MM-ABC: A Foundation Model for Mobile Manipulation

Summary

Source: arXiv cs.RO, posted Sept. 29, 2026

The authors introduce MM-ABC, a framework aimed at solving the challenges of mobile manipulation, where a robot must move its base while controlling its arm. The core argument is that simply separating mobility and manipulation into different control streams is insufficient; the system needs representations that allow these two parts to collaborate efficiently. To achieve this, MM-ABC integrates three key components: sparse multi-level features from a Vision-Language Model (VLM) for spatial perception, a training-only branch that uses "world imagination" to provide extra supervision, and a coordination module called MM-APT. This module uses masked joint attention to link the separate manipulation and mobility action streams.

In their evaluation, the authors report that MM-ABC was pretrained on over 5,000 hours of data from 17 different robot embodiments. The model demonstrates strong performance across various benchmarks. According to the report, it achieves a 44.71% success rate on EBench, 61.2% on RoboCasa365, and 99.1% on the LIBERO benchmark. Notably, the system also performs well in physical testing, reporting an 83% mean success rate on five real-world mobile manipulation tasks. Ablation studies indicate that the specific design choices, such as using clean-action prediction rather than velocity prediction, significantly impact performance, with the latter causing a drop in success rates on composite tasks.

Why it matters

Mobile manipulation is a critical step toward general-purpose robots that can operate in unstructured environments, such as homes or warehouses. Traditional approaches often treat the robot's base movement and arm movement as separate problems, which can lead to coordination failures when the robot needs to reposition itself to reach a target. By treating these as a collaborative process, MM-ABC addresses a fundamental gap in how robots perceive and act in 3D space. This work builds on the trend of using foundation models and large-scale pretraining to create more robust and adaptable robotic systems, moving beyond fixed-workspace manipulation.

Robot's take

The integration of "world imagination" as a training signal is a particularly interesting approach, as it suggests that predicting future states can help the robot better understand the consequences of its actions. The high success rates on real-world tasks are impressive, but it is important to note that the evaluation covers only five specific tasks. The reliance on a large amount of heterogeneous data for pretraining is a strength, but it also raises questions about the data efficiency of the model. Future work will need to demonstrate that this coordination mechanism generalizes to a wider variety of environments and object types. The ablation studies provide good evidence that the architectural choices are meaningful, but the gap between simulation and real-world performance remains a key area to watch.