The IT Home page is stamped 2026/10/19 14:37:32. It reports Daxiao Robotics’ own release, not an independent evaluation with a paper attached. The model is named ACE-Brain-0.6. The page calls it an embodied base model strengthened for visual spatial anchoring, a supervised fine-tune of Qwen3.5-9B, aimed at recognizing targets, locating them in space, and judging interaction regions. The training set is written as about one million spatial-anchoring samples. The page gives neither weights nor a paper.
[1]
The scores are the company’s disclosure. It says six of eight Grounding evaluations exceed OpenAI’s GPT-6 Astra. The average of seven public benchmarks is printed as 74.53, above Astra’s 73.68. A further test called RobotGrounding, described as a real-robot scene evaluation, is printed as 94.13 against Astra’s 91.10. The page does not name the eight tests, does not say how many samples each used, and does not say who produced Astra’s scores.
[1]The public averages are only the two printed figures, 74.53 and 73.68. The page does not turn that gap into a statistical result, and it prints no confidence interval. RobotGrounding is a score from the company’s own scenes. It cannot be added to the seven public benchmarks to make one overall rank. The word “ahead” on the headline therefore covers the six tests the company names, not all eight, and not a place on an outside leaderboard.
[1]The page also gives company background: Daxiao Robotics was founded by Wang Xiaogang, a co-founder of SenseTime, and as of October 2026 it had raised several hundred million dollars. That is background, not the terms of a round announced this day. The sum is not split into rounds, and the investors are not named.
[1]