INFIFORCE Raises ~1 Billion RMB Series A to Build AtomBrain, an Ego-Centric Embodied AI Data Moat

On August 14, INFIFORCE, a Chinese embodied-intelligence company, announced the close of its Series A and Series A+ rounds totaling nearly 1 billion RMB (~US$140M). The round was led by Dunhong Asset and a top-tier state-backed capital platform, with Zhejiang University Science and Innovation Group, Yandu State-Owned Holding, Lishui State Capital and others participating. Existing investor CCV (Creative Commons Venture) also increased its stake.

Proceeds will be directed to three areas:

1. AtomBrain, the company's causal world model 2. DataGrid, a full-stack AI training infrastructure 3. Mass delivery of multi-form robot bodies

Investor Composition: Four Capital Types at Once

The investor mix is itself a key signal:

• Dunhong Asset — hard-tech specialist in advanced manufacturing and embodied AI
• National strategic capital platform — long-horizon state capital reflecting industrial-policy intent
• Zhejiang University S&I Group — university-industry-research platform
• Yandu & Lishui SOEs — regional industrial capital from Jiangsu and Zhejiang
• CCV — existing investor continuing to back the team

Having financial capital + strategic state capital + regional industrial capital + university research all invest simultaneously is unusual in China's embodied-AI sector and signals that INFIFORCE is being positioned as a *region-level brain company* within China's local embodied-industry maps.

AtomBrain: An Engineering Bet on Ego-Centric Data

INFIFORCE is betting on a data path few have truly executed: Ego-Centric (first-person) data.

Most embodied models rely on one of two sources:

• Internet video (e.g., YouTube/Bilibili human-behavior clips — large volume, noisy labels)
• Teleoperated demonstrations (high quality, but expensive and limited in scale)

INFIFORCE chose a third path: let robots see and act with their own sensors, capturing genuine first-person interaction data. The analogy is human infant learning — not from videos, but from their own eyes, hands, force feedback, and trial-and-error.

The key advantage: the data reflects what the robot *actually perceives*, with no third-person annotation bias.

The Atom family of native embodied foundation models includes:

• AtomBrain — a causal world model that unifies visual observation, language instructions, action trajectories, and environmental feedback (force, position, velocity) in a single framework
• AIM world model (3rd gen) — reported 93.1% task success on the RoboTwin 2.0 benchmark
• Early Atom models — reported 97.0% task success on the LIBERO benchmark

Underpinning training is DataGrid, an in-house end-to-end intelligence-training system covering UMI handheld grippers, ego-collection rigs, and real-machine teleoperation, creating a closed loop from real-world experience to model capability.

Commercial Deployment: 30 Cities, 100 Scenarios

Models alone are not enough — embodied AI's real bottleneck is deployment in physical settings. INFIFORCE states it has built a product matrix around AtomBrain spanning dedicated, general-purpose, and humanoid robots, including:

• AstroDroid — wheeled humanoid
• UltraDroid — multi-form general-purpose platform
• XiaoYuanzi — bipedal
• FORCE series — dedicated task robots

Key figure: deployments in 30+ cities and 100+ real-world scenarios, covering commercial service, warehousing/logistics, and industrial manufacturing.

Even at 1–2 robots per scenario, this is 100–200 units of real deployment — large enough to spin an ego-data flywheel: more deployments → more data → better model → better deployment outcomes.

Team Background

• CEO Bai Huiyuan — former Alibaba Vice President; delivered the TEDxXuhui talk "The Force Awakens: The Rise of Embodied Intelligence"
• CTO Wang Yizhou — PhD from UC Berkeley, founding member of Baidu IDL's autonomous-driving team, core member of Nvidia DriveAV
• Chief Scientist Chen Jiayu — Assistant Professor at the University of Hong Kong, postdoctoral fellow at Carnegie Mellon
• R&D team is >75% master's/PhD level, with alumni from Alibaba, Huawei, Baidu, and Nvidia
• 100+ patents and software copyrights
• In July 2026, INFIFORCE-led national standard GB/Z "Specification for Crowdsourced Embodied-Intelligence Data Collection and Governance" was approved — China's first national data standard for embodied intelligence

Industry Signal: From "Who Has the Best Body" to "Who Has the Smartest Brain"

INFIFORCE's announcement states the crux: *"The core of embodied-intelligence competition has shifted from 'who can build the robot body' to 'who can make robots keep getting smarter.'"*

Over the past two years, funding has concentrated on robot-body companies — humanoids, arms, sensors. That logic is now being questioned: good hardware ≠ good intelligence. Even Tesla Optimus, with advanced hardware, still faces autonomy bottlenecks in factory tasks.

INFIFORCE's bet: brain + data. If body companies are "building the body," INFIFORCE is "building the brain and the real-machine data flywheel needed to train it."

Four Verification Points Over the Next 6–12 Months

1. AtomBrain's public third-party benchmark scores on RLBench, RoboTwin 2.0, and Calvin 2. How many of the 100 scenarios sustain stable operation for >6 months 3. Customer repurchase and renewal rates across the 30 cities 4. Any cross-embodiment transfer evidence — the same AtomBrain running on third-party robot bodies

---

Sources:

• INFIFORCE announcement
• Lansha Hard Tech 8-14 funding roundup
• Tencent Embodied Intelligence Daily 8-15
• Tonghuashun 8-15 report