ABot-M0: A Unified VLA Framework Solving the One‑Brain Many‑Forms Robotics Challenge

Overview

ABot-M0 is an open‑source unified framework designed to address the “one‑brain, many‑forms” problem in robotics, providing a Vision‑Language‑Action (VLA) foundation model that integrates perception and manipulation.

Key Contributions

UniACT dataset : Consolidates multiple fragmented robot manipulation datasets into a single, standardized training resource, enabling large‑scale, diverse data for better generalization.

Action Manifold Learning (AML) : Replaces noisy action prediction with direct prediction of smooth, physically plausible action trajectories on a low‑dimensional manifold, improving decoding speed and policy stability.

Plug‑and‑play dual‑stream perception : Combines a powerful VLM (Qwen3‑VL) for semantic understanding with an optional 3D module (e.g., VGGT) for geometric priors, allowing “what to do” and “where to do it” to be handled jointly without modifying the backbone.

Open‑source Release

The full code, data processing pipeline, and pretrained weights are released on GitHub ( https://github.com/amap-cvlab/ABot-Manipulation) together with a project website. The repository also links to the companion open‑source project starVLA for pretrained models.

Benchmark results show ABot‑M0 achieving first place on the LIBERO‑PLUS and RoboCasa‑GR1‑Tabletop leaderboards.

Problem Addressed

Current robot learning suffers from three major issues: insufficient data scale, inconsistent data quality, and non‑unified representations. ABot‑M0 provides an end‑to‑end pipeline that transforms heterogeneous raw data into efficient, generalizable policies, making cross‑hardware and cross‑task embodied intelligence feasible.