Chinese AI firm launches next-gen spatial perception models for robots

Robbyant said the launch represents a major step forward in robotic spatial perception, helping robots develop a more accurate understanding of the physical world and improve their ability to navigate complex environments.

Building on the foundation of LingBot-Depth, which introduced the Masked Depth Modeling (MDM) technique to address challenges in depth sensing for transparent and reflective surfaces, LingBot-Depth 2.0 brings expanded training capabilities and improved performance. 

The model was trained on 150 million samples and achieved leading results across 12 of 16 depth completion benchmarks, demonstrating stronger accuracy and reliability in real-world perception tasks.

AI model designed to overcome robotic depth sensing challenges

LingBot-Depth 2.0 delivers significant improvements in challenging indoor environments where depth perception is often limited. According to Robbyant, the model reduces depth error by more than half compared with its predecessor in scenarios with severe depth loss, lowering the RMSE score from 0.132 to 0.062.

The model also shows stronger performance in areas where traditional depth cameras typically struggle, including detecting and understanding glass surfaces, mirrors, and other transparent objects.

The Chinese firm attributes these advances to LingBot-Vision, a visual foundation model designed to improve how robots interpret and understand their surroundings. The model is the first in the industry to use “boundary structure” as a pre-training objective, allowing it to achieve sub-pixel-level boundary localization and stronger spatial structure understanding for more reliable robotic perception.

Despite being trained on a relatively smaller dataset of 160 million images, LingBot-Vision delivers strong performance compared with larger models. Robbyant said the model also provides stable object boundary detection, enabling robots to continuously track object edges and structures across video sequences.

LingBot-Depth 2.0 gains validation for commercial robotics applications 

Beyond powering LingBot-Depth 2.0, LingBot-Vision is designed as a flexible foundation model that can support a wide range of AI applications and downstream tasks. Robbyant said the model’s capabilities extend beyond depth perception, enabling broader use cases in robotic vision and intelligent systems.

For commercial deployment, LingBot-Depth 2.0 has received certification from the Depth Vision Laboratory of Orbbec. Tests using chip-level depth data from Orbbec’s Gemini 330 series stereo 3D cameras showed improvements in edge detection, object contours, recognition of smaller objects, long-distance depth estimation, and performance under challenging lighting and material conditions.

The collaboration with Orbbec will also extend to new hardware solutions for robotics data collection. As part of Orbbec’s newly launched Robot-Free Data Collection Hardware Platform, the RGB-D EGO device will feature a customized version of LingBot-Depth optimized for collecting high-quality training data. 

In future updates, the platform is expected to integrate an advanced commercial version of the model, further improving depth completion, object boundary detection, and spatial structure understanding. The goal is to provide a more accurate, stable, and practical data foundation for training embodied AI systems in real-world environments.