Galaxea

Chinese embodied AI company with the SOTA G0.5 VLA model — open-sourcing a unified autoregressive "think-while-act" architecture while scaling to $4.6B+ valuation.

1. Core Product / Service

Galaxea (星海图, Xinghaitu) builds embodied AI foundation models and the robots they run on. Its product ecosystem includes:

• G0 VLA model series: The flagship foundation model for robot control. G0.5 (released June 2026 at WDC 2026) uses a unified autoregressive architecture that generates reasoning tokens and action tokens from a single transformer — enabling "think-while-act" closed-loop behavior [1]. It decouples transferable basic action units (grasp, place, push, pull) to reduce data requirements, and ranks #1 on six global benchmarks [2].
• R1 robot platform: Wheeled dual-arm robots (R1 Lite, R1 Pro) as the primary deployment target for G0 models.
• A-series robotic arms: Single-arm manipulators (e.g., A1Z) for industrial tasks.
• Kengo: A bipedal humanoid robot unveiled at WDC 2026, representing the next embodiment frontier.
• Fast-WAM: A world-action model that predicts action outcomes before execution.

The company follows a three-layer strategy: (1) open-source VLA models to build an ecosystem, (2) sell robot hardware that runs the models, (3) invest in massive real-world data collection (100M-hour plan with Beijing Yizhuang) to feed model improvement [3].

2. Target Users & Pain Points

• Robot developers & integrators: Get a pre-trained, open-source VLA model that can be fine-tuned for specific manipulation tasks without starting from scratch.
• Industrial automation customers: Huawei Cloud, Volkswagen, Haier, Samsung, ByteDance — deploying Galaxea robots for factory and warehouse tasks [2].
• Research institutions: Stanford, MIT, and 150+ embodied AI development partners use Galaxea's open models and platforms for research [2].

The core pain point: current robot programming is brittle and task-specific. A general-purpose VLA model that works across embodiments could dramatically reduce deployment cost — and Galaxea's open-source strategy targets the academic/developer ecosystem that NVIDIA and OpenAI neglect.

3. Competitive Landscape

Galaxea's key differentiators: the unified autoregressive G0.5 architecture (ranked SOTA on multiple benchmarks), a fully open-source model + training framework, and the 100M-hour real-world data strategy — blending academic openness with industrial scale.

4. Unique Observations

• Open-source as a deliberate wedge. Unlike physical-intelligence (closed-source) and nvidia-groot|GR00T (open but CUDA-locked), Galaxea open-sources both weights and training framework under the OpenGalaxea GitHub org. This targets the global developer community in a way no US robotics company currently does — reminiscent of Meta's Llama strategy in LLMs.
• The 100M-hour data moat. Galaxea's joint venture with Beijing Yizhuang (亦数智能) aims to collect 100M hours of real-world robot data — 1M hours in 2026, scaling to 10M hours in 3 years [3]. If executed, this would be an order of magnitude larger than any existing robot dataset (GR00T's BONES-SEED is ~288 hours). Data, not model architecture, could be the real differentiator.
• Valuation velocity signals market frenzy. Galaxea went from ~$700M (mid-2025) to $4.6B+ (mid-2026) — 6.6x in ~12 months, with multiple rounds within quarters [4]. This mirrors the LLM fundraising wave of 2023–2024 and suggests embodied AI is now in its "capital-gravity" phase.

5. Financials / Funding

• Founded: September 2023, Beijing
• Total raised: >$500M (~¥3.5B) across ~10 rounds [4]
• Series B (Feb 2026): ¥1B ($144M) at ~$1.4B valuation, investors include Jinding Capital, BAIC, Hillhouse, Meituan, Capital Today [4]
• Series B+ (Apr 2026): ¥2B ($291M) at ~$2.9B valuation, ~20 institutions [4]
• Additional raise (mid-2026): ¥330M ($46M), valuation >$4.6B [4]
• Converted to joint-stock company (Jan 2026) — positioning for potential IPO

6. People & Relationships

• Founders: Gao Jiyang (高继扬), Xu Huazhe (许华哲), Zhao Hang, Li Tianwei — combining academic robotics research with industrial experience
• Key investors: Hillhouse (高瓴), Meituan, Ant Group, Capital Today (今日资本), IDG Capital, GL Ventures, Lens Technology, BAIC
• Partners: Ant Digital, Baidu AI Cloud, Haier, Beijing Yizhuang (joint data venture: 亦数智能)
• Key customers: Huawei Cloud, Volkswagen, Haier, Samsung, ByteDance, Physical Intelligence, Stanford, MIT
• Competitors: nvidia-groot, physical-intelligence, unitree, AGIBOT, Skild AI