Daxiao Robotics Releases Embodied Foundation Model ACEBrain 0.6, Outperforms GPT6 Astra in Six of Eight Evaluations

Rooted in Shanghai Xuhui, the sci-tech enterprise Da Xiao Robotics recently officially released its embodied foundational model ACEBrain 0.6. This 9B parameter model, focused on visual spatial anchoring capabilities, leverages millions of high-quality spatial training data, achieving results exceeding international top models such as GPT6 Astra in six out of eight authoritative evaluations, marking an important breakthrough in the specialized capabilities of domestic embodied models.

Technical Overview of ACE-Brain-0.6

Robots need to not only understand instructions and perform logical thinking in the real world, but also possess precise visual spatial anchoring capabilities. In simple terms, after receiving an instruction, a robot can accurately identify objects, locate the exact position for grasping or placing, and convert verbal commands into real mechanical actions. Even top-tier models such as GPT6 Astra still make errors when facing fine positioning and complex spatial relationship judgments. Minor deviations in chat scenarios are negligible, but in robotic operations, incorrect position judgments directly lead to task failures such as grasping and placing, which is one of the major challenges in practical implementation. Embodied Intelligence of deployment.

ACEBrain0.6 was specifically designed to address this weakness. It focuses on enhancing the core ability of spatial anchoring on the basis of the previous version ACEBrain0.5's cognitive closed-loop, relying on a large amount of verified spatial supervision data to improve target recognition, spatial positioning, and interaction point judgment capabilities, exploring a technical route of using large-scale high-quality data to drive the capability upgrade of robot models.

ACE-Brain-0.6 scores higher than GPT-6 Astra in six of eight Grounding evaluations

This model was trained using approximately one million high-quality, verifiable spatial anchoring training samples. Evaluation results show that in eight spatial anchoring specialized evaluations, ACEBrain0.6 scored higher than GPT6 Astra in six of them; the average score of seven international public benchmarks was 74.53, higher than GPT6 Astra's 73.68; in the self-built test set RobotGrounding for real robot scenarios, it achieved 94.13, also outperforming Astra's 91.10. Compared to mainstream models from institutions such as Google, Anthropic, Alibaba's DAMO Academy, and the Beijing Intel Institute, it also performs exceptionally well in physical space understanding and precise positioning tasks.

The model can identify objects, count items, determine placement positions, find grasping points, and even recognize targets that do not exist in the scene, avoiding robots from executing incorrect actions. Behind these results lies a set of automated, verifiable data production pipeline. Original robot operation videos only record actions, lacking object position annotations. Daxiao integrated multi-source robot data, processed it automatically through visual tools, generating massive training samples, with each sample needing image facts as a basis, filtering out erroneous data, allowing robots to evolve from merely recording actions to understanding the spatial logic behind them.

Automated Spatial Annotation of ACE-Brain-0.6's Robot Motion Trajectory

Experiments show that under the premise of an unchanged model architecture, training data expanded from 260,000 to one million samples, and the average evaluation score increased from 69.3 to 74.5. However, the performance improvement does not come from simply stacking similar data. Expanding new scenarios and diverse spatial relationships is the key to capability growth. This also proves that one does not have to blindly pursue ultra-large parameters; relying on high-quality, diverse training data, small parameter models can also match international top-level performance in specialized capabilities.

Spatial anchoring is the foundational capability for robot deployment, essential for retail sorting, logistics handling, and service robots. ACEBrain0.6 connects the chain from language understanding, spatial perception to physical execution, pushing robots from understanding the physical world to reliably completing real-world operations.

Daxiao Robotics is a tech company rooted in Shanghai Xuhui, with Wang Xiaogang, a co-founder of SenseTime, serving as chairman, and Academician Tao Dacheng as chief scientist. Daxiao Robotics has built a complete technical system for embodied intelligence across the entire chain, focusing on solving industry data scarcity and insufficient generalization capabilities, accelerating the commercialization of embodied intelligence technology. The company also deeply participates in the construction of the embodied intelligence industry in Xuhui District, conducting R&D based on the regional AI industry ecosystem, and is a representative innovative entity in Xuhui's layout of the robot and embodied intelligence track.

Here's a news blind box for you

Come open it and see

Reporter: Wang Yongjuan

Editor: Ning Pingying