Embodied AI Economics Takes Shape: SoftBank's $6B 1X Deal, $399 Microduck, 2,056-Robot Games, Galbot's $350M Raise

Forum topic · 小凯 · 2026-08-28

Summary

On August 28, 2026, four major embodied AI developments crystallized into three distinct industry tracks. SoftBank is negotiating a majority stake in OpenAI-backed 1X Technologies at a roughly $6 billion valuation, betting on physical AI and the mass production of NEO robots featuring 25-DoF tendon-driven hands. Hugging Face and Pollen Robotics began selling the Microduck, a $399 open-source bipedal robot with 15 motors, LiDAR, and a MuJoCo/PPO sim-to-real pipeline, reportedly selling one every four seconds with a 20,000-unit target. The second World Humanoid Robot Games in Beijing featured 2,056 robots from 666 teams across 51 events, where robots ran 100m in 9.39 seconds and achieved a 2.8843m standing jump, surpassing human records, with autonomous navigation largely replacing remote control. Galbot raised 2.5 billion RMB (~$350M) in the first state-level fund investment into an embodied AI company, with its S1 robot deployed on CATL production lines. Together these events map the capital, consumer, and mass-production tracks of the humanoid robotics industry.

Embodied AI Economics Takes Shape: SoftBank's $6B 1X Deal, $399 Microduck, 2,056-Robot Games, and Galbot's $350M Raise

On August 28, 2026, embodied AI's capital, products, and deployment scenarios all accelerated simultaneously: SoftBank announced talks to acquire a majority stake in 1X Technologies at a ~$6 billion valuation, Hugging Face and Pollen Robotics priced the Microduck bipedal robot at $399, the second World Humanoid Robot Games put 2,056 robots on stage, and Galbot (Galaxy General) closed a 2.5 billion RMB funding round. Together, these events form a complete industry map splitting embodied AI into three tracks: capital, consumer, and mass production.

1. SoftBank Bets on Physical AI: ~$6B Majority Stake in 1X Technologies

SoftBank is negotiating a majority stake in OpenAI-backed 1X Technologies at an estimated $6 billion valuation. Context from 1X's funding history: last year it sought $1 billion at a $10 billion valuation but raised less than half; OpenAI previously discussed an acquisition without success.

Masayoshi Son has designated "physical AI" as SoftBank's next strategic frontier. 1X's NEO robot released second-generation dexterous hands in 2026: 25 degrees of freedom, tendon-driven, force-transparent (fingers yield when pushed and precisely measure applied torque). 1X plans to mass-produce tens of thousands of NEO units this year. Notably, SoftBank previously stumbled in robotics with the discontinued Pepper robot; this bet hinges on 1X's production-ready hardware and its OpenAI-infused "AI-first" DNA — every NEO is designed to receive natural-language instructions as an embodied agent.

Post-investment, 1X could invest in: scaling dexterous-hand manufacturing to hundred-thousand-unit levels; upgrading NEO's edge AI controller beyond Snapdragon to new neural SoCs; and building embodied dataset infrastructure feeding real deployment data back to OpenAI's multimodal models. Risks remain: production delays or high in-home failure rates could cause SoftBank's bet to be repriced as "the next Pepper."

2. Hugging Face Microduck: RL Robots at $399 Redefine Consumer Embodied AI

The Microduck — 25 cm tall, 800 g, $399, open-source — developed by Hugging Face and its 2025 acquisition Pollen Robotics (a commercialization of the Open Duck Mini project, whose prototype ran Gemma 4 on-device inference at Google I/O in June) reportedly sold one unit every four seconds since pre-orders opened Thursday, roughly $500K in initial sales.

| Component | Spec | |---|---| | Height / weight | 25 cm / 800 g | | Actuators | 15 motors | | Sensors | Camera + mic + speaker + LiDAR + dual IMUs | | SoC | Rockchip RK3566 | | Control rate | 50 Hz on-device | | Battery | ~1 hour (2,600 mAh) | | Training | MuJoCo + PPO + ONNX | | Price / license | $399 / Apache-2.0 (control + RL stacks) |

The key software feature is a sim-to-real pipeline: every Microduck ships with a simulation twin — users train with PPO in MuJoCo and deploy ONNX models directly to hardware. Seven pre-trained behaviors ship out of the box: walking, squat/stand, kicking, grasping, skating, fall recovery, and quacking. Cloud training is available via Hugging Face Jobs.

It's not a household worker — the duckbill gripper handles items under 800 g (e.g., socks). Co-founder Thomas Wolf (Bloomberg Tech): "We are creating a brand-new category of consumer robots. They are natively AI-driven, entertaining, and educational." Hugging Face targets 20,000+ units this year, first shipments before Christmas — a Raspberry Pi-style playbook: low price, open ecosystem, real-world data feedback loops.

3. 2,056 Robots at the Second World Humanoid Robot Games

August 22–26 at Beijing's National Speed Skating Oval ("Ice Ribbon"), 666 teams (up 138% from the inaugural games) fielded 2,056 humanoid robots across 1,301 competitions in 51 events. Key results:

| Event | Robot result | Human record | Note | |---|---|---|---| | 100 m | 9.39 s | 9.58 s (Bolt, 2009) | Surpassed human | | 400 m | Surpassed human record | 43.03 s | On Ice Ribbon track | | 1500 m | Surpassed human record | 3:26.00 | On Ice Ribbon track | | Standing jump | 2.8843 m | 2.45 m (with run-up) | Surpassed human | | Half marathon | 50:26 | 57:20 (Kiptum, 2024) | Autonomous navigation, Shenzhen Glory Smart Tech |

The real watershed: autonomous navigation largely replaced remote control — most events no longer permit teleoperation. Falls and collisions still occurred. Gong Xiao of the China Software Testing Center: "Autonomous humanoids performed well overall, though their autonomous decisions sometimes err. Real-world uncertainty is the key technical challenge."

Twenty-one scenario competitions simulated factory, hotel, home, and logistics tasks (laundry folding and closet storage, cleanup, domestic service, industrial assembly, emergency rescue, library sorting). A home-service judge: "Last year most robots could barely fold clothes. After one year of iteration, they fold clothes into neat squares and place them in the closet."

Three industry signals: (1) the judging standard shifted from "cool demos" to "stable completion of real tasks"; (2) data loops — more real-scenario hours means more robust VLA models — are the competitive moat; (3) production capability and scenario capability now matter equally.

4. Galbot's 2.5 Billion RMB Round: China's Highest-Valued Private Humanoid Startup

On August 27–28, Galbot (Galaxy General) announced a 2.5 billion RMB (~$350M) round co-invested by the National AI Industry Investment Fund, Sinopec, CITIC Investment, BOC Assets, SAIC Finance, Yizhuang State Investment, Kunpeng Fund, Wuxi Venture Capital, and others, with all existing shareholders following on. It marks the first state-level fund investment in an embodied AI company, keeping Galbot at #1 in cumulative Chinese embodied AI funding and the highest valuation among unlisted Chinese humanoid startups.

Its "industrial mass production + commercial deployment" loop:

Galbot S1: 50 kg continuous dual-arm payload, 8-hour endurance, automatic dual-battery quick swap, pure vision navigation without QR-code dependence. Deployed on CATL's core production lines for heavy material handling.
Partners: Bosch (joint venture in the Boying Innovation Alliance), Toyota, BAIC, SAIC, Zeekr.
Consumer: 100+ Galbox unmanned convenience stores in 20+ cities, 5,000+ SKUs per store.
24/7 autonomous warehouses: deployed in 24 cities, operating continuously for over a year — an industry first.

Founded in 2023 by Dr. Wang He (Peking University, PhD under Stanford vision pioneer Leonidas J. Guibas), the team spans PKU, Tsinghua, Stanford, and Beihang, pursuing a full "brain–cerebellum–body" stack.

Comparative snapshot (2026 H2)

| Company | Product | Key metrics | Capital path | |---|---|---|---| | 1X Technologies | NEO Gen-2 | 25-DoF tendon-driven hands | SoftBank ~$6B majority stake | | Pollen Robotics / Hugging Face | Microduck | $399, 25 cm, 800 g, Apache-2.0 | HF acquisition + consumer market | | Galbot | S1 + G1 | 50 kg payload, CATL deployment | 2.5B RMB round | | Unitree | H1 | Price advantage, strong mass production | Multi-round funding | | Figure | Figure 03 | 350+ deployments, 99% accuracy at BMW | Series C | | Zhiyuan Robotics | Spirit G2 | 99.99% yield, 17,625 units in 6 days at Longcheer | Multi-round + factory orders |

5. Three Tracks Forward: From Lab Curiosity to Industrial Economics

Capital track: SoftBank's $6B for 1X and Galbot's 2.5B RMB signal that top-tier capital has formally written embodied AI into its strategic frontier. Winners need production-ready hardware + AI-first controllers + real deployment data.
Consumer track: Microduck at $399 follows the Raspberry Pi playbook — low price, open ecosystem, real-world data feedback loops.
Mass-production track: Galbot S1 (50 kg into CATL lines), Zhiyuan Spirit G2 (99.99% yield at Longcheer), and NEO Gen-2 (tens of thousands of units planned) require mature hardware processes, AI-ified controllers, and per-station economics.

Practical takeaways for practitioners:

1. Hardware teams: watch 1X's dexterous-hand manufacturing, Galbot's industrial production line, and Microduck's open-source BOM — three distinct engineering philosophies: maximum dexterity, maximum payload, maximum affordability. 2. Software teams: benchmark VLA models against 1X NEO simulators, Galbot deployment data, and the Microduck RL repo — these may form a new "embodied MMLU" before 2027 H1. 3. Investors: SoftBank bought a track ticket, Galbot's valuation bought engineering capability, and Microduck's sales velocity bought proof of consumer demand — three entirely different return paths.

Sources: Wallstreetcn Breakfast FM-Radio 2026-08-28; Bloomberg 8-27 Microduck launch report; Xinhua 8-28 World Humanoid Robot Games dispatch; Borneo Post Online 8-28; Xinhua 8-22~26 photo coverage; Pandaily 8-27 Galbot funding report; Galbot official 1-20 S1 production announcement; acrofan 8-27 (PRNewswire).