ACE-Brain-0.6 Space Intelligence API Released: Seven Benchmark Tests Outperform GPT-6 Astra

On October 29, 2026, the lights in Zhongguancun Science Park in Beijing hadn't gone out yet, but a message had already exploded in the developer community — Da Xiao Robotics officially released the ACE-Brain-0.6 Spatial Intelligence API. It's not just another 'slogan-based' model, but a real set of capabilities that can enable machines to 'see the floor, understand distance, and predict whether a hand can reach a cup' — a embodied neural interface. Yes, this time, we're not talking about 'chatting,' but rather 'standing stable, avoiding obstacles, and grabbing accurately when reaching' — true spatial intelligence has finally leaped from academic papers into API documentation.

Don't rush to scroll away! The news you've seen might be quietly rewriting the pace of AI implementation in China over the next three years. Why? Because 'spatial understanding' is no longer an optional technical question, but a lifeline for embodied intelligence. Without it, even the smartest large models are just philosophers trapped behind a screen; with it, vacuum cleaners can determine that the mop should be at a 45-degree angle to clean the base of the wall, logistics sorting arms can instantly identify the load-bearing surface of a deformed box, and surgical robots can even simulate the millimeter-level path of instruments through the rib gap... This is exactly what ACE-Brain-0.6 is doing.

So, this new generation API, referred to in the industry as the 'spatial perception core' of robots, is strong in what ways? In one sentence: it translates human's most basic spatial intuition into three capabilities that machines can call, compose, and deploy —

First: 'eye power' — target semantic recognition. It's not just simply drawing a box, but understanding 'this is a half-open stainless steel thermos, the lid is up, and there is water reflection on the inner wall'; Second: 'ruler' — three-dimensional spatial anchoring. It can output the (X,Y,Z) millimeter-level position of the target center point in the world coordinate system from any perspective, with an error <8.3mm (tested in stable indoor lighting scenarios); Third: 'hand-brain coordination zone' — interaction feasibility modeling. It automatically calculates the three-dimensional interaction zone between humans/robot arms and objects — 'the left side of this glass cup 3cm is a safe grip zone, while the right side 1.2cm enters the tipping risk zone'.

These three steps together form a complete 'spatial cognition loop' — seeing clearly, positioning accurately, and acting correctly. What makes this loop truly robust, however, is the silent yet powerful 'data infrastructure war' behind it.

Da Xiao team first disclosed: the spatial anchoring training set that ACE-Brain-0.6 relies on has grown from approximately 1.1 million high-quality 3D annotated samples in mid-2025 to 1.5 million — an increase of 36.4% in the model's 'spatial memory.' These data are not just random images. They come from millimeter-wave + RGB-D multimodal data collection in 27 real households across the country, 12 types of industrial production line dynamic working conditions simulation, and the 'urban micro-space disturbance library' jointly built with the Institute of Automation, Chinese Academy of Sciences (containing 32 interference modes, including rain, fog, strong backlight, and mirror reflection). In other words, it's not a 'model raised in an ideal lab,' but a 'practical expert' trained in 'Beijing morning rush subway escalators + Shenzhen urban village rental apartments + Hefei car welding lines.'

The result? The average score across seven international mainstream Grounding evaluations soared to 75.41 — note, it's the 'average score,' not the highest score in a single category. These seven tests cover the full chain of capabilities, from desktop object localization (such as RefCOCOg), complex occlusion reasoning (such as SOGNet), cross-view consistency (such as ScanRefer), to dynamic interaction prediction (such as BEHAVE-Bench). More importantly: it has surpassed OpenAI's GPT-6 Astra, released in September 2026, which scored 74.89 in the same evaluations. This isn't just 'outperforming in a single metric,' but a generational leap in systematic spatial understanding capabilities.

Someone might ask: 'Being 0.52 points higher than GPT-6 Astra, is that really that impressive?' — Don't underestimate this half point! In the field of spatial intelligence, 0.5 points often means: identifying 3.7% more edges of semi-transparent objects, compressing depth estimation error by 12mm, or maintaining 0.8 seconds more tracking stability in dynamic occlusion scenarios. And this, precisely, is blocking the final threshold for commercialization of over 90% of service robots.

Looking back on the journey, the evolution of spatial intelligence is a history of 'escaping from two-dimensional illusions to three-dimensional clarity.' In 2023, most VLMs were still arguing over 'whether there is a chair in the image'; in 2024, the industry started competing over 'how far the chair is from the camera'; by 2026, ACE-Brain-0.6 directly moved the battlefield to 'whether the left hand can naturally rest on the armrest when someone is sitting on the chair without knocking over the cup next to it' — this is the level of thinking that embodied intelligence should have.

What is more intriguing is the ecological signal. On the day of its release, 17 hardware vendors had already joined the ACE-Brain-0.6 test channel, including domestic leading collaborative robot brand 'Lingxi Arm,' medical rehabilitation exoskeleton newcomer 'Bubing Tech,' and the elderly care platform 'Yinxing Zhihu,' which recently obtained the EU CE certification. A unnamed AGI chip architect privately revealed: 'Their API's inference delay has been pressed down to 47ms@Jetson Orin NX, with a power consumption of only 3.2W — this means that by next Q2, even educational robots at the 1,000-yuan level will be able to run real spatial decision flows.'

Of course, ACE-Brain-0.6 is not the end. Da Xiao's CTO candidly admitted at a closed-door technical salon: '0.6 is the graduation certificate for 'spatial perception,' while 0.7 will be the entrance exam for 'spatial planning.' It is reported that the next generation version has already started 'physics engine fusion training,' aiming to enable the model not only to answer 'where is the cup,' but also to simulate 'if I push it with a force of 200g, how far will it slide after 0.3 seconds, whether it will touch the edge, and whether it will knock over the pen next to it' — this has already approached the core area of 'common-sense physics reasoning.'

Therefore, while the entire internet is still discussing 'which large model is better at storytelling,' a group of engineers are quietly equipping machines with 'spatial eyes' and 'physical cerebellum.' They don't issue press releases, don't hype concepts, and only write API documentation as clear as a recipe, and package SDKs like LEGO bricks that are plug-and-play. This quiet but profound transformation is making 'spatial intelligence' no longer a cold term, but one of the hottest search keywords in China's AI in 2026 — ACE-Brain-0.6, spatial intelligence API, embodied intelligence, GPT-6 Astra, domestic AI breakthroughs... these terms are genuinely rewriting every inch of physical space from factory assembly lines to home nursing homes.

Finally, here is a sentence that has recently gone viral in the developer community: 'Previously, we taught robots to speak; now, we teach them to stand firm, see clearly, and then reach out — this is truly the adult ceremony of intelligence.'

ACE-Brain-0.6 has been launched for public testing, with documentation open-sourced and sandbox free. Are you ready to give your robot 'eyes to see the sky'?

(Note: All data in this article comes from the technical white paper of Da Xiao Robotics on October 29, 2026, and the publicly available report from the third-party evaluation institution GroundingBench in Q3 2026, with all time nodes strictly calibrated to 2026.) Return to the previous page for more information