The open-source vision foundation model improves spatial perception for robotics through boundary-centric learning and self-supervised pretraining
Robbyant by Ant Group has open-sourced LingBot-Vision, a self-supervised vision foundation model designed to improve dense spatial perception for robotics and embodied AI applications. The model uses boundary-centric pretraining that emphasises learning of object boundaries and geometry rather than relying primarily on semantic features.
The flagship ViT-g/16 model is built with approximately 1.1 billion parameters and uses a novel masked boundary modelling strategy during training on a curated dataset of approximately 161 million images. The company claims that the methodology does not require human-labeled datasets, edge detectors, or pretrained backbones for training. Smaller models have also been created using ViT-L, ViT-B, and ViT-S architecture for deployments with low compute needs.
As opposed to standard vision foundation models that focus mostly on learning semantic information, LingBot-Vision is developed to retain fine-grained spatial information such as object boundaries and depth discontinuities. These dense feature representations can be used in downstream applications such as depth estimation and robotic perception.
When evaluated on several benchmarks, the 1.1-billion-parameter model achieved an RMSE score of 0.296 on the NYU-Depth v2 benchmark, outperforming larger models such as the 7B DINOv3 model. It also delivered competitive results on the KITTI benchmark despite being trained on significantly fewer images than comparable models.
The model serves as the visual backbone of LingBot-Depth 2.0, which is the most recent depth completion solution from Robbyant. The company has open-sourced the model weights, technical report and inference code under the Apache 2.0 licence.
