六项评测超越GPT-6 Astra！大晓ACE-Brain-0.6以数据Scaling推动具身空间智能跃迁

Bottom line: Daxiao Robotics' ACE-Brain-0.6 model, trained on roughly one million verifiable spatial data points, allows a 9B-level model to outperform OpenAI GPT-6 Astra on six embodied spatial perception benchmarks.

1. Data is the key: Increasing training data from 260k to 1M samples raised the average score across seven public benchmarks from 69.3 to 74.5; each doubling of data added about 2.4 points on average. But adding more of the same type of data yielded less than 1 point of improvement, so continual introduction of new scenes, spatial relationships, and supervision formats is essential.

2. Practical value: Accurate target recognition, grasp-region localization, and placement-space understanding directly support retail picking, logistics operations, and service robots. The model can also determine when a target does not exist, preventing robots from acting recklessly.

3. Verifiable data: Every training answer has a clear visual basis, reducing erroneous supervision and helping robots move from merely understanding the world to precisely locating targets.

Bottom line: For brand owners, this article shows that competition in embodied intelligence products is shifting from parameter scale to the precision of spatial perception. Brands should treat visual spatial anchoring as the technical cornerstone of product experience differentiation.

1. Product development: ACE-Brain-0.6 proves that even a 9B-level compact model can surpass top-tier general-purpose models with high-quality data. This means brands don't need to blindly scale parameters; instead they should focus on user-perceptible capabilities such as object recognition, spatial localization, and interaction-position prediction.

2. User experience: In retail sorting, logistics operations, and service robots, a robot's ability to locate targets accurately, judge object functions, and understand available space directly affects service efficiency and consumer trust. In particular, nonexistent targets are rejected rather than falsely located, improving safety.

3. Competitive narrative: Using GPT-6 Astra as a benchmark and demonstrating superior results through public benchmarks and real-robot evaluations can serve as credible evidence for technology brand communication. Separately, the pattern of roughly 2.4-point average gains from doubling data size suggests that refining products requires continuously expanding new-scene data, not simply adding more of the same.

Bottom line: For sellers, the most direct takeaway is that embodied intelligence robots are entering retail sorting, logistics operations, and service scenarios, potentially reshaping the cost structure of warehouse fulfillment and after-sales service.

1. Opportunity: ACE-Brain-0.6 can more accurately identify target objects, grasp points, and placement space, reducing mispicks and omissions. The model can also determine when a target is absent, lowering the risk of damage from needless robot movement.

2. Data asset value: The large volumes of first-person interaction videos and order-operation videos generated in daily seller operations can be converted into training data through key-frame filtering and spatial annotation. In the future, sellers may collaborate with robot manufacturers or data platforms and create new data-driven revenue streams.

3. Risk note: Model capability does not come from scale alone; adding more of the same type of data improves performance by less than 1 point. If partners cannot continuously provide data from new scenes and new spatial relationships, performance growth will stall. Therefore, when choosing a technology partner, sellers should check whether its data pipeline covers diverse, real-world scenarios.

Bottom line: For factories, the core insight is that robots in manufacturing must not only understand objects but also pinpoint target positions and operable areas. Spatial anchoring is the key to making automation work on the ground.

1. Product and design requirements: Tasks such as folding clothes, placing goods, and grasping parts require robots to extract contours, estimate sizes, and compute orientation and position. When deploying robots, factories should prioritize fine-grained localization capability over large-model parameter counts.

2. Digitalization and data assets: Robot operation videos, gripper states, and motion trajectories in factories can become high-quality training data. Using the article's approach, key frames can be filtered, an object vocabulary built, and automated annotation tools applied to generate spatial question-answer data, forming a verifiable data pipeline.

3. Business opportunity: A 9B-level model can reach high capability with around one million training samples, which means small and medium-sized factories do not need to pursue extreme compute. Jointly building real-scene datasets with model providers, or creating private data in-house, can improve sorting, loading/unloading, and quality-inspection efficiency of production-line robots.

Bottom line: For solution providers, this article offers a technical fix for a class of client pain points and provides verifiable delivery evidence: robots often make mistakes in fine-grained geometric localization and complex spatial-relationship judgments, and ACE-Brain-0.6 improves this with visual spatial anchoring.

1. Client pain points: Even top-tier general-purpose models can misjudge the empty space in front of a specific object and point to an adjacent area. This has little effect on question answering, but it deeply impacts operational chains such as robotic grasping and placement. Service providers can use this to diagnose why client systems fail.

2. Solution: Deliver an embodied model built on spatial anchoring, unifying object recognition, multi-instance localization, spatial relationships, functional-region recognition, and interaction-position prediction. When a target does not exist, the model explicitly rejects it, reducing operational errors.

3. Delivery results and data methods: Across eight evaluations, six exceeded GPT-6 Astra, with an average public score of 74.53 across seven benchmarks and 94.13 on real-robot scenarios. Expanding training data from 260k to 1M samples raised the average score by 5.2 points. Service providers can reference its automated data pipeline—key-frame filtering, vision-tool annotation, and answer validation—to reduce data-building costs.

Bottom line: For marketplace/platform operators, this article highlights a platform-scale opportunity: build a large-scale, automated, and verifiable data pipeline around spatial anchoring data, and use it to attract embodied-model developers and robotics companies to the ecosystem.

1. Platform approach: Integrate more than 30 open-source robot data families with proprietary ego data, providing unified data ingestion, key-frame filtering, scene-specific object-vocabulary construction, and automated annotation tools such as Grounding DINO and SAM 3, yielding a spatial-supervision data production platform.

2. Operations and governance: Built-in mechanisms for annotation validation, answer balance, and evaluation-data exclusion are needed to prevent erroneous supervision and training shortcuts. The RobotGrounding benchmark can be opened up with rankings covering target localization, small-object recognition, multi-instance counting, target-absence detection, and cross-camera viewpoint correspondence, attracting ecosystem participants.

3. Risk avoidance: More data does not necessarily lead to linear capability gains; simply expanding existing types of samples has limited value. The platform should encourage new scenes and spatial relationships, continuously broaden data coverage, and prevent developers from gaming with low-information incremental data.

Bottom line: This article frames an important thesis for embodied AI research: spatial anchoring is a key link between high-level reasoning and physical execution. Data scaling can overcome parameter-size limits, and a 9B model can surpass GPT-6 Astra on six grounding evaluations.

1. New direction: Capability gains come not only from model architecture but also from verifiable spatial supervision data. ACE-Brain-0.6 achieves a leap in spatial perception with about one million samples, 940k images, and 2.7M question-answer pairs, revealing a technical path from language understanding to physical action.

2. New question: Data scaling shows approximately log-linear returns, with each doubling of data giving an average gain of about 2.4 points, while adding only samples of existing types gives less than 1 point. This raises a research question: how to quantify the contributions of data coverage, information density, and diversity of spatial relationships in order to avoid homogeneous supervision signals.

3. Suggestions and implications: The field should build unified data-construction pipelines and benchmarks, such as RobotGrounding with 2,591 questions across 14 task types, covering target localization, counting, spatial relations, affordances, and target-absence detection. It should also encourage open-sourcing multi-source robot data, automated annotation, and answer-validation mechanisms, pushing embodied AI from precise perception toward reliable execution.