ACE-Brain-0.6 fine-tune beats GPT-6 Astra on four grounding benchmarks
Original title: ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models
Fine-tuning Qwen3.5-9B on about 1.5 million samples scored 75.41 versus Astra's 73.68.
Researchers scale visual grounding for embodied operation by converting robot recordings and other visual data into verifiable supervision. Fine-tuning Qwen3.5-9B on roughly 1.5 million samples yields ACE-Brain-0.6. The model beats GPT-6 Astra on four of seven public grounding benchmarks. Under a fixed fine-tuning setup, aggregate scores rose as later data releases covered more grounding scenarios.
Unconfirmed: Benchmark comparisons are the authors' reported scores against the named models.
- Overall average was 75.41 for ACE-Brain-0.6 and 73.68 for GPT-6 Astra.
- The analysis also cites Claude Opus 5.5 among general multimodal models.
- Grounding scores improved as successive data releases expanded the training collection.
From the original article
1 min read at huggingface.co
General-purpose multimodal models, including GPT-6 Astra and Claude Opus 5. 5, demonstrate broad capabilities in embodied manipulation. Our analysis indicates that this generality is accompanied by strong spatial perception, as Astra can determine where to interact with task-relevant entities amid visual distractors.
By Yinhe Wang, Tao Tao, Zexin He · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.