LingBot-Depth 2.0

What is LingBot-Depth 2.0?

LingBot-Depth 2.0 is a next-generation foundational model for robotic spatial perception released on July 7, 2026, by Robbyant, a subsidiary of Ant Group. It is not positioned as a large language model, but rather as a foundation model for 3D vision, primarily designed to enable robots to ”see the three-dimensional world clearly.”

GPT is the brain of AI, while LingBot-Depth is the robot's eyes.It addresses one of the most fundamental issues in robotics—Depth Perception.

Technical Principles of LingBot-Depth 2.0

The breakthroughs achieved by LingBot-Depth 2.0 are primarily attributable to the following two core technologies:

• Spatial Native Vision Platform (LingBot-Vision): It is the industry’s first visual foundation model to use “boundary structure” as a pre-training objective. By introducing geometric constraints and a boundary-centered masking mechanism, it possesses sub-pixel-level boundary localization and spatial structure understanding capabilities, providing robust foundational support for depth estimation.
• Masked Deep Modeling (MDM) and Big Data: The model uses the “depth gaps” naturally created by depth sensors on transparent and reflective objects as masks, requiring the model to “fill in” the complete depth using only RGB color images and the remaining valid depth information. At the same time, the scale of its training data has expanded significantly from 3 million in version 1.0 to 150 million, creating a virtuous cycle between data and the model.

Technological Advantage:

Key Features of LingBot-Depth 2.0

• In-Depth Autocomplete for Complex Materials: It effectively overcomes the “hard scenes”—such as glass, mirrors, and transparent objects—where traditional depth cameras often fail, and is capable of generating 3D depth maps with complete structures and clear boundaries.
• Precise Spatial Awareness: It delivers outstanding performance in edge clarity, recognition of small objects (such as cables and thin rods), and long-range depth estimation, thereby preventing obstacle-avoidance blind spots caused by the robot missing objects.
• High Precision and High Robustness: In the most challenging indoor scenarios involving large-scale depth loss, the depth error (RMSE) was halved compared to the previous generation (from 0.132 to 0.062), and the system ranked first in 12 out of 16 depth completion benchmark evaluations.

Use Cases for LingBot-Depth 2.0

• Embodied Intelligence and Robotics: Provides a reliable spatial data foundation for robot navigation, obstacle avoidance, and precise grasping (such as grasping transparent storage bins, glass cups, and champagne towers).
• industrial automation: Enables tasks such as precision assembly, alignment of highly reflective metal parts, and tracking the edges of cardboard boxes in logistics and warehousing.
• AR/VR and Autonomous Driving: Provides precise depth information to enhance safety and user experience in environments with complex lighting and materials.

LingBot-Depth 2.0 project repository

• Project website::https://technology.robbyant.com/lingbot-vision
• GitHub repository::https://github.com/robbyant/lingbot-vision
• HuggingFace Model Library::https://huggingface.co/collections/robbyant/lingbot-vision
• arXiv Technical Paper::https://github.com/robbyant/lingbot-vision/blob/main/paper.pdf

How do I use LingBot-Depth 2.0?

• Get the source code: The model weights for LingBot-Vision (available in four versions: ViT-G, L, B, and S) and its technical report have been fully open-sourced; developers can access them via GitHub, Hugging Face, or ModelScope.
• Hardware and SDK Integration: Ant Lingbo has entered into a deep partnership with Orbbec. Users can enable spatial perception capabilities with a single click using Orbbec’s SDK, or plan to purchase an all-in-one “3D camera + spatial perception” camera product—which integrates the commercial version of this model—by the end of the year for out-of-the-box use.
• Cloud Deployment: Developers can also leverage AI computing platforms such as Baidu Baige to quickly run models and perform inference using the official Docker images provided.

Comparison of similar products