Hugging Face Releases MM-ABC: A Groundbreaking Foundation Model for Mobile Manipulation Robots

Summary:Hugging Face has released MM-ABC, a foundation model for mobile manipulation robots, designed to address the challenges of spatial perception and heterogeneous action coordination in robot interactions. MM-ABC leverages sparse multi-level Vision-Language Model (VLM) features for spatial perception and incorporates a future branch and MM-APT module to enhance perception and manipulation-intent prediction. Experimental results demonstrate MM-ABC's strong performance across multiple benchmarks, inc

---

Overview

Hugging Face has recently released MM-ABC, a foundation model for mobile manipulation robots, aimed at addressing the core challenges of spatial perception and heterogeneous action coordination in robot interactions. MM-ABC achieves this through the following three key components:

1. Sparse Multi-level VLM Features: Used for spatial perception, enabling precise spatial localization and understanding under continuous ego-motion.
2. Future Branch: Utilized only during training, it enhances perception and manipulation-intent prediction by using world imagination and geometric intent as extra supervision.
3. MM-APT Module: Coordinates separate manipulation and mobility streams through masked joint attention and clean-action prediction.

Technical Highlights

• Efficient Cross-Stream Collaboration: MM-ABC not only decouples manipulation and mobility actions but also supports efficient cross-stream collaboration, thereby improving the overall learning signal.
• Multi-level Supervision: With the introduction of the future branch and geometric intent, MM-ABC excels in perception and manipulation-intent prediction.
• Large-scale Pretraining: MM-ABC has been pretrained on over 5,000 hours of heterogeneous robot data, covering 400K+ episodes, 12 datasets, and 17 different robot embodiments.

Experimental Results

MM-ABC demonstrates strong performance across multiple benchmarks:

• EBench: Success rate of 44.71%.
• RoboCasa365: Success rate of 61.2%.
• LIBERO: Success rate of 99.1%.
• LIBERO-Plus: Success rate of 82.8% without perturbation training.
• Real-world Tasks: Average success rate of 83%.

Industry Impact

The release of MM-ABC marks a significant advancement in the field of mobile manipulation robotics. Its efficient spatial perception and action coordination capabilities make it highly effective in complex tasks, providing a powerful tool for researchers and developers. Additionally, the architectural design of MM-ABC offers new insights for future development of robotic agents, particularly in the areas of multimodal interaction and complex environment adaptation.

Developer Recommendations

• Model Application: Researchers and engineers are encouraged to use MM-ABC in mobile manipulation robot projects to improve task execution efficiency and success rates.
• Extended Research: Further exploration of MM-ABC's performance in different robot embodiments and task scenarios is recommended, along with optimization through other AI technologies such as reinforcement learning.
• Community Participation: Developers are encouraged to participate in the MM-ABC open-source community, share usage experiences and improvement suggestions, and contribute to the model's development and application.

---

Source: Hugging Face Daily Papers (2026-09-28)

— END —

Tags: #Hugging Face Follow #Mobile Manipulation Follow #Robotics Follow #VLM Follow #AI Model Follow