ACE-Brain-0.6 beats GPT-6 Astra on six of eight grounding tests, but leads the seven-benchmark public average by only 0.85 points

ACE Robotics released ACE-Brain-0.6 on Oct 19, a 9B robot foundation model the company describes as built for visual spatial grounding. In ACE's reported tests it beat OpenAI's GPT-6 Astra on six of eight grounding evaluations, including margins of 5.45 points on Where2Place and 5.27 on RoboSpatial. The model follows ACE-Brain-0.5 and is based on Qwen3.5-9B.

The seven-benchmark public average is the tighter number: 74.53 for 0.6 against 73.68 for Astra, a lead of 0.85 points. ACE's scaling experiment puts the gain in the data pipeline. With architecture and training method held fixed, growing the corpus from 260,000 samples to about 1 million raised that average from 69.3 to 74.5, roughly log-linear at about 2.4 points per doubling. More samples of the existing data types added less than a point. The growth came with new scenes, new spatial relations and new forms of supervision. That makes the pipeline, not the 9B model, the part worth testing.

One million samples, 2.7 million question-answer pairs

The training set covers about 940,000 distinct images and 2.7 million question-answer pairs. ACE combined more than 30 open-source robot dataset families with private egocentric data, human first-person interaction data, dense object and part annotations and 3D spatial data. For manipulation video, the pipeline picks key frames from gripper state and motion amplitude, builds a scene object vocabulary automatically, and uses Grounding DINO and SAM 3 for localization, region recognition and semantic correction. It then generates referring, counting, spatial-relation, interaction-location and absent-target question-answer pairs. Annotation checks, answer balancing and exclusion of evaluation data are there to limit wrong supervision and shortcut learning. In training, the vision encoder stays frozen and only the language-model parameters are supervised fine-tuned, ACE says.

ACE says its analysis found that Astra's visual grounding improved markedly over GPT-5.5, while fine-grained geometric localization and complex spatial relations still trip it, such as placing "the free spot in front of an object" in an adjacent area. The eighth evaluation is the company's own: RobotGrounding, 2,591 questions across 14 task types including target localization, small-object recognition, multi-instance counting, absent-target judgment and cross-camera correspondence. ACE-Brain-0.6 leads Astra there by 3.03 points, narrower than its margins on Where2Place and RoboSpatial, both public benchmarks. ACE points to the absent-target cases, where the model should report that an object is absent rather than output a plausible but wrong location. It also compared 0.6 with Claude Opus 5.5, Gemini 4 Argon, models from Alibaba DAMO Academy and BAAI, and several open-source embodied AI models; ACE says 0.6 led those comparisons.

Every score in this article is ACE-reported. The two public benchmarks where Astra finished ahead are not named in the accounts from Leiphone, Xinmin Evening News or IT Home, and the ACERobotics Hugging Face page listed ACE-Brain-0 and ACE-Brain-0.5 on Oct 19, with no 0.6 model.