LingBot-Depth 2.0 - A spatial perception model launched by Ant Lingbo Technology

What is LingBot-Depth 2.0

LingBot-Depth 2.0 is a spatial perception model launched by Ant Lingbo Technology for real-world scenarios, with the vision foundation model LingBot-Vision also open-sourced. Training data was expanded from 3 million to 150 million, and it ranked first in 12 of 16 depth completion benchmark evaluations. In indoor scenarios with large areas of missing depth, the RMSE error dropped from 0.132 to 0.062. The model performs outstandingly in scenarios with transparent objects such as glass and mirrors, has passed professional certification by Orbbec, and builds the capability chain for robots from "understanding" to "seeing accurately."

Main features of LingBot-Depth 2.0

• Depth completion: Based on RGB images and sparse depth input, it completes and generates high-precision dense depth maps.
• Transparent object perception: Achieves complete 3D structure reconstruction for transparent/reflective objects such as glass, mirrors, and faucets.
• Fine edge recognition: Possesses sub-pixel-level boundary localization capability, accurately recognizing object contours and fine structures.
• Temporal stability: Maintains temporal consistency and spatial accuracy of depth estimation in video streams.
• Cross-scenario generalization: Covers real-world scenarios including indoor, outdoor, complex lighting, and various materials.

Technical principles of LingBot-Depth 2.0

• Dual-model collaborative architecture: LingBot-Depth 2.0 is not a single network; it adopts a two-level collaborative architecture of "foundation model + task head." The underlying LingBot-Vision is responsible for extracting boundary and spatial structure features from RGB images, while the upper-layer LingBot-Depth completes the mapping from sparse or noisy depth to dense depth maps based on these features. The decoupled design allows visual representation and depth estimation to each focus on the part they are best at, building a complete capability chain for robots from "understanding" to "seeing accurately."
• LingBot-Vision: Boundary structure pretraining: LingBot-Vision is the industry's first vision foundation model that uses "boundary structure" as an explicit pretraining objective. It adopts geometric modeling to learn structural knowledge such as "where objects are, how edges are distributed, and how space is organized," equipping the model with sub-pixel-level boundary localization capability. With pretraining on only 160 million images, its recognition clarity and stability for object contours, surface intersections, and fine structures outperform general vision foundation models that rely on billion-level data.
• Depth completion mechanism: The depth completion network of LingBot-Depth 2.0 is based on the boundary-aware features provided by LingBot-Vision. It identifies depth discontinuity regions through an encoder, then performs multi-scale feature fusion through a decoder, preserving sharp object edges while reasonably interpolating smooth regions. The key is that the model uses explicit boundary structure features as constraints for depth propagation, effectively preventing the "depth bleeding" problem common in traditional depth completion, ensuring correct truncation between foreground and background at object boundaries, while introducing a temporal feature alignment mechanism to ensure inter-frame depth consistency under video stream input.

Follow WeChat and reply "open source" to join theAI open source project communication group

How to use LingBot-Depth 2.0

• Obtain the model: Download the open-source LingBot-Vision weights from HuggingFace or ModelScope.
• Clone the code: Via the GitHub repository https://github.com/robbyant/lingbot-vision Get the code and technical report.
• Environment Setup: Configure the dependency environment according to the documentation and prepare the RGB-D input data.
• Run Inference: Load the model for depth completion inference, or train downstream vision tasks based on LingBot-Vision.
• On-Device Integration: Use Orbbec Gemini 330 series cameras to access LingBot-Depth commercial edition capabilities through the SDK.

Core Advantages of LingBot-Depth 2.0

• Leading Data Scale: Training data reaches 150 million, a 50-fold expansion over the previous generation, covering richer real-world scenarios.
• Excellent Evaluation Results: Won first place in 12 out of 16 depth completion benchmark evaluations.
• Significantly Reduced Error: In indoor scenarios with large areas of missing depth, RMSE error is halved compared with the previous generation.
• Breakthrough in Transparent Scenarios: Particularly outstanding in scenarios where traditional depth cameras fail, such as glass and mirrors.
• Backed by Hardware Certification: Passed professional certification by the Orbbec Depth Vision Laboratory, with significantly improved edge clarity and robustness.

Project Addresses for LingBot-Depth 2.0

• Project Official 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

Comparison of LingBot-Depth 2.0 with Similar Competing Products

Application Scenarios of LingBot-Depth 2.0

• Home Service Robots: Achieve precise navigation and operation in environments such as kitchens and bathrooms where glass, mirrors, and faucets are present.
• Industrial Quality Inspection: Perform three-dimensional dimensional measurement and defect detection on transparent packaging and reflective metal parts.
• Warehousing and Logistics: Identify cargo edges and small objects in warehouses with complex lighting to assist robotic arm grasping.
• Autonomous Driving: Complete depth information in road scenes and improve perception of transparent obstacles such as glass curtain walls.
• Data Collection Devices: Integrated into RGB-D collection devices such as Orbbec EGO to provide precise real-world data for embodied intelligence training.