--- title: "Spirit AI Raises 1.5 Billion Yuan A+ Round to Build the \"Brain\" for Humanoid Robots" date: 2026-11-05 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-robot-spirit-ai-moz1-funding language: en --- # Spirit AI Raises 1.5 Billion Yuan A+ Round to Build the "Brain" for Humanoid Robots > Hangzhou startup Spirit AI has raised roughly 4.5 billion yuan in six months, betting that a force-controlled humanoid called Moz1 plus its own vision-language-action models is the fastest route to useful embodied AI. Folding a T-shirt is trivial for a person and brutal for a robot: the fabric shifts shape with every touch, so the machine cannot rely on a fixed script. A humanoid in Hangzhou does it anyway — grabbing the shirt, sensing the wrinkles through force-controlled joints, flattening, folding and stacking without stopping, even when someone throws interference at it mid-task. The robot is Moz1, built by Spirit AI (千寻智能), and the company around it has just become one of the most capitalized embodied AI (具身智能) startups in China. In June 2026 it announced a 1.5 billion yuan (≈ US$210 million / HK$1.6 billion) A+ round; industry tallies put its first-half 2026 fundraising at about 4.5 billion yuan (≈ US$630 million / HK$4.9 billion), the largest in China's smart-device sector. ## Key takeaways - Spirit AI raised a 1.5 billion yuan (≈ US$210 million / HK$1.6 billion) A+ round in June 2026, taking first-half 2026 fundraising to roughly 4.5 billion yuan — top of China's smart-device sector. - Its full-stack bet pairs the Moz1 force-controlled humanoid with self-trained Spirit VLA models and self-collected real-world data. - Its RoboArena "global first" claim drew industry challenges — a reminder that in embodied AI, paid deployments outrank leaderboards. ## The bet: the model and the body must grow up together Founded in early 2024 by Han Fengtao, co-founder and former CTO of industrial-robot maker Rokae, Spirit AI took a deliberately different path from startups that buy arms from suppliers. It builds both layers of the stack: the Spirit family of vision-language-action (VLA) foundation models, and Moz1, a 26-degree-of-freedom humanoid with fully force-controlled joints and a 1:1 payload-to-weight ratio. Why does force control matter? Because most manipulation failures — crushed eggs, dropped shirts, dented sheet metal — happen when a robot applies position commands to a physical world that pushes back. Moz1's torque-controlled joints let the model feel resistance and adjust in real time, which is what makes fabric handling or delicate assembly plausible at all. The data strategy follows the same logic: rather than training mainly in simulation, the company collects real-world manipulation data with its own handheld capture devices, then trains Spirit models on that corpus. It says this lets the model and the hardware co-tune each other — the algorithm improves the body's skills while the body exposes what the algorithm still can't do. - Moz1: 26 DoF, fully force-controlled humanoid, released June 2025 - Spirit VLA models trained on self-collected real-world data - Spirit v1.5 open-sourced in 2026 - Long-horizon tasks: folding clothes, tidying desks, clearing tables - Stated goal: a robot in 10 percent of households within ten years ## Who is writing the checks — and what they signal The investor list reads like a map of Chinese tech capital: JD led the pre-A+ round in mid-2025; Prosperity7 Ventures, the venture arm of Saudi Aramco, led the earlier pre-A; Huawei's Hubble, state funds, and investors linked to CATL appear among backers. The June 2026 A+ round added leading dollar funds and state-affiliated capital, per the company's announcement. Two facts stand out. First, the money is large by any global standard for a two-year-old robotics startup, and it concentrated in one company — analysts noted humanoid-robot startups absorbed about 17 billion yuan across 25 deals in China in the first half of 2026, with Spirit AI at the top of the table. Second, industrial investors (JD, component-adjacent funds) suggest the near-term use case is factories and logistics, not home robots: industry research lists manufacturing and logistics as the company's first paid deployments. The company also says it has formed a strategic partnership with Bosch China. Reports describe Moz1 units being used for data collection and industrial task validation rather than volume production work — an honest snapshot of where the whole sector is. ## The benchmark flap every reader should know about In June 2026 Spirit AI announced that Spirit v1.6 topped the RoboArena leaderboard — a community-run evaluation platform backed by UC Berkeley, Stanford and NVIDIA researchers — ahead of NVIDIA's Cosmos and Physical Intelligence's Pi0.5. It would have been the first Chinese model at the top of that board, and the announcement landed the same day as the A+ round. The claim immediately drew scrutiny from industry researchers, and the episode became a case study in how benchmark hype can outrun verification in embodied AI. Ranking on physical-robot platforms is genuinely hard: sample sizes are small, hardware varies, and evaluation rules are still maturing. The durable takeaway is not who was number one on which day; it is that model-only claims — however impressive — still need to survive contact with paying customers. ## Honest limitations - Funding figures come from company announcements and industry tallies; Spirit AI is private, and round sizes are not independently audited. - The RoboArena "global first" claim was a company announcement that industry researchers challenged; this article treats it as contested, not established. - Deployment evidence — Bosch China partnership, paid pilots in manufacturing and logistics — is thin and largely company- or research-reported; no independent verification of volume commercial use exists. - Force-controlled humanoid hardware is promising in demos but unproven for cost-effective, all-day industrial operation; reliability data is not published. - The "10 percent of households in ten years" goal is a stated ambition, not a forecast with evidence behind it. ## What readers can do now - If you evaluate embodied-AI vendors, ask for the failure videos, not just the demo reel — how a robot recovers from a dropped object tells you more than the success clip. - If you follow robotics investment, treat benchmark rankings as marketing until the evaluation methodology and sample sizes are published. - If you work in warehouse or light manufacturing operations, force-controlled manipulation pilots are the near-term jobs to watch — fabric, food and mixed-bin handling are the first commercial beachheads. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-spirit-ai-moz1-funding Free to quote with attribution and a link to the original.