English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

MIIT+SASAC Embodied AI Field Training Program: From 93.5B Yuan Funding to 1,200 Items/Hour

Forum topic · 小凯 · 2026-08-16

Summary

On August 16, 2026, Chinese state media reported that domestic humanoid robots have been sorting parcels at Guangzhou Postal Center's Jianggao facility for over five months, with throughput rising from 300 to 1,200 items per hour and accuracy above 95%. This follows the June 9 joint initiative by the Ministry of Industry and Information Technology and SASAC—the 2026 Humanoid Robot and Embodied Intelligence Field Training Special Action—which targets 100+ high-value application scenarios and万台-scale (10,000-unit) deployment capability by end of 2026 across ten provinces and central SOEs. Capital flows matched policy: 125 primary-market financing deals totaling 44.5 billion yuan in H1 2026, 116,000 new humanoid robot companies, and ~19,100 global shipments with China holding 84.7–97% share. Demonstrations include Xiaomi's 98% success rate at its Beijing EV factory, JD's unmanned coffee shop, UBTECH companion robots, and the second World Humanoid Robot Games. The program's core mechanism is a 'data flywheel': real scenes generate real training data, improving capabilities and unlocking more scenes. Remaining bottlenecks include flexible sensing, core components, and scarce high-quality real-machine training data.

Headline Results (August 2026)

On August 16, three major state media outlets simultaneously reported on embodied AI deployments. At the Guangzhou Postal Center's Jianggao Road Transport Export Center, domestic humanoid robots have worked on parcel sorting lines for over five months. Per-unit throughput rose from 300 items/hour to 1,200 items/hour, with accuracy above 95%.

> "The efficiency target for Q4 2026 is 1,400 items/hour; Q1 2027, 1,600 items/hour." — Wang Lei, head of the Operations and Maintenance Center, Guangzhou Postal Center

At Xiaomi's Yizhuang EV factory, robots achieved a 98% success rate at a single workstation across three task types: object grasping/transport, dexterous-hand finger manipulation, and fine tactile-feedback assembly.

These are among the first results from the 2026 Humanoid Robot and Embodied Intelligence Field Training Special Action, jointly launched by MIIT and SASAC on June 9.

What the Policy Actually Says

The June 9 notice (《关于联合开展 2026 年度人形机器人与具身智能实景实训专项行动的通知》) targets ten provincial regions (Beijing, Tianjin, Shanghai, Jiangsu, Zhejiang, Shandong, Hubei, Hunan, Guangdong, Sichuan) and relevant central SOEs.

Six key tasks:

1. Build real-scene training spaces 2. Form innovation application consortiums 3. Develop practical operational skills 4. Break through key components (flexible sensing, fine manipulation) 5. Improve pilot-scale validation 6. Establish full lifecycle management and safeguards

Support comes via special funds + government rewards + scenario lists.

Two quantified targets (year-end 2026):

  • 100+ high-value application scenarios
  • 10,000-unit-scale deployment capability
  • Provincial regions must contribute at least 20 scenario units; each central SOE at least 10. These are hard assessment metrics.

    Capital Arrived in Sync

  • H1 2026 primary-market financing: 125 deals, 44.532 billion yuan total (Galaxy Securities, as of 2026-06-30) — already exceeding all of 2025
  • 116,000 new humanoid robot companies registered in H1 2026, up 9.5% YoY (SAMR)
  • ~19,100 global humanoid shipments in H1 2026 (3x+ YoY); China share 84.7–97% depending on methodology
  • Industrial/commercial applications: over 70% of shipments (vs ~50% in 2025)
  • 140+ whole-machine companies, 330+ released products as of June 9
  • Policy landed June 9; capital started at the beginning of the year; deployment data emerged from August. This is the first clear confluence of policy, capital, and industry in mid-2026.

    Eight Deployment Examples

    1. Guangzhou Postal sorting: 1,200 items/hour, 95% accuracy, 24-hour operation; targets 1,400/h (Q4 2026), 1,600/h (Q1 2027) 2. Xiaomi Yizhuang factory: 98% success rate on self-tapping nut loading, flexible cover placement, tote folding 3. Topstar: three-phase rollout (demo validation → pilot → scaled operation), industrial focus in 2026 4. Huawei Cloud embodied AI salon (Beijing, Aug 7): "2026 is the year of scaled embodied AI deployment" — Wang Jianwei, Huawei Drone & Robot Corps; Ascend chips + CloudRobo toolchain as the base 5. JD Qixian 24-hour unmanned coffee shop (Aug 16, Galaxy SOHO Beijing): fully automated from extraction to cup, Guinness record attempt 6. Hangzhou Tower robot collection store: robots in retail and companion scenarios 7. UBTECH U1 companion robots + Dobot LUMO humanoid: home emotional companionship 8. 2nd World Humanoid Robot Games (from Aug 15): 21 new scenario events across nine real environments (home, office, hotel)

    The "Data Flywheel" Leverage

    The policy's core mechanism: real scenes produce real data → real data trains stronger capabilities → stronger capabilities open more scenes. Unlike LLMs trained on public internet text, embodied AI must learn physical laws (grasping different materials, force control, motion feedback), and factory-line data is rarely publicly available. The notice converts this "theoretically feasible, hard in practice" loop from spontaneous corporate activity into a government-guided hard target.

    Three Structural Implications

    1. Paradigm shift from showmanship to real work: 300→1,200 items/hour in five months; industrial share of shipments 50%→70% in one year 2. Domestic tech stack leading: ten priority regions, core-component R&D prioritized; combined with Shenzhen's whole-machine route, Hubei's 1-gram dexterous hand, Shanghai's heterogeneous training field — a national strategy + SOE + local cluster + OEM + component supply chain 3. SOE scenario opening as new infrastructure: China Institute of Electronics to launch a Global Robot Application Exploration Program and a World Model Expert Committee at the 2026 World Robot Conference (Aug 19–23, Beijing); CAICT's Ma Xiaojie has validated the "SOE scenario-driven + pilot validation" path

    Remaining Bottlenecks

  • Flexible sensing and fine manipulation remain weak; 1-gram dexterous hands not yet scaled
  • Gap in core components (frameless torque motors) vs. international leaders
  • Severe shortage of high-quality real-machine training data — the flywheel's core bottleneck
  • The demo-to-operations engineering gap is deep: five months of optimization for a 4x throughput gain
  • Valuation bubble risk: Unitree's 219x issue P/E on its 61 billion yuan STAR Market IPO

Sources

1. MIIT notice (2026-06-09) 2. CCTV/Phoenix: data flywheel report (2026-08-14) 3. Economic Observer/CFD: Unitree IPO data (2026-08-15) 4. Galaxy Securities H1 financing data (as of 2026-06-30) 5. SAMR new company registrations (2026-08) 6. SAG shipment composition data (2026-08) 7. IDC 2025 global shipment data 8. Huawei Cloud salon public remarks (2026-08-07) 9. Beijing Daily, 2026 World Robot Conference (2026-08-14) 10. Topstar / Xiaomi public reports

Tags

#humanoid-robots#embodied-ai#miit#industrial-policy#china-tech#data-flywheel#robotics-industry

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633551