In a strategic move to democratize advanced spatial perception, Robbyant’s CEO announced open-sourcing LingBot-Depth and forming partnerships with hardware leaders to lower barriers for embodied AI across homes, factories, and warehouses.

Robbyant, an embodied AI company under Ant Group, released LingBot-Depth as a high-precision depth sensing and 3D environment understanding model to boost robots’ performance in complex real-world settings.

LingBot-Depth achieves robustness by leveraging Orbbec’s Gemini 330 stereo cameras and MX6800 depth engine, reconstructing missing data for challenging lighting and reducing sensor latency through on-device computation.

The model is designed to be hardware-compatible with existing sensors, avoiding the need for form-factor changes to accelerate adoption.

Len Zhong of Orbbec emphasizes the close coupling between Orbbec’s depth data from Gemini 330 and LingBot-Depth, underscoring stable, high-fidelity data as foundational.

Robbyant trained LingBot-Depth on roughly 10 million raw samples, building a curated 2 million RGB-depth pairs dataset, with plans to open-source the dataset to spur broader community innovation.

The model tackles depth gaps from transparent or reflective surfaces with Masked Depth Modeling, using RGB texture, contours, and scene context to reconstruct missing information.

Robbyant intends a strategic partnership with Orbbec to embed LingBot-Depth into Orbbec’s next‑generation depth cameras for embodied intelligence applications.

Beyond Orbbec, Robbyant seeks to broaden its ecosystem by sharing spatial perception capabilities with additional hardware partners to enable real-world deployment of intelligent robots in dynamic environments.

Key resources include the LingBot-Depth codebase on GitHub, a technical report, and a HuggingFace page detailing the project.

Orbbec contributed hardware resources and expertise, with LingBot-Depth co-optimized on Orbbec platforms and validated in Orbbec’s Depth Vision Laboratory.

On standard benchmarks like NYUv2 and ETH3D, LingBot-Depth outperformed major rivals by cutting indoor scene relative error by over 70% and lowering RMSE by about 47% on sparse Structure-from-Motion tasks.