**Robbyant’s LingBot‑Vision: Boundary‑Aware, Self‑Supervised Vision for Spatial Reasoning**
Robbyant, the embodied‑AI organization within Ant Group, has open‑sourced **LingBot‑Vision**, a family of self‑supervised Vision Transformers designed for dense spatial perception. Released under Apache‑2.0 on Hugging Face, the models (ViT‑giant, ViT‑large, ViT‑base, ViT‑small) include weights, a technical report, and inference code.
Unlike most foundation models that prioritize semantic “what” and discard fine‑grained spatial detail, LingBot‑Vision treats object boundaries, contours, and depth discontinuities as a **core pre‑training signal**. This boundary‑centric design enables a 1B‑parameter backbone to match or outperform models up to seven times larger on dense spatial tasks, including those where 7B DINOv3 sets the state of the art.
### How Masked Boundary Modeling Works
Building on DINO/iBOT self‑distillation, LingBot‑Vision uses a teacher–student framework where the teacher (an EMA‑averaged copy of the student) generates online targets. The key innovation is **boundary‑forcing**: boundary‑bearing tokens identified by the teacher are explicitly added to the student’s mask (M⁺ = M ∪ B). Boundary tokens are modeled via a categorical distance‑and‑orientation field, regressed as a per‑pixel classification problem with 32 bins. This setup yields two complementary learning objectives:
– Semantic reconstruction from random masking
– Geometric supervision at boundary locations
The combined loss integrates:
– DINO-style distillation
– iBOT boundary modeling
– Boundary-specific geometric loss
– KoLeo consistency loss
### Performance Highlights
When evaluated with frozen patch‑token features and simple linear heads, LingBot‑Vision achieves strong results across multiple benchmarks:
– **NYU‑Depth v2**: best RMSE (0.296), outperforming 7B DINOv3 (0.309) with roughly 7× fewer parameters
– **KITTI**: best performance below 2B parameters
– **ADE20K / Cityscapes / VOC**: competitive segmentation results, improving over same‑size DINOv2 by >4 mIoU and matching distilled DINOv3 variants
– **Video object segmentation**: 70.0 J&F on DAVIS‑2017 and 73.5 on YouTube‑VOS, on par with DINOv3‑H and 7B DINOv3
The model also distills effectively: a 0.3B ViT‑L student matches 7B DINOv3 on NYU‑Depth with about 23× fewer parameters, while ImageNet‑1K linear probing and k‑NN results show the trade‑off toward spatial fidelity versus global image‑level invariance.
### Use and Deployment
LingBot‑Vision’s frozen patch tokens support dense downstream tasks such as:
– Depth estimation
– Semantic and instance segmentation
– Video object segmentation
– Depth‑completion training initialization
The models are available in four sizes and can be loaded directly via the LingBot Vision library, with PyTorch ≥2.0 and Python ≥3.10 recommended for GPU deployment.
### Downstream Impact: LingBot‑Depth 2.0
By replacing DINOv2 with LingBot‑Vision and scaling training data from 3M to 150M samples, LingBot‑Depth 2.0 achieved leading results on 14 depth‑completion benchmarks. Gains compound: LingBot‑Vision–initialized training continues to improve with more data, while DINOv2 saturates earlier.
### Key Takeaways
– Boundary information is treated as a native pre‑training signal, learned without labels or edge detectors
– Boundary‑forcing with a categorical field enables joint emergence of geometry and semantics
– A parameter‑free Number‑of‑False‑Alarms (NFA) test validates detected boundaries
– High‑quality dense representations can be obtained efficiently, with strong performance even after distillation
– Spatial strengths persist even when models are scaled, making them well suited for embodied and robotic applications
**Reference:** Ant Group Research. “Vision Pretraining for Dense Spatial Perception.” *Robbyant*, https://example.com/lingbot-vision.
*This article is based on Ant Group’s technical release notes and aligns with the open‑source publication details for LingBot‑Vision.*
