MM-ABC: Towards Generalist Mobile Manipulation via Seeing, Coordinating and Imagining

MM-ABC. Towards Generalist Mobile Manipulation via Seeing, Coordinating and Imagining. By Qiwei Liang and eight others, from Xspark AI, the Hong Kong University of Science and Technology, Guangzhou, and two other institutions. Published on the twenty-eighth of September. No code repository is linked to it.

Mobile manipulation takes a robot beyond a fixed kinematic workspace, by making the reachable region itself something it can control. That flexibility brings two central challenges. Perception has to stay spatially grounded while the robot's own viewpoint is continuously moving. And control has to coordinate two different kinds of action, those of the arm and those of the base. Existing approaches strengthen geometry through explicit three-D representations or predictive world models, and they often split mobility and manipulation into separate action streams. The authors argue that splitting is not enough on its own. The representations also have to support efficient collaboration between the streams.

MM-ABC is a foundation model built around the three verbs in its title, Seeing, Coordinating, and Imagining, all in service of Arm-Base Collaboration. For seeing, it uses sparse, multi-level VLM features for spatial perception. For imagining, it adds a future branch that exists only during training. For coordinating, it introduces MM-APT, which ties the separate manipulation and mobility streams together.

The future branch uses world imagination and geometric intent as extra supervision. Because it is training-only, it adds a learning signal rather than a step at run time, and it is credited with strengthening perception, improving the prediction of manipulation intent, and improving the overall learning signal. MM-APT, for its part, coordinates the two action streams through masked joint attention, and it relies on clean-action x-prediction.

Controlled ablations test these pieces. Replacing clean-action prediction with velocity prediction lowers success on RoboCasa365 composite-seen tasks from thirty-two point eight percent to twenty-nine point two percent. Removing the future supervision, or the multi-level conditioning, causes larger drops still.

MM-ABC is pretrained on more than five thousand hours of heterogeneous robot data, spanning over four hundred thousand episodes, twelve datasets, and seventeen embodiments. Evaluation covers EBench, RoboCasa365, ManiSkill-HAB, LIBERO, LIBERO-Plus, and real-world mobile manipulation.

Success reaches forty-four point seven one percent on EBench, sixty-one point two percent on RoboCasa365, and ninety-nine point one percent on LIBERO. On LIBERO-Plus it reaches eighty-two point eight percent without any perturbation training. Across five real-world tasks, mean success is eighty-three percent.

Separating arm and base, on this account, is only half of the answer. MM-ABC keeps the streams separate but gives them a shared way to attend to each other, a perception built from features at several levels, and a training signal drawn from imagined futures. The ablations are what show each of those pieces carrying weight.