Ant Group Open Sources LingBot-Vision, Giving Robots Spatial Awareness
In embodied intelligence, enabling robots to perceive physical space with human-like precision remains a fundamental challenge. Robbyant, an Ant Group subsidiary specializing in embodied AI, has officially open-sourced the LingBot-Vision model family. This suite of self-supervised vision Transformers achieves exceptional results in dense spatial perception through an innovative "boundary modeling" strategy, outperforming larger models across multiple benchmarks despite having fewer parameters.
Most current visual foundation models prioritize "object recognition," focusing on identifying what is in an image while often overlooking critical physical interaction cues like object boundaries, contours, and depth. LingBot-Vision inverts this approach by treating "boundaries" as primary pre-training signals. By leveraging mask boundary modeling, the model identifies the most information-dense regions in an image and centers its training on them. This method allows the model to acquire semantic understanding while simultaneously developing robust geometric spatial perception.

Regarding performance, the flagship LingBot-Vision model, ViT-g/16, contains only 1.1 billion parameters yet achieves state-of-the-art results in depth estimation tasks, including NYU-Depth v2. It surpasses DINOv3, which has 7 billion parameters, while using a training dataset roughly one-third the size. For resource-constrained deployment scenarios, the series offers distilled versions ranging from 3 billion parameters down to smaller sizes, ensuring top-tier dense prediction performance across various hardware configurations.
To showcase the technology's practical value, the development team also upgraded the depth completion system to LingBot-Depth 2.0. Tests indicate significant accuracy improvements when handling transparent objects, a traditional perception blind spot. As data volume increases, LingBot-Vision's performance continues to scale without the saturation typical of traditional models, further validating the potential of its boundary-centric spatial perception architecture in complex real-world environments.
LingBot-Vision is now fully open-sourced on Hugging Face under the Apache-2.0 license, including weights and inference code for four model sizes, from large to small. This technology empowers developers to equip robots with more sensitive physical perception capabilities at lower computational costs, paving the way for a more precise and interactive future in embodied intelligence.
Related article
Universal Music, ElevenLabs Forge Co-Created AI Music Platform for Fans and Artists
Universal Music Group (UMG) announced on the 10th (US time) that it has entered into a multi-year licensing agreement and a broader strategic partnership with ElevenLabs, an AI audio research and product company. This marks ElevenLabs' first collabor
How does Yahoo Japan SEO ranking work in 2024?
AI-powered image search has evolved from a novelty into a daily essential for millions, yet most users operate under outdated assumptions.The disconnect between what this technology actually achieves and public perception is surprisingly vast.While A
Zuckerberg expands Meta’s enterprise AI vision beyond agents
In June, Meta launched a new AI agent for enterprises, designed to assist with customer service, support, and daily operations. However, CEO Mark Zuckerberg revealed to investors on Wednesday’s second-quarter earnings call that the company’s enterpri
Related Special Topic Recommendations
Comments (0)
0/500
In embodied intelligence, enabling robots to perceive physical space with human-like precision remains a fundamental challenge. Robbyant, an Ant Group subsidiary specializing in embodied AI, has officially open-sourced the LingBot-Vision model family. This suite of self-supervised vision Transformers achieves exceptional results in dense spatial perception through an innovative "boundary modeling" strategy, outperforming larger models across multiple benchmarks despite having fewer parameters.
Most current visual foundation models prioritize "object recognition," focusing on identifying what is in an image while often overlooking critical physical interaction cues like object boundaries, contours, and depth. LingBot-Vision inverts this approach by treating "boundaries" as primary pre-training signals. By leveraging mask boundary modeling, the model identifies the most information-dense regions in an image and centers its training on them. This method allows the model to acquire semantic understanding while simultaneously developing robust geometric spatial perception.

Regarding performance, the flagship LingBot-Vision model, ViT-g/16, contains only 1.1 billion parameters yet achieves state-of-the-art results in depth estimation tasks, including NYU-Depth v2. It surpasses DINOv3, which has 7 billion parameters, while using a training dataset roughly one-third the size. For resource-constrained deployment scenarios, the series offers distilled versions ranging from 3 billion parameters down to smaller sizes, ensuring top-tier dense prediction performance across various hardware configurations.
To showcase the technology's practical value, the development team also upgraded the depth completion system to LingBot-Depth 2.0. Tests indicate significant accuracy improvements when handling transparent objects, a traditional perception blind spot. As data volume increases, LingBot-Vision's performance continues to scale without the saturation typical of traditional models, further validating the potential of its boundary-centric spatial perception architecture in complex real-world environments.
LingBot-Vision is now fully open-sourced on Hugging Face under the Apache-2.0 license, including weights and inference code for four model sizes, from large to small. This technology empowers developers to equip robots with more sensitive physical perception capabilities at lower computational costs, paving the way for a more precise and interactive future in embodied intelligence.
Universal Music, ElevenLabs Forge Co-Created AI Music Platform for Fans and Artists
Universal Music Group (UMG) announced on the 10th (US time) that it has entered into a multi-year licensing agreement and a broader strategic partnership with ElevenLabs, an AI audio research and product company. This marks ElevenLabs' first collabor
Zuckerberg expands Meta’s enterprise AI vision beyond agents
In June, Meta launched a new AI agent for enterprises, designed to assist with customer service, support, and daily operations. However, CEO Mark Zuckerberg revealed to investors on Wednesday’s second-quarter earnings call that the company’s enterpri





Home






