Galaxea (星海图), a Beijing-based embodied AI unicorn, held its inaugural Global Developer Conference (Galaxea WDC 2026) in Beijing Yizhuang on June 16, unveiling a comprehensive technology and ecosystem roadmap that positions the company as China's most vertically integrated embodied AI player.
G0.5 VLA Foundation Model: Open-Source and Top-Ranked
The centerpiece of the event was the release and open-sourcing of G0.5, Galaxea's next-generation Vision-Language-Action (VLA) foundation model. Built on a unified autoregressive VLA architecture, G0.5 integrates visual understanding, language reasoning, and action generation into a single pipeline, distilling transferable foundational action primitives.
According to Galaxea CEO Gao Jiyang, G0.5 has achieved top-1 ranking in China across seven major global benchmarks, placing it firmly in the global first tier. The model可以实现 zero-shot generalization — handling unfamiliar objects, novel scene layouts, and new language instruction combinations without task-specific fine-tuning.
The company also revealed its roadmap: G0.7 targeting long-horizon bimanual manipulation, and G1.0 aiming for general-purpose bimanual operation intelligence.
Fast-WAM World Model: 190ms Inference Latency
Alongside G0.5, Galaxea introduced Fast-WAM, a world model that compresses single-step inference latency to just 190 milliseconds — over 4x faster than traditional architectures. Rather than first imagining future video frames then executing actions, Fast-WAM directly generates action sequences during inference, enabling real-time operational intelligence.
Kengo: The Bipedal Humanoid Debut
The company also showcased Kengo (行客), its first fully self-developed bipedal humanoid robot. With 80% of its powertrain components (modules, gears, motors) developed in-house or through co-development with supply chain partners, Kengo demonstrated both high-difficulty whole-body movements (e.g., quadruple consecutive kicks) and practical bimanual tasks such as handing objects, carrying boxes, and folding clothes.
With Kengo's debut, Galaxea becomes the only Chinese company that simultaneously possesses both top-tier foundation models and self-developed robot hardware — closing the "full-stack + intelligence" strategy it set three years ago.
1 Million Hours: The Data Ecosystem
Recognizing real-world data as the critical bottleneck for embodied intelligence, Galaxea announced the formation of Yishu Intelligence (亦数智能), a data company jointly established with Beijing Yizhuang Robot Company and Yizhuang State Investment. The entity launched the "1 Million Hours Ultra-High-Quality Real-World Data Initiative," targeting one million hours of data collection this year and scaling to ten million hours over three years.
A Data Ecosystem Alliance was simultaneously formed with 15 founding members including Ant Digital Technologies, Baidu AI Cloud, and Haitiansheng, covering the full data pipeline from collection and annotation to application.
Xingtu Plan: Startup Incubation
Galaxea also launched the "Xingtu Plan" (星途计划) in partnership with Cathay Capital, targeting incubation of 100 embodied AI early-stage startups over five years with 1 billion RMB in committed investment.
Gao Jiyang made a bold prediction: "Leveraging China's dual advantages in data supply chain and hardware supply chain, China's embodied foundation model capabilities will surpass the US to become the world's best within two to three years."
The event featured an international ecosystem roundtable with robotics leaders from Germany, Japan, and the US, underscoring Galaxea's global ambitions.
