Chinese robotics company Daxiao Robotics has announced new benchmark results for its ACE-Brain-0.6 model, showing that a 9-billion-parameter system can outperform OpenAI's GPT-6 Astra on several visual spatial grounding tasks. The work builds on SenseTime's recent GPT-6-Astra embodied capability evaluation, in which Daxiao Robotics participated in real robot testing. The results underscore how precise spatial perception is becoming a decisive factor in translating general model reasoning into dependable physical actions.
A Shift from General Reasoning to Spatial Precision
Daxiao's analysis of GPT-6 Astra examined tasks such as placing a fruit basket and folding a T-shirt, where the model had to identify edges, estimate dimensions, and compute movement parameters. It found that stronger visual spatial anchoring, rather than advanced reasoning alone, is what allows the model to convert abstract instructions into concrete robot actions. Small spatial errors may matter little in text responses, but they can disrupt grasping, placement, and entire manipulation sequences in the physical world.
Six Benchmark Wins Over GPT-6 Astra
ACE-Brain-0.6 was evaluated on eight grounding benchmarks covering object reference, visual counting, spatial localization, functional understanding, and interaction region prediction. It outperformed GPT-6 Astra on six of them, including Where2Place, PixMo-Count, RoboSpatial, RoboAfford, ShareRobot, and RobotGrounding. Across seven public benchmarks, ACE-Brain-0.6 recorded an average score of 74.53, surpassing Astra's 73.68, and on the RobotGrounding set it scored 94.13 versus Astra's 91.10.
Competitive Standing Against Global Models
Daxiao says the model also led comparisons against Google Gemini 4 Argon, Anthropic Claude Opus 5.5, and representative systems from Alibaba DAMO Academy and the Beijing Academy of Artificial Intelligence. The performance is notable because ACE-Brain-0.6 is a 9 billion parameter model competing against much larger general purpose systems. These results suggest that focused spatial supervision can help smaller architectures close or reverse gaps in embodied perception tasks that require precise object and region localization.
Building a Verifiable Spatial Data Pipeline
To train the model, Daxiao Robotics integrated more than 30 open source robot data families with private egocentric data and converted heterogeneous robot interactions into spatial supervision signals. The pipeline selects key frames from manipulation videos based on gripper states and motion, then uses visual tools for object localization and semantic correction. It generated roughly one million training samples, involving about 940,000 images and 2.7 million question answer pairs, each with a checkable visual basis.
What Data Scaling Actually Delivers
Controlled experiments revealed that expanding training samples from 260,000 to one million lifted the seven public benchmark average from 69.3 to 74.5, with an approximately logarithmic linear trend. Daxiao reports an average gain of about 2.4 points per doubling of data volume. The research also found that repeating existing data types added less than one point, while adding new scenes, spatial relationships, and supervision formats contributed most of the improvement.
These findings position visual spatial anchoring as a practical foundation for robot manipulation in retail, logistics, and service settings. ACE-Brain-0.6 shows that a compact model, trained with high quality and verifiable spatial data, can reach or surpass top general systems on grounding tasks that directly affect execution reliability. As embodied AI moves from understanding commands to acting in physical spaces, such grounding capabilities are likely to shape the next wave of industrial deployment and model development.