Space‑Native Vision: LingBot‑Vision & Depth 2.0 Advance Embodied Intelligence

Background and Motivation

On June 8 the Ministry of Industry and Information Technology together with the State‑owned Assets Supervision and Administration Commission issued a joint notice requiring robots to shift from a "performance mode" that tests motion to an "operational mode" that tests intelligence by the end of the year. The Stanford AI Index 2026 report shows that while robots achieve ~90% success on the laboratory RLBench benchmark, the success rate drops to only 12.4% on the real‑world BEHAVIOR‑1K suite of 1,000 household activities, highlighting a gap in spatial perception.

Space‑Native Vision: LingBot‑Vision

Ant LingWave (蚂蚁灵波) released the next‑generation spatial perception model LingBot‑Depth 2.0 and open‑sourced the LingBot‑Vision foundation model, which they describe as the world’s first "space‑native" visual foundation model. The core idea is to embed spatial structure directly into the training objective rather than relying on random masking.

Boundary‑Centric Masked Modeling

Traditional vision pre‑training masks patches randomly, ignoring the fact that object boundaries carry the most information. LingBot‑Vision introduces Boundary‑centric Masked Modeling : during training the model predicts where image boundaries lie, forces those boundary patches into the masked set, and compels the model to reconstruct geometry from context. Boundary tokens are supervised as a classification of discrete distance‑and‑direction categories, generated online by a teacher model, so no manual labels are needed.

The authors illustrate a simplified example where a sparse set of corner points anchors the decoding of random boundary fields, producing coherent line segments even when the boundary field values are random. This "bootstrapping" trick lets the model first guess coarse geometry from corners and then refine details during training.

To avoid collapse when regressing continuous geometry, boundary prediction is cast as a classification problem and filtered with an a‑contrario statistical test, which discards insignificant pseudo‑boundaries and keeps training targets clean.

Training Efficiency and Benchmarks

LingBot‑Vision is a ViT‑g/16 model with ~1.1 B parameters, trained on 161 M images (selected from 2 B raw images), i.e., an order of magnitude fewer images than DINOv3, and with less than one‑third the training iterations. Despite this, on NYUv2 depth estimation LingBot‑Vision achieves RMSE 0.296 versus DINOv3’s 0.309, and on KITTI it is the strongest model under 2 B parameters.

On three segmentation benchmarks LingBot‑Vision matches DINOv3‑ViT‑H (0.8 B) and outperforms DINOv2 by over 4 percentage points. For video object segmentation it reaches top‑tier performance without any temporal fine‑tuning. Classification on ImageNet is a relative weakness (lower than DINOv3/DINOv2) due to the model allocating more attention to local geometry rather than pure semantics, which aligns with the "space‑native" design goal.

LingBot‑Depth 2.0 and Masked Depth Modeling

Building on the vision backbone, LingBot‑Depth 2.0 adopts the same training pipeline but replaces the encoder with LingBot‑Vision. It uses Masked Depth Modeling (MDM) : RGB and raw depth are tokenized separately, depth‑missing regions (e.g., reflective or transparent surfaces) are forced into the mask, and the model learns to infer depth purely from visual context. This is essentially a masked auto‑encoder for RGB‑D data.

Two changes drive the new version: (1) the encoder is initialized with LingBot‑Vision instead of DINOv2, and (2) the training data scale expands from 3 M to 150 M images. Results show LingBot‑Vision initialization consistently outperforms DINOv2/DINOv3 across benchmarks, especially on the hardest "block‑mask" scenario. On the DIODE‑Indoor dataset, ViT‑L RMSE drops from 0.152 (DINOv2) to 0.094, and ViT‑g from 0.118 to 0.083.

Data‑scaling experiments reveal that while both initializations improve with more data, DINOv2 plateaus after ~20 M images (D₁₀₂ ≈ 0.755), whereas LingBot‑Vision continues to climb (0.777 → 0.795), indicating a strong starting point amplifies benefits of larger datasets.

Real‑World Validation

LingBot‑Depth 2.0 has been certified by OBIR Zhongguang’s depth‑vision lab and integrated into their EGO‑RGBD data‑capture device, with an SDK released. Qualitative results show the model fills missing depth on glass, mirrors, and reflective floors, producing smooth, stable point clouds across frames.

Implications for Embodied Intelligence

The open‑source release of LingBot‑Vision provides the robotics community with a spatial‑native visual foundation, potentially lowering the barrier for downstream perception tasks (classification, detection, segmentation, depth) to share a single representation. This unified approach reduces engineering cost on compute‑constrained robot platforms. The authors argue that the shift from "can the robot perform a demo?" to "can the robot reliably perceive complex real‑world environments?" marks the transition from performance to operational mode, and that space‑native models are the essential groundwork for this evolution.