SoftBank's $6B 1X Move, $399 Microduck, WRC's 2,056 Robots, Galbot's RMB 2.5B Raise: Embodied AI's 'Three Tracks' Form in a Single Day
On August 28, 2026, capital, product, and deployment in embodied AI all shifted into high gear: SoftBank announced plans to take majority control of 1X Technologies at a ~$6B valuation, Hugging Face and Pollen Robotics put the Microduck bipedal robot on sale at $399, the second World Humanoid Robot Games concluded with 2,056 robots competing, and Galbot completed a RMB 2.5 billion funding round. These are not four isolated stories — together they form the first complete map of the industry split into a capital track, a consumer track, and a mass-production track.
1. SoftBank writes 'Physical AI' into its strategic frontier: $6B for 1X Technologies
SoftBank is negotiating to acquire a majority stake in OpenAI-backed 1X Technologies at an estimated $6 billion valuation. The number looks aggressive but fits 1X's trajectory: last year 1X sought $1B at a $10B valuation and raised less than half its target; OpenAI previously discussed an acquisition without a deal.
SoftBank's Masayoshi Son has defined 'Physical AI' as the company's next strategic frontier. 1X's NEO robot shipped a second-generation dexterous hand in 2026 — 25 degrees of freedom, tendon-driven, force-transparent (fingers yield naturally and measure applied torque when pushed). 1X is preparing to mass-produce tens of thousands of NEO units this year. The bet reflects two things SoftBank values: production-ready hardware, and the 'AI-first' DNA instilled by OpenAI — each NEO is designed to receive natural-language instructions directly. 1X sells not 'a machine that walks' but 'LLM-commandable labor.'
Risks remain: if mass-production ramps slip or NEO's failure rate in real homes exceeds thresholds, SoftBank's $6B could be repriced as 'the next Pepper.'
2. Selling an RL robot for $399: Microduck redefines consumer-grade embodiment
Hugging Face and Pollen Robotics' Microduck — 25 cm tall, 800 g, $399 — sold roughly one unit every 4 seconds after pre-orders opened, reaching about $500K in sales within days. A commercialized version of the Open Duck Mini project, it was shown running on-device Gemma 4 inference at Google I/O in June.
| Spec | Value | |---|---| | 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 | | License | Control & RL stacks: Apache-2.0 |
The killer feature is the sim-to-real pipeline: each unit ships with a simulation twin, users train PPO policies in MuJoCo, then deploy ONNX models directly to hardware. Seven pre-trained behaviors ship out of the box (walking, squatting, kicking, grabbing, skating, fall recovery, quacking). Cloud training is available via Hugging Face Jobs.
Microduck is not a household chore robot — its duck-bill gripper handles small sub-800 g objects. It is a low-cost physical AI experimentation platform. Co-founder Thomas Wolf: "We are creating an entirely new consumer robot category — natively AI-driven, entertaining, and educational." The plan: 20,000+ units this year, first shipments before Christmas, echoing Raspberry Pi's playbook of seeding a category with cheap open hardware.
3. 2,056 robots at the starting line: the 2nd World Humanoid Robot Games
From August 22–26 at Beijing's National Speed Skating Oval and nearby venues, 666 teams fielded 2,056 humanoid robots across 1,301 events in 51 disciplines — a 138% increase in teams over the inaugural Games.
| Event | Robot result | Human record | Note | |---|---|---|---| | 100 m | 9.39 s | 9.58 s (Bolt, 2009) | Surpasses human | | 400 m | Beyond human record | 43.03 s | — | | 1,500 m | Beyond human record | 3:26.00 | — | | Standing high jump | 2.8843 m | 2.45 m (with run-up) | Surpasses human | | Half marathon | 50:26 | 57:20 (Kiptum, 2024) | Autonomous navigation |
The watershed is not the results but the replacement of teleoperation with autonomous perception-decision-action loops. Falls and collisions remain frequent. As China Software Testing Center deputy director Gong Xiao put it: "Autonomous humanoids performed well overall; their autonomous decision-making still sometimes errs. Real-world uncertainty is the key technical challenge."
Three industry signals: (1) judging criteria have shifted from flashy demos to stable completion of real tasks; (2) data flywheels — real-scenario hours — are the competitive moat for VLA models; (3) mass production and scene capability now matter equally.
4. Galbot's RMB 2.5B raise: China's most valuable private humanoid company extends its lead
Galbot announced a RMB 2.5 billion (~$350M) round backed by the national AI industry investment fund, Sinopec, CITIC, BOC Asset Management, SAIC Finance, and others — the first time a state-level fund has invested in an embodied AI company. Cumulative funding now leads China's embodied AI sector.
- Galbot S1: 50 kg continuous dual-arm payload, 8-hour runtime, automatic dual-battery swap, pure visual navigation (no QR codes). Deployed in CATL's core production lines for heavy material handling.
- Partners: Bosch (JV), Toyota, BAIC, SAIC, Zeekr.
- Consumer: 100+ unmanned Galax Capsule stores across 20+ cities, 5,000+ SKUs per store.
- 24/7 autonomous warehouses in 24 cities, running over a year continuously — an industry first.
| Company | Product | Key metrics | Capital path | |---|---|---|---| | 1X Technologies | NEO Gen-2 | 25-DoF hand, tendon-driven | SoftBank $6B control stake | | Pollen / Hugging Face | Microduck | $399, 25 cm, Apache-2.0 | HF acquisition + consumer | | Galbot | S1 + G1 | 50 kg payload, CATL deployment | RMB 2.5B round | | Unitree | H1 | Price and volume leadership | Multiple rounds + A-share | | Figure | Figure 03 | 350+ deployments, 99% BMW accuracy | Series C | | Zhiyuan (Agibot) | Sprite G2 | 99.99% yield, 17,625 units in 6 days | Multiple rounds + factory orders |
5. Three tracks: embodied AI moves from lab spectacle to industrial economics
Capital track: SoftBank's $6B for 1X and Galbot's RMB 2.5B show top-tier capital writing embodied AI into strategic frontiers. Winners need production-ready hardware + AI-first controllers + real deployment data.
Consumer track: Microduck brings reinforcement-learning robots into homes at $399, replicating Raspberry Pi's category-creation playbook.
Mass-production track: Galbot S1 in CATL lines, Agibot G2 at 99.99% yield in 6 days, NEO Gen-2 targeting tens of thousands of units annually. Winners must nail manufacturing maturity, AI-ified controllers, and per-station economics.
Practical takeaways: hardware teams should study 1X's dexterous hand, Galbot's production lines, and Microduck's open BOM; software teams should benchmark VLA models on NEO simulators, Galbot deployment data, and the Microduck RL repo; investors should note these are three entirely different return paths — a track ticket, an engineering capability, and consumer demand.
Sources: Wallstreetcn FM-Radio 2026-08-28; Bloomberg 8-27 Microduck report; Xinhua 8-28 WRC dispatch; Borneo Post Online 8-28; Pandaily 8-27 Galbot funding report; Galbot official announcements.