ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models

ACE-Brain-0.6. Scaling Spatial Grounding for Physical Foundation Models. By Yinhe Wang and thirteen others. Published on the nineteenth of October. Its code repository is public but slight: five source files.

General-purpose multimodal models, GPT-6 Astra and Claude Opus 5.5 among them, show broad capability in embodied manipulation. The authors' analysis points to something underneath that generality: strong spatial perception. Astra can work out where to interact with the entities a task cares about, even when the scene is full of visual distractors. Spatial perception matters here because it is the link between high-level reasoning and the physical scene. A plan is only as good as the model's sense of where things actually are. The question the paper sets itself is whether an open model can be given that perceptual foundation by scaling visual grounding.

The bet is on data rather than architecture. If grounding is what connects reasoning to the world, then more and broader grounding supervision should strengthen it, provided the supervision can be checked. The work therefore centres on producing verifiable grounding supervision at scale, drawn from many kinds of source, and then fine-tuning an open model on it.

A unified data pipeline does the producing. It combines robot manipulation recordings with complementary visual data from diverse sources. The robot recordings are first annotated automatically, and the pipeline then converts everything it has collected into grounding supervision that can be verified. Fine-tuning Qwen3.5-9B on roughly one point five million samples from this pipeline yields ACE-Brain-0.6. The base model stays fixed. What changes is the breadth of what it is taught to locate.

Against GPT-6 Astra, ACE-Brain-0.6 comes out ahead on four of seven public grounding benchmarks. Its overall average score is seventy-five point four one.

Astra's is seventy-three point six eight. A second result speaks to the scaling claim directly. Under a fixed Qwen3.5-9B fine-tuning setup, aggregate grounding performance improves steadily as successive data releases expand the training collection to cover a broader range of grounding scenarios. The model and the recipe hold still, the data grows, and the score rises with it.

The argument runs in a short chain. Strong general models turn out to be strong spatial perceivers. Spatial perception is the bridge from reasoning to the physical scene. And that bridge can be built up in an open nine billion parameter model by scaling grounding supervision that is verifiable and diverse in origin, without changing the model itself.