Embodied Intelligence Breakthrough: AutoNavi Fully Open-Source General Robot Base Model ABot-M0

null

Embodied AI has entered a milestone-level breakthrough. Today, AOTD officially announced the full open-sourcing of the world’s first robot embodied operation base model, ABot-M0, built on a unified architecture. The model’s core positioning is to achieve “one general-purpose brain that can adapt to multiple robot form factors,” aiming to break down barriers between heterogeneous hardware and accelerate embodied AI from the lab to industrial and home scenarios.

Core technologies and performance

ABot-M0 has demonstrated outstanding performance across multiple authoritative industry benchmark tests. Data shows that the model achieves a task success rate of 80.5% on the Libero-Plus benchmark, improving by nearly 30% compared with Pi0, the industry benchmark方案 from before. In addition, it has also set new SOTA (the industry’s top-tier) records in tests such as Libero and RoboCasa.

Full open-sourcing across three dimensions

To address the long-standing pain points in the embodied AI field—“data silos” and “deployment difficulties”—AOTD’s open-sourcing this time covers three dimensions: underlying data, core algorithms, and pre-trained models:

Data layer: It open-sources UniACT, currently the largest general-purpose robot dataset. This dataset integrates over 6 million real-world manipulation trajectories and provides a complete end-to-end processing pipeline, including transforming heterogeneous data into standardized training data.

Algorithm layer: It simultaneously releases the model architecture and training framework. Key highlights include AOTD’s innovative Action Manifold Learning (AML) algorithm and a dual-stream perception architecture, endowing robots with excellent spatial understanding and action execution capabilities.

Model layer: It provides end-to-end pre-trained models and a complete toolchain. Developers can get “plug-and-play” functionality without having to build the framework from scratch, significantly lowering the barrier to adapting robots for industrial collaboration or home service.

Industry impact

AOTD’s technical负责人 of ABot-M0 stated that truly general-purpose embodied intelligence requires the joint refinement of developers worldwide. AOTD’s open-sourcing of ABot-M0 is not only a technical share, but also an effort to build a bridge connecting academic research and industrial applications—so that every robot of a different form factor can have a smart, reliable, and general-purpose “brain.”