Galaxea preps G0.5 MAX and a new robot as it chases 10,000-unit deliveries
At the World Robot Conference in Beijing, Galaxea AI (星海图) laid out the most concrete plan yet for turning embodied-AI models into an actual production system — and put numbers on it. Chief scientist Hang Zhao said the company's G0.5 foundation model beat industry SOTA across eight benchmarks including LIBERO and RoboTwin 2.0, and confirmed that a stronger G0.5 MAX version is about to ship. Just as notable: the startup is opening up, launching a "G0.5 reproduction program" that gives outside researchers full model weights, inference interfaces, evaluation suites, and fine-tuning tools — a bet that open ecosystem-building, not secrecy, wins the embodied race.
Two more pieces of the stack got upgrades. Galaxea previewed Fast-WAM, a world model that skips generating future video frames entirely and cuts single-step inference latency from roughly 800ms to 190ms — over 4x faster — with a pretrained version to follow; fast world models are what let robots react at physical-world speeds rather than demo speeds. And its new G-Fleet distributed RL system closes the loop for deployed fleets: policy rollout on real robots, data flowing back, retraining, and automatic version releases.
The hardware side got the spotlight too: Nexo, a next-gen wheeled dual-arm flagship with 20kg peak payload per arm, 8-hour battery with hot-swap, plus the legged Kengo for rough terrain (energy, inspection) and Lemo, a cheap desktop dual-arm rig for developers. COO Tianwei Li said Galaxea has already delivered thousand-unit-scale orders of model-driven robots and targets 10,000+ units next year. Roundtables added a useful dose of skepticism — DynaRobotics' Yang Shiyuan noted that pretraining on a million hours of human video actually lowered test accuracy, a reminder that watching humans and doing tasks are very different skills.
What to watch: whether G0.5 MAX's benchmark gains survive contact with real customer sites — Galaxea says mass commercial deployment is close, and 10,000 units is the number that will prove or break that claim.
If robots are finally clocking in as real workers, which job would you trust one with first?
Sources: 量子位
