ACE Robotics, Founded by SenseTime Co-Founder, Releases Spatial Perception Model

ACE Robotics, a company founded by SenseTime co-founder Xiaogang Wang, released the ACE-Brain-0.6 embodied AI model on October 19. The model is designed to help robots connect language and visual information with physical space.

ACE-Brain-0.6 was built through supervised fine-tuning of Qwen3.5-9B. Its main functions are identifying objects, locating those objects in images, and determining where a robot can interact with them. Those capabilities are central to spatial grounding, which lets a robot translate a request such as 'pick up the cup' into a specific object and a reachable point in the physical world.

The company said it trained the model with about 1 million spatial-grounding samples. That data volume is meant to give the model broad exposure to object categories, spatial relationships and interaction contexts. ACE Robotics has not disclosed the full composition of the training set or the annotation process.

In the evaluations reported by the company, ACE-Brain-0.6 outperformed GPT-6 Astra on six of eight grounding benchmarks. The comparison positions the model against a general-purpose system on tasks that require linking language, images and robot action points. The results are company-reported and have not been independently verified.

ACE-Brain-0.6 scored an average of 74.53 across seven public benchmarks. On RobotGrounding, an evaluation based on real robot scenarios, it scored 94.13. GPT-6 Astra recorded 73.68 and 91.10 on the same two measures, respectively.

The benchmark gap on RobotGrounding is notable because real robot scenarios introduce variation in lighting, camera angle, object placement and clutter. A model that performs well in such settings must do more than recognize objects in clean images; it must also estimate usable interaction points.

ACE Robotics is one of a growing number of companies building embodied AI models that combine vision-language pretraining with robot-specific data. The approach often uses a general foundation model as a starting point, then fine-tunes it on task data. ACE-Brain-0.6 follows that pattern by adapting Qwen3.5-9B through supervised fine-tuning.

Spatial perception is a practical bottleneck for robots. A robot may recognize an object but still fail to grasp it if it cannot locate a handle, a rim or a clear approach path. Models that output interaction points aim to reduce that gap by turning perception into actionable coordinates.

The release also highlights the role of benchmark design in embodied AI. Public benchmarks allow comparisons across models, while real-robot evaluations test whether those gains survive physical conditions. ACE Robotics reports strong results on both, but independent testing will be needed to confirm how the model performs across different robot bodies and environments.

The company has framed ACE-Brain-0.6 around spatial grounding, rather than broad conversational ability. That focus reflects a common division of labor in robotics: language models interpret instructions, vision models identify objects, and grounding models connect those outputs to locations where a robot can act.

ACE Robotics was founded by Xiaogang Wang, a co-founder of SenseTime. The company's connection to SenseTime gives it a background in computer vision and AI research, though ACE Robotics is developing its own embodied AI products. The October 19 release is its latest public step in that effort.

TechNode reported the release, citing Pandaily. The available details focus on model architecture, training scale and benchmark scores. Questions about deployment, robot compatibility, inference cost and real-world reliability remain open.

For robotics developers, the appeal of a model like ACE-Brain-0.6 lies in reducing the engineering work needed to connect perception with action. If the reported performance holds, such models could make it easier to teach robots new objects and interaction targets without rebuilding the entire perception stack.

The results should be read with the usual caution around vendor-reported benchmarks. Benchmark scores can depend on prompt design, evaluation protocols and the match between training data and test scenarios. ACE Robotics has released headline numbers but not yet a full independent evaluation.

ACE-Brain-0.6 is therefore a model to watch rather than a settled solution. Its significance lies in the combination of a large language-model backbone, a million-scale spatial-grounding dataset and a focus on interaction points. Those are the ingredients that embodied AI companies increasingly see as necessary for robots that can act on instructions in real spaces.