A 9-billion-parameter model that rivals OpenAI's flagship on robot spatial tasks would let robot makers run perception on cheaper hardware, if ACE Robotics releases weights for outside testing.
ACE Robotics (大晓机器人), the embodied AI company chaired by SenseTime co-founder Wang Xiaogang, released ACE-Brain-0.6 on October 19, a model built to tell robots where objects are and where they can be grasped, placed or moved. The company says the model beat OpenAI’s GPT-6 Astra on six of eight spatial grounding benchmarks, averaging 74.53 across seven public tests against Astra’s 73.68, according to Jiemian and IT Home. On RobotGrounding, a test ACE Robotics built itself, it reported 94.13 against Astra’s 91.10.
The model is small by frontier standards. ACE-Brain-0.6 is a supervised fine-tune of Qwen3.5-9B, a 9-billion-parameter model, trained on roughly one million spatial grounding examples. Spatial grounding is the step where a model links a word such as "basket" or "shelf edge" to an exact location in a camera image, which a robot then turns into a movement.
Every figure comes from ACE Robotics. A Shanghai Observer report carried by the 21st Century Business Herald said the results came from company material and still await third-party evaluation. None of the reports said whether the model weights will be published, and the company’s Hugging Face page listed only earlier ACE-Brain versions when checked.
The company published five public benchmark scores alongside its own test
The Shanghai Observer report gave the most detail. On Where2Place, which asks a model to pick a free spot for an object, ACE-Brain-0.6 scored 76.55 against Astra’s 71.10. On RoboSpatial it scored 74.82 against 69.55, and on RoboAfford, which tests whether a model can identify the part of an object a robot should interact with, 88.51 against 84.94. Its margins were narrower on PixMo-Count, an object-counting test, at 89.41 against 88.10, and on ShareRobot, at 49.70 against 46.60.
RobotGrounding, the in-house test, contains 2,591 questions across 14 task types drawn from real robot scenes. The other two public benchmarks, where by that count Astra scored higher, were not named in any report EastFrontier reviewed. ACE Robotics also said the model led Google’s Gemini 4 Argon, Anthropic’s Claude Opus 5.5 and models from Alibaba’s DAMO Academy and the Beijing Academy of Artificial Intelligence, but published no scores for those comparisons, according to the same report and Science and Technology Daily.
The choice of rival is deliberate. ACE Robotics told Beike Finance that its analysis found GPT-6 Astra locates objects in images far more accurately than GPT-5.5, and that this accounts for much of Astra’s ability to handle physical tasks. The company described tasks such as setting down a fruit basket, where the model must find the basket’s rim and the surface beneath it, and folding a T-shirt, where it must trace the garment’s outline and work out folding distances. OpenAI’s own launch page for Astra reports no robotics benchmarks; its closest measure is ScreenSpot-Pro, a test of locating elements on a computer screen, where Astra scored 92.7%.
ACE Robotics credits data diversity for most of the gain
The architecture changes are modest. According to the Shanghai Observer, ACE Robotics froze Qwen’s image encoder and retrained only the language portion. The roughly one million samples cover about 940,000 images and 2.7 million question-and-answer pairs, assembled from more than 30 families of open robot datasets plus the company’s own first-person camera footage, with labeling tools GroundingDINO and SAM 3 used to mark object locations.
ACE Robotics says growing the training set from about 260,000 to about one million samples lifted the seven-benchmark average from 69.3 to 74.5. The company added that piling on more examples of familiar data types gained less than one point; most of the improvement came from new scenes and new kinds of spatial relationships.
Version 0.6 has no public release yet
ACE Robotics has published earlier models in the series openly. A technical report on arXiv dated July 5 describes ACE-Brain-0.5 as an 8-billion-parameter model evaluated on 15 benchmarks, and both that model and ACE-Brain-0 are downloadable. Until ACE-Brain-0.6 weights or a paper appear, outside researchers cannot rerun the scores, including the two losses the company has not named.
The company has the funding to keep iterating. An angel round led by Ant Group, with SenseTime Guoxiang Capital adding to an existing stake, was reported by EqualOcean in February, and IT Home put total funding at several hundred million US dollars by October. The release fits a wider move among China’s embodied-AI investors from robot bodies to robot brains. The next marker is a model card for ACE-Brain-0.6 on Hugging Face listing all eight benchmark results.