Gaode officially fully open-sources ABot-M0: a brain adaptable to multi-form embodied robots

Tech Report April 1: GaoDe announced that it has officially fully open-sourced ABot-M0, the world’s first robot embodiment operation base model based on a unified architecture.

The model implements a universal “brain” that adapts to various forms of embodied robots, providing an entirely new universal technical foundation for the development of the embodied intelligence field.

ABot-M0 model architecture diagram

ABot-M0’s technical strength has been fully validated in authoritative tests. It achieves the best performance across multiple benchmark tests, including Libero, Libero-Plus, RoboCasa, and others. In the Libero-Plus benchmark, the task success rate reaches 80.5%, which is nearly a 30% improvement over the industry’s previous benchmark solutions. Its performance in space understanding and task execution is especially impressive.

ABot-M0 evaluation results on Libero-Plus

This open-sourcing covers three core dimensions—data, algorithms, and models—breaking through key industry development challenges in an all-around way.

On the data layer, the open-sourced UniACT, the largest-scale general-purpose robot dataset, integrates over 6 million real-world operation trajectories. With standardized processing, heterogeneous robot data from around the world can be used uniformly;

On the algorithm layer, it innovatively introduces the motion manifold learning algorithm and the dual-stream perception architecture. The former allows the model to directly predict a sequence of actionable motions, improving decoding efficiency. The latter addresses the shortcomings of 3D inference and strengthens spatial understanding capabilities;

On the model layer, it open-sources end-to-end pretrained models and a complete toolchain, enabling developers to quickly adapt to diverse scenarios such as industrial and home environments.

ABot-M0’s unified architecture successfully verifies the feasibility of a brain-driven approach that can support robots in multiple forms, providing an out-of-the-box solution for deploying embodied intelligence technology.

【This article ends】If you need to reprint it, you must indicate the source: Tech Report

Editor: Sui Xin