Ant Lingbo Open-Sources First Spatial-Native Embodied Vision Foundation Model

Ant Lingbo, Ant Group's embodied AI and robotics team, announced on July 7 that it has open-sourced what it calls the first spatial-native embodied vision foundation model. The model is purpose-built for robot vision, with a core focus on enabling machines to perceive and understand 3D spatial structures rather than merely processing 2D image data.

The spatial-native design philosophy means 3D spatial perception is baked into the model's architecture from the start, rather than being retrofitted onto existing vision models. According to Ant Lingbo, traditional vision models often lack native depth understanding for tasks like robot grasping, navigation, and object manipulation, a gap this open-source model aims to fill at a fundamental level.

Embodied intelligence is one of the frontier directions in AI research, emphasizing that AI systems must be capable of interacting with the physical world. Vision perception is a critical foundational module — without an accurate understanding of their 3D environment, robots struggle with autonomous operation and navigation. By open-sourcing this model, Ant Lingbo allows researchers and developers to freely access, use, and modify the foundation model.

From an industry perspective, the lack of spatial-native vision models has long constrained embodied AI development. Most vision models are trained for 2D image recognition and have limited capacity for processing distance, occlusion, and object volume information in 3D space. Ant Lingbo's open-source effort aims to provide a more targeted starting point for research and engineering teams in robotics.

Open-source is a consistent strategy for Ant Group in the AI space. The company has previously contributed open-source projects in financial AI, privacy computing, and other domains. This move by Ant Lingbo extends its open-source footprint into embodied vision, signaling continued investment in AI infrastructure.

For those tracking embodied intelligence and robotics, the next watchpoint is how the model performs on real-world robotic platforms and whether the community can rapidly build applications on top of it. Community contributions and feedback will directly influence the model's iteration speed and ecosystem growth.