The walls of a transparent glass on a champagne tower look shattered, hollow, and merged with the background to a traditional depth camera. To LingBot-Depth 2.0, the cup's boundaries, layers, opening direction — even the transparent stream of water pouring from a spout — are crystal clear.

On July 7, Ant Group's embodied AI subsidiary LingBot Tech officially released the spatial perception model LingBot-Depth 2.0, and simultaneously open-sourced the vision foundation model LingBot-Vision — the industry's first space-native vision foundation model built for real robotic tasks. Across 12 depth completion benchmarks, LingBot-Depth 2.0 took top results across the board. More critically, its backbone LingBot-Vision, with 1 billion parameters, matches or outperforms general-purpose vision foundation models roughly 7x its size (including 7B DINOv3) on dense spatial tasks like NYU-Depth v2.

Why Does Robotic Vision Need "Space-Native"?

For the past few years, embodied AI companies have relied on a "borrow-and-use" approach: taking general-purpose computer vision foundation models (like DINOv3, CLIP, SAM) and applying them directly to robots. These models excel at image classification, semantic segmentation, and object detection — but they fall apart in robotic scenarios.

Why? Because general vision models are trained to answer "what is this?" — recognizing cats, dogs, cars, pedestrians. But robots need vision that does more than recognize; they need to know "where is it, where are its boundaries, how far away is it, can I grasp it, will I crash if I walk there?" These are two completely different visual understanding paradigms.

The LingBot team found a critical breakthrough: boundaries. A boundary isn't just an object outline — it simultaneously signals depth discontinuity, occlusion relationships, spatial segmentation, and interactable zones. The rim of a cup determines whether a robotic arm can reach inside; the edges of a doorframe determine whether a robot can pass through; the contour of a tabletop determines whether objects placed on it will fall.

So the team proposed masked boundary modeling — a self-supervised training method specifically designed to make models learn boundaries and geometric structure. Standard masked modeling randomly masks patches of an image and has the model predict them, but random masking can hit low-information-density regions like walls or sky. LingBot-Vision takes a more aggressive approach: a teacher model identifies boundary tokens in the image online, forces these boundary tokens into the mask, and makes the student model learn to reconstruct the hardest, most spatially informative regions.

This is the boundary-forcing mechanism. The model doesn't just learn "what is this" — it learns "where are its boundaries, what's its shape, how does it relate to the surrounding space?"

Four Challenges from Lab to Real World

Robotic vision in real scenarios fears four types of scenes, and LingBot-Depth 2.0 delivers impressive results on all of them.

The first is transparent and reflective materials. This is robotic vision's worst nightmare — glass cups, champagne towers, mirrors, polished metal are visible in RGB, but depth sensors frequently return voids or incorrect distances in these areas. Through geometric priors learned by the LingBot-Vision backbone, LingBot-Depth 2.0 stably preserves geometric relationships among rims, walls, and bodies of cups, letting robotic arms know where to reach without shattering adjacent glassware.

The second is small and distant objects. As a mobile robot moves forward, it can't wait until obstacles are inches away to react — it needs to see distant people, pets, and barriers early. In outdoor scenes, LingBot-Depth 2.0 stably identifies distant tennis balls (objects smaller than a dog), with spatial layers among ground, walls, trees, and animals preserved intact.

The third is complex indoor scenes. Glass doors, curtains, TV stands, balcony clutter at home — combined with constantly changing lighting as the camera moves, traditional models often produce large broken patches at boundaries between glass, bright light, and shadow. The depth maps output by LingBot-Depth 2.0 are more continuous; walls, doorframes, and tabletop edges are more complete.

The fourth is edge deployment. LingBot-Vision's 0.3B distilled student model achieves accuracy comparable to 7B DINOv3 on NYU-Depth v2, with roughly 23x fewer parameters. That means no need to haul around a massive model — it can run on ordinary robot edge computing chips.

On the commercial side, LingBot-Depth 2.0 has been certified by Orbbec's Depth Vision Lab, and the two are collaborating on EGO-RGBD data capture devices, SDKs, and future integrated camera products. Model weights are open-sourced on HuggingFace and ModelScope; code is available on GitHub.

Over the past year, the embodied AI race has primarily focused on "bodies" (motors, joints, motion control) and "brains" (LLM integration, task planning). Ant LingBot's open-source release reminds the industry: visual perception is the first gateway through which robots interact with the physical world. Seeing without seeing accurately means all downstream planning and control is a house of cards. From Nvidia's ASPIRE skill library to LingBot's vision foundation model, embodied AI is moving from "impressive demos" to "actually getting work done."

See you tomorrow.