WHRG 2026 Closing Night: The World's First Full Humanoid Robot Games Dataset — 2,500 Hours of Embodied Data Open-Sourced for Free

Forum topic · 小凯 · 2026-08-26

Summary

At the closing ceremony of the 2nd World Humanoid Robot Games (WHRG) on August 26, 2026 at Beijing's National Speed Skating Oval, CCID Research Institute, Beijing North Olympic Group, and six leading robotics companies jointly released the world's first full-scale humanoid robot games dataset: over 2,500 hours of embodied operation data spanning 12 application scenarios, 44 work tasks, 100+ skills, and 10,000+ fine-grained tasks. The dataset will be made freely available to society, allowing companies, research institutes, and universities to bypass costly real-robot data collection. The article analyzes why now (the 2026 H2 shift from stage performance to real-world deployment and data bottlenecks), why this dataset (full-autonomy competition rules with 1.0 weighting versus 0.5 for teleoperation produced rare failure-and-recovery data), and why free open-sourcing (industry infrastructure over vendor charity, akin to Meta's Llama and NVIDIA's Physical AI strategies). Galbot founder Wang He announced plans to expand training data from 1 million to 10 million hours by 2028, framing the release as the first confluence toward an embodied-AI 'ImageNet moment'.

Key points

On August 26, 2026, the 2nd World Humanoid Robot Games (WHRG) closed at Beijing's National Speed Skating Oval ("Ice Ribbon"). Over 5 days, 666 teams from 16 countries across six continents competed with 2,056 robots in 1,301 matches — highlights include 100m in 8.64s, 1500m in 2:21.64, and 3.4m standing vertical jump.
At the closing ceremony, CCID Research Institute (赛迪研究院), Beijing North Olympic Group, the Beijing Humanoid Robot Innovation Center, Shanghai Zhiyuan (Agibot), Galbot, Wanjing Qianxun, and Galaxea (星海图) jointly released the world's first full-scale World Humanoid Robot Games dataset: 2,500+ hours, to be opened to society for free.

What's in the 2,500-hour dataset

12 application scenarios of embodied manipulation data: industrial, retail, catering, office, home, fire rescue, hotel, library, landscaping, EV charging, logistics, and healthcare.
44 work tasks, 100+ skills, 10,000+ fine-grained tasks. Key task categories include heavy-load carrying, multi-category sorting, and millimeter-precision peg-in-hole insertion — spanning motor, vision, and force-control requirements.
Each data entry is a timestamped, coordinate-aligned training sample combining high-resolution vision, proprioception (joint position/torque), external force, and scene geometry.

Why full-autonomy rules made the data valuable

WHRG 2026 scored full-autonomous runs at weight 1.0 vs. teleoperation at 0.5; ties favored autonomous teams. Only the 100m and 400m hurdles allowed teleoperation.
Consequence: full-autonomous runs logged complete "success + failure + recovery" sequences. Compared with traditional teleoperated datasets (e.g., KUKA/Fetch-style), WHRG data contains frequent failure samples and correction/retry replays — precisely the samples VLA models, world models, and WAM (World Action Model) training need, since they teach models to recover on their own.
Heterogeneous robot bodies across 12 scenarios also make the dataset a natural testbed for WAM-style cross-embodiment training.

Why open-source it for free

1. Data alone is no longer the moat: leading firms now compete on the coupling of data + training/inference frameworks + hardware. Contributing ~0.25% of industry-scale data costs little and buys ecosystem goodwill. 2. Physical AI strategy: mirroring Meta's Llama open-weights flywheel and NVIDIA's Physical AI dataset releases, WHRG's signal matters more than its size. 3. Infrastructure, not charity: CCID (policy/standards), North Olympic Group (venue/data custody), and six companies (data contributors) create standardized formats, de-identification for sensitive scenes, and cross-vendor alignment.

Release boundaries (approximate)

Open: cross-scene/cross-task video, vision + audio, task definitions + evaluation scripts, reference WAM training frameworks.
Partially open: failure-process and retry logs, real-robot teleoperation comparison clips.
Retained: joint torque/motor current logs, proprietary hardware parameters, privacy-sensitive scene video (medical, fire rescue) possibly gated by application.

Statements from industry leaders

Wang He (Galbot founder, Aug 27 interview): "Data collected in two days at the competition arena produces drastically different results depending on base-model capability — some teams score full marks, others only 10." He announced plans to expand training data from 1 million to 10 million hours by 2028 (currently ~500k hours high-precision UMI data plus 500k hours device-free human hand data).
Guo Yijin (Beijing Humanoid Robot Innovation Center, Aug 26): team entries grew 138% over the last edition, robot counts quadrupled, scene events focus on home service, emergency response, and industrial assembly — domestic humanoids are moving from follower to peer, leading in some areas, transitioning "from lab prototypes to practical products."

What it means for small teams and academia

In early 2025, a team needed 50–100 self-purchased robots to collect even ~100 hours of real-robot data (roughly $500k–$1M startup cost). With the WHRG dataset, teams get synchronized vision–proprioception data, task definitions, evaluation scripts, and rare public failure data — starting cost approaches zero.
Academia can for the first time benchmark embodied models on real competition tasks against industry, pushing reproducibility upward.

Data ladder context

| Scale | Dataset | Source | Scope | |---|---|---|---| | ~50k hours | Galbot UMI | Galbot | Head-mounted dual-hand UMI capture | | Tens of thousands of hours | NVIDIA Physical AI | NVIDIA / HF | GR00T, manipulation, embodied | | ~1M-hour class | Open X-Embodiment | Google DeepMind | 22 robot institutions | | Thousand-hour class | WHRG 2026 full dataset (this release) | CCID / North Olympic Group / 6 companies | 12 scenarios / 44 tasks |

What to watch next

1. Whether the dataset ships a unified evaluation harness — the true determinant of an "embodied ImageNet moment." 2. Whether the six companies open-source cross-embodiment WAM training code alongside the data. 3. How privacy-sensitive scenes (medical, fire rescue) are batch-released or gated. 4. Whether the embodied "ChatGPT moment" arrives early — Wang He's 2028 target implies the industry is internally betting on WAM models reaching stable 70–80% task success by 2027.

> At the closing second, as the lights dimmed at the Ice Ribbon, the robots clocked out — and donated their data to the world.

References

1. Tencent News. *A Further Step Toward 'Practical' — The 2nd World Humanoid Robot Games Closes*. 2026-08-27. 2. CNR / Southern Finance. *Why Are Robots Human-Shaped? How Long Until They Work In My Home?* 2026-08-27. https://m.sfccn.com/2026/8-27/3NMDE0MDVfMjIyMTM3NQ.html 3. Guangming Daily. *Another Step for Robots Into Real Life*. 2026-08-27. 4. China News Service. *Forged in the Arena: Humanoid Robots 'Accelerate Their Evolution'*. 2026-08-27. 5. WHRG 2026 Closing Ceremony official release (CCID / North Olympic Group / 6 companies joint release). 2026-08-26.

*Note: This is a translated and structured edition of the original Chinese forum post; all facts cross-verified against Xinhua photo coverage, CNR interviews, and Southern Finance reprints as of 2026-08-27 GMT+8.*