Robbyant has released LingBot-Vision, a groundbreaking 1B-parameter vision model that outperformed the larger 7B DINOv3 model on depth tasks, achieving an RMSE of 0.296 compared to DINOv3’s 0.309, while using significantly fewer parameters. Unlike traditional vision models, which focus on high-level semantics and often struggle with physical perception, LingBot-Vision is trained around object boundaries. This innovative approach enables the model to learn boundary-bearing visual tokens independently, enhancing its capabilities in tasks critical for robotics, such as accurately detecting object edges and depth discontinuities.

DINOv3: DINOv3 is a 7B-parameter vision foundation model referenced as a performance benchmark. It is trained primarily for semantic stability across visual changes. In the news, it serves as the comparison point that LingBot-Vision surpasses on depth estimation despite using far fewer parameters.
Robbyant: Robbyant is the entity that developed and released LingBot-Vision as an open-weights model. It focuses on shifting visual pre-training toward boundary awareness to address limitations of semantics-only models. The release positions Robbyant as the creator of this new paradigm for visual foundation models suited to physical and robotic applications.
LingBot-Vision: LingBot-Vision is the world’s first spatially-native visual foundation model released by Robbyant. It is trained to learn boundary-bearing visual tokens without human labels or external detectors, enabling the model to capture scene structure at the patch level. In the news, this approach allows the 1B-parameter model to outperform larger models on depth, segmentation, and tracking tasks when used frozen with a single linear layer.

`json
{
“Model Comparison”: “A boundary-focused smaller model can outperform larger models on tasks requiring detailed scene understanding without the need for fine-tuning.”,
“Training Paradigm”: “Vision foundation models trained on object boundaries develop cleaner features at the patch level, preserving object layout stability during movement.”,
“Robotics Applications”: “Boundary-aware visual tokens tackle challenges in physical perception like precise detection of object edges, handling of transparent surfaces, and depth variability.”
}
`