Daxiao Robot releases ACE-Brain-0.6, tops GPT-6 Astra on six spatial tests

The notable point is not the benchmark win but the method: a 9B model fine-tuned on roughly one million spatial anchoring samples edging out a frontier system on grounding, which suggests embodied spatial reasoning may reward targeted data curation more than raw scale. The margin is thin, and the strongest number comes from Daxiao's own RobotGrounding evaluation, so independent replication is the open question. Watch whether grounding-focused, smaller models become the default for robotics teams and whether GPT-6 Astra's camp responds with comparable spatial benchmarks.