Robbyant has released LingBot-Depth, an open-source depth refinement model that takes raw, incomplete output from commodity depth sensors and converts it into metric-accurate 3D geometry. The model processes RGB images and depth maps together in a shared latent space, producing both a refined depth map and a calibrated point cloud. Pretrained weights are available on Hugging Face under the Apache 2.0 license.

The underlying method is masked depth modeling, a self-supervised approach that trains a Vision Transformer encoder to reconstruct missing depth from partial observations. The encoder uses cross-modal attention to pull geometric context from RGB when depth data is absent or noisy. The interesting part is the metric-scale preservation: unlike relative depth methods that output normalized values, LingBot-Depth returns measurements in meters, meaning downstream systems can use the output directly without a separate calibration step.

Robbyant trained the model on a 3 million sample RGB-D dataset, split between 2 million real-world captures from residential, office, and commercial environments and 1 million photo-realistic simulated renders. The capture rig combined Intel RealSense, Orbbec Gemini, and Azure Kinect sensors. The full dataset is publicly available on Hugging Face and ModelScope.

The team demonstrated four applications: filling incomplete depth maps, high-fidelity indoor scene reconstruction, 4D point tracking of dynamic objects in metric space, and dexterous robotic grasping. Grasping tests covered steel cups, glass cups, storage boxes, and toy cars. The 4D tracking demonstrations used gym equipment, including scooters, rowing machines, and pull-up bars, as dynamic targets.

The repository includes eight example scenes and a PyTorch inference API requiring PyTorch 2.0 or later and a CUDA-capable GPU. A depth-completion-optimized variant, lingbot-depth-postrain-dc-vitl14-v0.5, is available alongside the base pretrained model. The paper was accepted to ECCV 2026 in June. Robbyant has not published head-to-head benchmark results against other metric depth completion methods in the public documentation, which means practitioners evaluating the model against commercial or research alternatives will need to run their own comparisons.