ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models
Abstract Overview
This paper studies whether scaling visual grounding data can strengthen the spatial perception needed for embodied physical foundation models. The authors build a unified pipeline that converts robot manipulation recordings, egocentric interaction data, and other visual grounding sources into verifiable question-answer supervision, including automatically annotated robot scenes. Using this pipeline, they fine-tune Qwen3.5-9B on about 1.5 million samples to produce ACE-Brain-0.6. The evaluation focuses on public grounding benchmarks and a held-out robot-scene benchmark, with additional analysis of how performance changes as the amount and coverage of supervision increase.
Novelty
The distinctive contribution is a data-centric approach to embodied spatial grounding: a unified pipeline that automatically annotates robot recordings and merges heterogeneous visual sources into a common, verifiable grounding format. The paper also emphasizes controlled scaling analysis, holding the model and training recipe fixed to isolate the effect of supervision scale and coverage.
Results
ACE-Brain-0.6 achieves a 75.41 average across seven public grounding benchmarks, compared with 73.68 for GPT-6 Astra, and it outperforms Astra on four of the seven benchmarks. On the authors' RobotGrounding benchmark, ACE-Brain-0.6 reaches 93.42 versus Astra's 91.10. Across data releases, performance improves roughly log-linearly with sample count over the tested range, and the evidence suggests that broader supervision coverage contributes more than simply adding more samples from existing sources.
Key Points
- The training pipeline combines more than thirty robot dataset families with egocentric, object-centric, and spatial data, then converts them into verifiable grounding QA supervision.
- Fine-tuning Qwen3.5-9B on approximately 1.54 million samples yields strong gains over the base model and competitive performance against closed models on grounding tasks.
- The scaling study indicates that expanding the diversity and coverage of grounding supervision is more important than repetition from existing data sources for improving grounding performance.