Daxiao Robotics Unveils ACE-Brain-0.6 Embodied Model, Claims Edge Over GPT-6 Astra

Daxiao Robotics has introduced ACE-Brain-0.6, a vision-spatial-anchoring-enhanced embodied base model. Built on Qwen3.5-9B through supervised fine-tuning, it is designed to improve object recognition, spatial localization and interaction-region judgment—core capabilities for grounding language and perception in physical environments.

According to Daxiao, the model was trained with roughly one million spatial-anchoring samples. These samples are intended to help ACE-Brain-0.6 link visual inputs to physical locations and actionable regions, moving beyond general image-text alignment. The company disclosed the details through Jiemian.

In grounding evaluations, ACE-Brain-0.6 beat OpenAI's GPT-6 Astra on six of eight tests. Grounding tasks test whether a system can locate objects or regions referred to in language, a prerequisite for manipulation and navigation. On seven public benchmarks, it averaged 74.53 points versus 73.68 for GPT-6 Astra. On Daxiao's self-built RobotGrounding benchmark, it scored 94.13. The company says the results show stronger spatial perception.

The release highlights spatial anchoring as a route to more capable embodied AI. Because the comparisons include a self-built benchmark and are company-reported, independent testing would be needed to determine how well the gains transfer to real robot tasks. As a base model, it would still require task-specific adaptation for deployment.