Robbyant запустила LingBot-Depth 2.0 для восприятия стекла роботами

Robbyant, a division of Chinese company Ant Group focused on embodied artificial intelligence, has launched two new computer vision models: LingBot-Depth 2.0 and LingBot-Vision. These models are designed to solve one of robotics' long-standing and technically complex challenges — the ability of robots to correctly perceive and interact with transparent objects, including glass, mirrors, and other glass surfaces.

How robots missed glass

For a robot's computer vision, glass and mirrors are a classic "blind spot." Traditional machine vision systems rely on infrared sensors (active IR systems) and LiDARs, which emit light or radio waves and analyze their reflection from objects. Transparent materials transmit or scatter these signals unexpectedly, making depth perception and object location inaccurate.

As a result, a robot in the real world can collide with a glass door or miss a glass cup when attempting to grasp it. This problem is particularly acute for robotic applications in logistics, e-commerce, and home automation, where glass is ubiquitous.

What the new Robbyant models can do

LingBot-Depth 2.0 is the next generation of a spatial perception model that relies on computer vision rather than hardware sensors alone. The model is trained to recognize and perceive glass, mirrors, and semi-transparent objects, enabling robots to build a more accurate three-dimensional map of their surroundings.

LingBot-Vision is a foundational computer vision model that serves as the basis for developing other specialized visual AI systems. By analogy with how large language models like GPT are used as the foundation for creating specialized assistants, LingBot-Vision can be adapted for various robotic tasks, from warehouse automation to home robotics.

• LingBot-Depth 2.0 — next-generation spatial perception model for perceiving transparent objects
• LingBot-Vision — foundational computer vision model for robotics
• Target problem: traditional sensors cannot see glass and mirrors
• Approach: shift to computer vision instead of full dependence on IR/LiDAR
• Developer: Robbyant (a division of Ant Group, China)

Context: the race in embodied AI

The launch of LingBot marks an intensification of global competition between AI labs in developing perception capabilities for physical robots. As computer vision models become increasingly powerful (especially multimodal models like GPT-4V or Claude 3.5 Vision), robotics companies are moving away from full dependence on specialized hardware sensors toward more universal vision-based perception systems.

This strategy suits Ant Group well, which has significant experience in logistics, e-commerce (through Alibaba), and fintech. All these areas require mass robot automation, including manipulation of small objects and navigation in complex environments.

What it means

The launch of LingBot-Depth 2.0 and LingBot-Vision shows that robot computer vision has matured enough to solve real, specific engineering problems. This could lead to cheaper and more versatile robotic systems that rely on powerful AI models instead of expensive specialized sensors. In the future, robots will be able to work in ordinary domestic and industrial environments without specialized equipment.