Summary
ABot-M0 (Alibaba AMAP CV Lab) addresses the “one-brain, many-forms” challenge in embodied AI by building a systematic data curation pipeline and unified training strategy over six public datasets. The resulting UniACT-dataset contains over 6 million trajectories and 9,500 hours of data spanning 20+ robot morphologies. The core technical contribution, Action Manifold Learning (AML), predicts clean actions directly instead of denoising noise, yielding a more stable and efficient policy representation.
Key Contributions
- UniACT-dataset: 6M+ trajectories, 9,500+ hours, 20+ robot embodiments, curated from 6 public datasets
- Action Manifold Learning (AML): direct clean-action prediction replacing standard noise-prediction diffusion
- Plug-and-play 3D spatial understanding module for improved precision on complex manipulation tasks
- Joint optimization of data curation pipeline, model architecture, and training strategy
Significance
ABot-M0 demonstrates that systematic data curation combined with AML can rival or surpass larger models with ad-hoc training, positioning Alibaba’s approach as a scalable path to cross-embodiment generalist manipulation.