Breaking

Ant-backed Robbyant open-sources depth completion model LingBot-Depth

June 29, 2026 at 01:59 EDT

Open Source
Robotics
Research & Papers

Robbyant, an embodied-AI company under Ant Group, has open-sourced LingBot-Depth, an AI model that turns noisy, incomplete depth sensor data into high-quality, metric-accurate 3D measurements. Since the model and paper release in January 2026, the team added a large-scale dataset in March and shipped a v0.5 bug-fix release in June. The model is available immediately under the Apache 2.0 license via Hugging Face, ModelScope and GitHub.

January 27, 2026 · Robbyant (Ant Group)

LingBot-Depth: Turning Noisy Sensors Into Clean 3D

An open-source (Apache 2.0) spatial-perception model that fuses raw RGB-D data with visual context to repair missing depth on glass, mirrors and low-texture surfaces — outputting metric-scale depth maps and 3D point clouds out of the box.

40%+

RMSE improvement in hard cases vs major models

3.02M

Training samples (real-world + simulation)

Apache 2.0

Model + dataset fully open-sourced

Training data mix — ~3.02M samples

1.4M

RobbyReal

990k

RobbySim

580k

RobbyVla

38k

Validation

How it works — the pipeline

Inputs
RGB image + metric depth + camera intrinsics

→

RGB-D fusion
ViT-L/14 encoder, DINOv2 backbone, cross-modal attention

→

Outputs
Refined metric depth + 3D point cloud

What works well

Handles glass, mirrored & low-texture surfaces
Runs out of the box with RealSense, Orbbec, ZED
Strong gains on sparse depth completion

Caveats

Inference cost of a ViT-Large-class model
Training data skews indoor-centric
Requires a CUDA GPU (PyTorch 2.0+)

Where it lands

Recommended model v0.5 (general purpose), with a DC variant for sparse completion. Demonstrated on grasping steel cups, glass and toys, and 4D tracking of dynamic human motion. Accepted to ECCV 2026, with chip-level depth affinity via the Orbbec Gemini 330.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.