Galaxea packages a full embodied-AI stack: open weights, a sub-200ms real-world simulator, and three humanoids

At the 2026 World Robot Conference, Xinghai Tu (星海图), a Beijing-based embodied-AI lab, packaged an unusually full stack into one announcement: an open foundation model, a sub-200ms world model, a fleet-training system, and three new humanoids.

G0.5 is a vision-language-action model: a single autoregressive stream that links perception, language, reasoning, and action. The company says it surpassed state-of-the-art on eight benchmarks including LIBERO and RoboTwin 2.0. Galaxea opened the weights, inference interface, and fine-tuning toolkit through a reproduction plan, with code on GitHub and weights on Hugging Face; the paper covers the architecture. Fast-WAM, the world-model piece, skips future-video generation at inference and reportedly cut single-step latency from ~800ms to 190ms, a more-than-4x speedup the company is now scaling into a larger pretrained model.

The training loop is G-Fleet, a distributed reinforcement-learning system that links cluster-level policy dispatch with real-machine rollout and real-scene data return. The hardware is three new platforms: Nexo, a 20kg-payload wheel-arm humanoid for 24/7 retail, logistics, and industrial work; Kengo, a bipedal unit for unstructured environments like inspection and exploration; and Lemo, a desktop dual-arm developer kit.

COO Li Tianwei said the company has shipped at thousand-unit scale and is targeting 10,000+ units in 2027, both company-stated figures, not independently measured. JD's Tian Ruilin attached a 2028 target of one million robots across a general-base + vertical-model architecture, also a corporate forecast. Dematic's Wang Kai argued overseas ROI cycles are shorter and pushed a "go overseas first" strategy. The panel itself flagged ROI computability and shop-floor reliability as the real test.