Galaxy General's Yin Fangming Talks Embodied AI as Chinese Team Probes GPT-6 Astra Limits

Galaxy General chairman Yin Fangming spoke for China's embodied AI sector at the SCO Digital Economy Forum, with open collaboration as his core argument.

His proposed path has two layers: exporting mature technology foundations and product solutions, and working with local industry partners to co-develop robot solutions adapted to each country's manufacturing and service needs.

Set against OpenAI and peers pouring tens of billions into embodied AI, the remarks point to how Chinese embodied AI should address its data-scale gap.

Less than two weeks after GPT-6 Astra's release, the first team to systematically map its capability limits was Chinese. The test report was written by a PhD student of Professor Wang He, founder and CTO of Galaxy General.

The report paired OpenAI's GPT-6 Astra with π0.5 in a hybrid architecture, scoring 62.60 points and a 48% success rate across ten dual-arm tasks on the Robo Dojo benchmark — nearly twice the top score on the public leaderboard.

π0.5 is built on physical-space interaction and action priors, while GPT-6 leans toward semantic understanding. The division of labor has GPT-6 deciding what to do and whether it is right, and π0.5 deciding how to do it in physical space.

The comparison experiments were strictly aligned to closed-loop control. The report reveals two things: US large models have made a breakthrough in generalization for embodied manipulation, and the US is advancing faster than expected at the embodied AI brain layer.