World Humanoid Robot Games Closes with Free 2,500-Hour Embodied AI Dataset Release
At the closing ceremony of the second World Humanoid Robot Games (WHRG) on August 26, 2026, CCID (赛迪研究院), Beijing Olympic Group, and leading robotics firms including the Beijing Humanoid Robot Innovation Center, AgiBot (上海智元), Galbot (银河通用), and others jointly announced the world's first full-volume humanoid robot games dataset — more than 2,500 hours of real-robot embodied manipulation data — to be released free of charge to the public.
Key points
- Scale of the Games: 666 teams from 16 countries across six continents fielded 2,056 robots completing 1,301 events over 5 days, including a 100m sprint record of 8.64s and a 1500m record of 2:21.64.
- Dataset contents: 12 application scenarios (industrial, retail, catering, office, home, firefighting/rescue, hotel, library, landscaping, EV charging, logistics, medical), 44 task types, 100+ skills, and 10,000+ fine-grained tasks.
- Autonomy rule drove data quality: fully autonomous runs scored at weight 1.0 vs. 0.5 for teleoperation (with ties broken in favor of autonomous teams). Only the 100m/400m hurdles allowed teleoperation. This produced abundant failure + recovery + replanning data — more valuable for VLA/world-model training than human demonstrations.
- Why open-source: the post argues that real moats have shifted from data alone to the coupling of data + training/inference frameworks + hardware. Opening the data functions like Meta's Llama or NVIDIA's Physical AI strategy: building an ecosystem flywheel around "physical AI."
- Governance structure: a neutral coordinator (CCID for standards/policy, Beijing Olympic Group for venue/data custody) plus six corporate data contributors standardizes formats, alignment, and privacy de-identification across heterogeneous robot bodies — framing the dataset as industry infrastructure rather than vendor charity.
- For small teams: free access to real-robot synchronized vision + proprioception data, evaluation harnesses, and training frameworks cuts embodied-AI startup costs (previously ~$500k–$1M to collect even 100 hours) to near zero.
- For academia: researchers can benchmark models on the same real-world task sets as industry for the first time.
- Data scaling targets: Galbot founder Wang He stated the company holds 500k hours of high-precision UMI data plus 500k hours of human-hand data (1M total), and plans to scale training data from 1M to 10 million hours by 2028, betting on an embodied "ChatGPT moment" at 70–80% task success rates.
- Industry context: entries grew 138% over the first Games; Beijing Humanoid Robot Innovation Center's Guo Yijie noted domestic humanoid robots are moving from "lab prototypes" to "practical products" across real scenarios.
What is open (and what is not)
| Level | Status | |---|---| | Cross-scenario/task video, vision + audio, task definitions + evaluation scripts, WAM training reference implementations | Open | | Failure processes + retry records, teleoperation comparison segments | Partially open | | Joint torque/motor current, proprietary hardware parameters, privacy-sensitive scene video (medical, firefighting) | Retained / restricted |
Why it matters
What to watch next
1. Whether the released dataset ships a unified evaluation harness (the key to an "embodied ImageNet moment"). 2. Whether cross-vendor WAM/VLA training frameworks are co-released with the data. 3. How privacy-sensitive scenario data (medical, firefighting) will be batch-released or gated by application. 4. Whether the embodied "ChatGPT moment" arrives ahead of the 2028 target, in 2027.
*The 2,500 hours sit mid-way in the embodied data ladder — below Galbot's UMI collection and NVIDIA's Physical AI datasets in some categories, but uniquely positioned as competition-grade, fully-autonomous, multi-body, real-scene data bridging "real competition" and "real factory" domains.*