Boundaries are the least redundant, most informative regions of an image.
At a moment when artificial vision has grown vast and costly, Ant Group's robotics division has released a model that achieves more by attending to less — not the broad semantic sweep of an image, but the precise edges where surfaces meet and depth changes. LingBot-Vision, a 1.1-billion-parameter encoder trained on boundaries as primary signals rather than afterthoughts, outperforms models seven times its size on the spatial tasks that matter most to robots and autonomous systems. It is a quiet argument that in perception, as in wisdom, knowing where things end may matter more than knowing what they are.
- Vision foundation models have grown so large and semantically focused that they routinely discard the fine boundary and depth information that robots and autonomous vehicles actually depend on.
- LingBot-Vision inverts this priority entirely, treating object edges as native training signals rather than downstream problems — forcing geometry and semantics to develop together rather than compete.
- A novel masked boundary modeling technique targets the least redundant regions of an image, using statistical validation to ensure only meaningful structure becomes a teaching signal, preventing training collapse.
- On standard benchmarks, the 1.1B-parameter model beats 7B competitors on depth estimation and matches them on video segmentation, while training on a dataset an order of magnitude smaller.
- Distilled variants as small as 86 million parameters preserve most of the performance, opening practical deployment paths for robotics hardware that cannot afford the compute costs of today's dominant models.
Ant Group's robotics division has released LingBot-Vision, a 1.1-billion-parameter vision model that outperforms systems seven times its size — not by doing more, but by paying attention to what others discard. Most large vision models optimize for semantic understanding, learning to identify objects while quietly throwing away the boundary and depth information that a robot's gripper or a self-driving system actually needs. LingBot-Vision treats those boundaries as the primary signal from the start.
The training method, called masked boundary modeling, extends standard self-supervised techniques with a key insight: not all image patches are equally hard to reconstruct. A patch in the middle of a flat wall is trivially recoverable from context; a patch straddling an object boundary is not. By explicitly masking boundary-bearing regions and assigning them geometric targets alongside semantic ones, the model is forced to develop both types of understanding simultaneously rather than sacrificing one for the other.
The implementation represents boundaries as dense fields encoding each pixel's distance and angle to the nearest edge, then recasts prediction as classification across discrete bins — a move that stabilizes training and enables a statistical validation step that filters out unsupported structure before it can corrupt the learning signal.
The results are concrete. On NYU-Depth v2, LingBot-Vision achieves an RMSE of 0.296, beating the 7-billion-parameter DINOv3's 0.309, while training on roughly one-tenth the data. On video object segmentation, it matches much larger models using only frozen features and cosine similarity, with no temporal supervision. The one acknowledged trade-off is image-level recognition, where the model trails competitors that allocate capacity to semantic invariance rather than spatial structure.
Downstream, the gains compound. Swapping LingBot-Vision into a depth-completion system and scaling training data cut indoor RMSE in half and achieved near-perfect results on transparent objects — a domain where conventional depth sensors routinely fail. Distilled variants as small as 86 million parameters preserve most of the performance, making the approach viable for the constrained hardware that real robotics deployments actually run on. The model is available now under Apache 2.0, and the field is left to consider whether boundary-centric pretraining has quietly become the right foundation for any system that needs to understand physical space.
Ant Group's robotics division has released an open-source vision model that does something counterintuitive: it gets better at understanding space by ignoring what most AI systems prioritize. LingBot-Vision, a 1.1-billion-parameter encoder, outperforms models seven times its size on depth estimation and object segmentation—the kinds of tasks that matter to robots and autonomous systems navigating the physical world.
The model arrives at a moment when vision foundation models have grown enormous and expensive to train. Most of them optimize for semantic understanding: they learn to identify what's in an image while actively discarding the fine-grained spatial details—object boundaries, depth discontinuities, the precise edges where one surface meets another—that a robot's gripper or a self-driving car's navigation system actually needs. LingBot-Vision inverts that priority. It treats boundaries not as a downstream problem to solve after the fact, but as a native signal to learn from the start.
The training method, called masked boundary modeling, builds on existing self-supervised techniques but adds a crucial twist. Standard masked image modeling hides random patches and asks the model to recover them from context. The insight here is that some patches are harder than others: a patch in the middle of a flat wall is easy to reconstruct from its neighbors, but a patch straddling an object boundary contains information that context alone cannot supply. Boundaries are the least redundant regions of an image. LingBot-Vision forces the model to pay attention to them by explicitly masking boundary-bearing tokens and giving them geometric targets in addition to semantic ones. This dual routing prevents the two types of information from competing; instead, they co-emerge.
The technical implementation uses a categorical boundary field—representing edges as line segments lifted into a dense map where every pixel stores its distance to the nearest boundary and three angles locating it. Rather than regressing this directly, which causes training to collapse, the model discretizes each channel into 32 bins and recasts boundary prediction as classification. This move has an elegant side effect: it enables a parameter-free validation test based on statistical null hypotheses, so unsupported structure never becomes a teaching signal. The teacher network decodes candidate segments, validates only the statistically significant ones, and re-renders them into the target field.
On the NYU-Depth v2 benchmark, a standard test for depth estimation, LingBot-Vision achieves an RMSE of 0.296—better than the 7-billion-parameter DINOv3 at 0.309. The training was also efficient: the model learned from 161 million curated images selected from a 2-billion-image web pool, an order of magnitude smaller than DINOv3's dataset, and consumed less than a third of DINOv3's training samples. The model ships in four sizes—ViT-giant, ViT-large, ViT-base, and ViT-small—under Apache 2.0 on Hugging Face.
The performance holds up across other spatial tasks. On video object segmentation, LingBot-Vision reaches 70.0 J&F on DAVIS-2017 and 73.5 on YouTube-VOS, matching the much larger DINOv3 models. The boundary tokens themselves are stable enough to track through video using only frozen features and cosine similarity, with no temporal supervision. The trade-off is image-level recognition: on ImageNet-1K, the model reaches 86.32 with linear probing, trailing DINOv3-7B, which allocates its capacity to image-level invariance rather than spatial structure.
Where the architecture really proves its worth is in downstream applications. The team upgraded its depth-completion system to LingBot-Depth 2.0 by swapping in LingBot-Vision as the encoder and scaling the training data from 3 million to 150 million samples. The results compound: on block-masked DIODE-Indoor, RMSE dropped from 0.132 to 0.062—cut in half. On transparent objects, where active depth sensors typically fail, the system achieves 0.010 RMSE. Crucially, as training data grows, the LingBot-Vision-initialized curve keeps improving while the older DINOv2-initialized baseline saturates. The better starting point amplifies rather than washes out with scale.
The distilled variants matter for real deployment. A 300-million-parameter ViT-L student matches the 7-billion-parameter DINOv3 on NYU-Depth v2 with roughly 23 times fewer parameters. An 86-million-parameter ViT-B variant leads dense prediction within its size class. For robotics systems and embodied AI applications running on constrained hardware, this efficiency opens doors that were previously closed. The model is available now, and the question for the field is whether boundary-centric pretraining becomes the standard for any system that needs to understand space.
Citações Notáveis
Most vision foundation models are trained for semantic invariance: they learn to answer what is in an image while discarding exactly the fine-grained spatial structure that robots and other physically embodied systems depend on.— LingBot-Vision technical report
A Conversa do Hearth Outra perspectiva sobre a história
Why does a robot care more about boundaries than about what something is?
Because a robot lives in space. It needs to know where one object ends and another begins—not to name them, but to move through the world without collision. A boundary is geometry. A label is semantics. They're different problems.
So most vision models have been solving the wrong problem?
Not wrong. Just misaligned with what embodied systems need. A model trained to recognize "chair" learns to ignore the exact contour of the chair. That's useful for image classification. It's useless for a gripper trying to grasp the chair.
How does masking boundaries force the model to learn them?
By making them harder to ignore. If you mask a patch in the middle of a wall, the model can guess it from context. If you mask a patch where two surfaces meet, context isn't enough. The model has to learn the structure itself. That's where the signal lives.
The categorical boundary field—why discretize instead of regress?
Regression collapses during training. But if you turn it into classification—32 bins per channel—you inherit the stability machinery that modern self-distillation uses. And you get something free: a statistical test that validates whether a boundary is real or noise.
What does it mean that advantages survive distillation?
It means the spatial understanding isn't fragile. You can compress the model 23 times and keep the depth performance. That's the difference between a trick that works at scale and a principle that works at any scale.
Why does more training data help LingBot-Vision but not DINOv2?
Because DINOv2 wasn't built for spatial tasks. It saturates on semantic understanding. LingBot-Vision was built to extract geometry from images. More images mean more geometry to learn. The architecture and the data align.