MM-ABC: Towards Generalist Mobile Manipulation via Seeing, Coordinating and Imagining
MM-ABC coordinates arm and base control for generalist mobile manipulation in simulation and on real robots.
MM-ABC is a mobile-manipulation foundation model that combines sparse multi-level VLM features, a training-only future-imagination branch, and MM-APT coordination of separate arm and base action streams. It is pretrained on more than 5,000 hours of heterogeneous robot data spanning over 400,000 episodes, 12 datasets, and 17 embodiments. Reported success rates are 44.71% on EBench, 61.2% on RoboCasa365, 99.1% on LIBERO, 82.8% on LIBERO-Plus without perturbation training, and 83% mean success on five real-world tasks. In ablations, replacing clean-action prediction with velocity prediction lowers RoboCasa365 composite-seen success from 32.8% to 29.2%.
- Pretrained on 5,000+ hours, 400K+ episodes, 12 datasets, and 17 embodiments
- Reports 61.2% on RoboCasa365, 99.1% on LIBERO, and 82.8% on LIBERO-Plus
- Reaches 44.71% on EBench and 83% mean success on five real tasks
- Velocity prediction drops composite-seen success from 32.8% to 29.2%
From the original article
1 min read at huggingface.co
Mobile manipulation extends robot interaction beyond a fixed kinematic workspace by making the reachable region itself controllable. This flexibility introduces two central challenges: spatially grounded perception under continuous ego-motion and coordinated control of heterogeneous arm and base actions.
By Qiwei Liang, Guangyu Chen, Shaolong Zhu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.