ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models
A 9B model beat GPT-6 Astra on four of seven spatial grounding benchmarks, the authors report.
ACE-Brain-0.6 averaged 75.41 across the benchmarks, compared with Astra’s 73.68. The team fine-tuned Qwen3.5-9B on roughly 1.5 million samples combining robot manipulation recordings with visual data from other sources. Its pipeline automatically annotates the recordings and converts the collected data into verifiable grounding supervision. With the fine-tuning setup held fixed, aggregate grounding performance improved as successive data releases broadened the training scenarios.
HF Daily Papers' note
ACE-Brain-0.6 averaged 75.41 across the benchmarks, compared with Astra’s 73.68. The team fine-tuned Qwen3.5-9B on roughly 1.5 million samples combining robot manipulation recordings with visual data from other sources. Its pipeline automatically annotates the recordings and converts the collected data into verifiable grounding supervision. With the fine-tuning setup held fixed, aggregate grounding performance improved as successive data releases broadened the training scenarios.
HF Daily Papers' note
score 5