────────────┐ │
│   │  Depth Cam   │  │   LiDAR     │  │   Millimeter │ │
│   │ (RGB‑D, ~30Hz)│ │ (~20Hz)    │ │   Radar (~50Hz)│ │
│   └───────┬─────┘  └───────┬─────┘  └───────┬─────┘ │
│           │                │                │       │
│           └───────┬────────┼────────────────┘       │
│                   ▼        ▼                        │
│          ┌───────────────────────────────────────┐ │
│          │          Sensor Fusion (Kalman, DL)    │ │
│          │ • 3D map construction & localization │ │
│          │ • Semantic understanding (road/obj) │ │
│          │ • Dynamic target tracking            │ │
│          └───────────────────────────────────────┘ │
└─────────────────────────────────────────────────────┘

Challenges included rapid lighting changes, weather variations, dynamic obstacles (staff, other robots, spectators), and terrain classification (grass, cement, plastic track).

2.5 Embodied Large Model – From Preset Programs to Autonomous Decision‑Making

2026 is dubbed the "Embodied Intelligence Year". The VLA (Vision‑Language‑Action) model integrates a vision encoder (ViT + temporal modeling), a language model enhanced for embodiment, and an action decoder.

┌─────────────────────────────────────────────────────┐
│               Embodied Large‑Model Stack            │
├─────────────────────────────────────────────────────┤
│   ┌───────────────────────────────────────────────┐ │
│   │            Foundation Model Layer                │ │
│   │ • Vision encoder (ViT + temporal)               │ │
│   │ • Language model (LLM + embodiment)           │ │
│   │ • Action head (joint torque commands)          │ │
│   └───────────────────────┬───────────────────────┘ │
│                           ▼                       │
│   ┌───────────────────────────────────────────────┐ │
│   │               Task‑Planning Layer               │ │
│   │ • Long‑term task decomposition (e.g., "from start to finish" → "go straight 200 m", "turn left", …) │
│   │ • Real‑time instruction generation (vision → next action) │
│   │ • Failure handling (fall detection → self‑righting) │
│   └───────────────────────┬───────────────────────┘ │
│                           ▼                       │
│   ┌───────────────────────────────────────────────┐ │
│   │               Low‑Level Control Layer           │ │
│   │ • Joint‑space torque commands                  │
│   │ • Foot‑force control                           │
│   │ • Whole‑body coordination                    │
│   └───────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────┘

Key advantages of the end‑to‑end VLA model are:

End‑to‑end learning: Directly maps raw pixels to motor commands, eliminating hand‑crafted perception‑decision pipelines.

Long‑context understanding: Handles 21 km of varied scenes (curves, slopes, obstacles) by maintaining extended temporal context.

Real‑time inference: Maintains >100 Hz control loops, requiring millisecond‑scale model execution.

2.6 Hardware Supply Chain – China’s Integrated Ecosystem

The competition showcased a fully domestic supply chain covering motors, gearboxes, sensors, MCUs, and batteries, with most components achieving >90% domestic content and international‑level performance.

Core Component      Major Suppliers          Domestic Share   Tech Level
-----------------  ----------------------   --------------   ---------
Frame‑less torque motor   Buke, Lvde Harmonic   95%+            International lead
Harmonic gearbox          Lvde Harmonic, Laifu 90%+            International lead
Planetary screw          Nanjing Gongyi, Bot Precision 80%+   International advanced
Depth camera              Orbbec, RealSense      70%+            International advanced
LiDAR                     SuTeng, Hesai          85%+            International lead
High‑performance MCU      Zhaoyi, Lingdong       60%+            Domestic lead
Power battery             CATL, BYD             98%+            International lead

Industry Analysis – Why China Leads the Humanoid‑Robot Track

Omdia reports 2025 global shipments of ~13,000 humanoid units, with China accounting for **90%** of volume and occupying the top six slots on the sales ranking.

2025 Global Humanoid Market Share
--------------------------------
China   90% ███████████████████████████████████████████
USA     6%  ███
Japan   2%  ██
Europe  1%  █
Other   1%  █

Cost advantage in Shenzhen enables a full‑size humanoid for roughly 2 million CNY thanks to:

One‑stop component procurement ecosystem.

Rapid prototyping and iteration enabled by local electronics manufacturing.

Concentration of robotics engineers and control‑algorithm experts.

Substantial municipal subsidies and policy support.

Competitive Landscape – Head‑Runner Technical Roadmaps

Key participants and their technical focuses:

Beijing Glory "Lightning": Integrated joint + liquid cooling (champion).

Fourier Intelligent "Ultra": Full‑size humanoid with self‑developed joints (finished).

Yushu Tech "H1": Open‑source ecosystem + reinforcement learning (finished).

Mi "Iron‑Da": Deep imitation + large‑model perception (finished).

ZhiYuan "Goat": Lightweight (<50 kg) + dual‑arm collaboration (finished).

Technical Route Comparison
---------------------------
[Fourier]  Industrial‑grade reliability → Full‑size (1.7 m+), 400 Nm+ torque joints, industrial use.
[Yushu]    Open‑source rapid‑iteration → Mid‑size (1.8 m), community‑driven, RL fast loops, research/education.
[Mi]       AI‑large‑model empowerment → Full‑size (1.77 m), VLA perception‑decision, home service.
[ZhiYuan]  Lightweight low‑cost → <50 kg, cheap mass‑production, dual‑arm tasks, commercial service.

Application Scenarios – From Track to Factory

Current deployments (progress shown by stars):

Industrial manufacturing: ★★★★☆ – automotive assembly, 3C electronics; market ~500 B CNY by 2028.

Logistics handling: ★★★☆☆ – warehouse sorting, last‑mile delivery; market ~800 B CNY by 2030.

Commercial services: ★★★☆☆ – hotel delivery, restaurant serving; market ~300 B CNY by 2028.

Special operations: ★★☆☆☆ – power‑line inspection, hazardous environments; market ~200 B CNY by 2028.

Home companionship: ★☆☆☆☆ – household chores, elder care; market >1 T CNY by 2030.

Medical rehabilitation: ★☆☆☆☆ – surgical assistance, rehab training; market ~150 B CNY by 2028.

Key challenges to move from "can run" to "can work":

Fine‑grained manipulation – current hand dexterity comparable to a 3‑4 year‑old child.

Generalization – robots excel at narrow tasks but struggle with open‑world adaptability; embodied large models are expected to bridge this gap.

Cost‑benefit – unit cost 0.5–2 M CNY, ROI 3–5 years; mass production needed to lower price.

Safety & reliability – human‑dense environments demand rigorous safety validation, low failure rates, and regulatory standards.

Future Roadmap – 2026‑2030 Outlook

2026 roadmap highlights progressive milestones from motion control to complex task execution, with current achievements marked as ✅, ongoing work as 🔄, and planned breakthroughs as ⏳.

2026 Roadmap
------------
Q1: Motion control breakthroughs – walking, running, jumping.
Q2: Perception‑decision enhancements – vision, navigation.
Q3: Fine‑operation upgrades – dual‑hand dexterity.
Q4: Complex‑task generalization – embodied large models.

Milestones:
✅ Implemented: indoor biped walking, simple autonomous navigation, voice interaction, basic manipulation.
🔄 In progress: rugged terrain traversal, fine manipulation, long‑duration operation.
⏳ Planned: open‑world generalization, multi‑robot collaboration, autonomous learning.

Market forecasts:

End‑2026: industrial deployment worth ~50 B CNY.

2027: mature embodied large models enable complex tasks – market ~150 B CNY.

2028: unit cost drops below 500 k CNY, commercial services expand – market ~500 B CNY.

2029: fine‑operation breakthroughs drive full manufacturing penetration – market ~1 T CNY.

2030: household service robots become mainstream – market >2 T CNY.

Key Data Summary – AI Industry Snapshot

Metric                              Value                Note
-------------------------------------------------------------------
Humanoid half‑marathon champion time 50 min 26 s          "Lightning" result
Robots entered the race            300+                Total participants
Fully autonomous teams             38%                 No human pilots
Global market share (2025)          90% (China)         Omdia data
Cost of a full‑size robot (Shenzhen) 2 M CNY            Approx. 200 万 RMB
AI Hospital diagnostic accuracy    98.5%               Tsinghua Agent Hospital
Smart guide‑dog service users       50 k people‑times   Gaode + Alibaba
Yushu H1 top speed                  10 m/s              2026‑Apr data

Conclusion – From "Can Run" to "Can Work"

The 2026 Beijing Yizhuang humanoid half‑marathon proved that six core technologies—motion control, autonomous navigation, energy management, multi‑sensor perception, embodied large models, and a domestic supply chain—are now viable at scale. While true human‑task replacement still requires advances in fine manipulation, generalization, cost reduction, and safety standards, the technical pathways are clear and commercialisation has already begun.

Upcoming milestones:

End‑2026: large‑scale industrial deployment.

2027: mature embodied large models enable complex task execution.

2028: robot cost falls below 500 k CNY, commercial services proliferate.

2030: household service robots enter the market, unlocking a trillion‑plus opportunity.

2026 marks the official start of the "Embodied Intelligence" era.