Daxiao Robotics has officially launched ACE-Brain-0.6, an embodied foundation model enhanced for visual-spatial anchoring. The 9-billion-parameter system outperformed OpenAI's GPT-6 Astra in six of eight visual-spatial anchoring benchmarks, Gasgoo reports.
Image credit: Daxiao Robotics
Digging into the numbers, ACE-Brain-0.6 averaged 74.53 across seven international public benchmarks — topping GPT-6 Astra's 73.68. In the real-world RobotGrounding test, the model scored 94.13, well ahead of GPT-6 Astra at 91.10.
The model also holds an edge in broader head-to-head comparisons. It outperformed Google's Gemini 4 Argon, Anthropic's Claude Opus 5.5, and a range of other leading domestic and international models.
Built on Qwen3.5-9B, ACE-Brain-0.6 freezes the visual encoder while applying supervised fine-tuning to language model parameters. The architecture centers on visual-spatial anchoring, weaving object recognition, multi-instance localization, spatial relationship understanding, and interaction prediction into a single perception system. The goal: sharper fine-grained geometric localization and more accurate judgments of complex spatial relationships.
To build its training data, Daxiao Robotics drew from more than 30 open-source robotics datasets and proprietary sources, generating roughly 1 million training samples — including some 940,000 images and 2.7 million question-answer pairs. Controlled experiments reveal a clear scaling trend: holding architecture and training methods constant, expanding the dataset from 260,000 to 1 million samples lifted the average benchmark score from 69.3 to 74.5. Every doubling of data volume yielded an average gain of roughly 2.4 points.
Through multi-source data fusion and automated spatial annotation, Daxiao Robotics has established a unified pipeline for embodied spatial understanding. The system converts heterogeneous data into verifiable spatial-anchoring supervision signals. By scaling high-quality data, the company aims to bridge language understanding, spatial perception, and physical action — pushing embodied AI toward precise perception and reliable execution.
