Hugging Face Releases MambaAct: Efficient Flow-Matching Vision-Language-Action Model Solution
Summary:Hugging Face introduces MambaAct, addressing the inference efficiency bottleneck in flow-matching Vision-Language-Action (VLA) models for real-time control. By replacing the traditional Transformer-based action expert with a compact Mamba architecture while retaining the visual-language prefix encoder, MambaAct reduces the action expert's parameter size by a factor of 10 and significantly cuts down inference latency. In LIBERO and real-world deployments, MambaAct achieves a 64.8% and 58.8% laten
Technical Mechanism Analysis
The core innovation of MambaAct lies in replacing the traditional Transformer-based action expert with a Mamba-based compact action expert. The Mamba architecture is inherently advantageous in handling long-range dependencies, making it particularly suitable for the iterative evaluation of action experts in flow-matching VLA models. Specifically, MambaAct achieves efficient performance through the following mechanisms:
- Mamba Action Mixer: Mamba serves as the action mixer, modeling dependencies across the noisy continuous-action sequence, thus replacing the traditional Transformer attention mechanism.
- Lightweight Context Compressor: To avoid repeated token-level cross-attention computations, MambaAct introduces a lightweight Context Compressor that maps VLM prefix tokens into a compact conditioning signal for gating and modulating each Mamba action block.
Engineering Trade-offs and Performance
MambaAct achieves significant optimizations in several aspects:
- Parameter Size and Computational Efficiency: The action expert's parameter size is reduced to 1/10 of the original, significantly lowering computational costs.
- Latency Reduction: In the LIBERO benchmark, MambaAct reduces flow-matching denoising latency by 64.8%; in real-world deployment, it achieves a 58.8% reduction.
- Task Success Rate: Across three independent fine-tuning seeds, MambaAct improves the average success rate on LIBERO by 1.08 percentage points and maintains performance gains on LIBERO-Plus and real-world benchmarks.
However, some trade-offs accompany MambaAct's optimizations. For instance, the Mamba architecture may face memory bottlenecks when handling extremely long sequences. Additionally, while latency is significantly reduced, the robustness of MambaAct in certain extreme scenarios needs further validation.
Developer Deployment and Implementation Recommendations
For developers, the release of MambaAct offers several opportunities:
- Real-time Control Applications: MambaAct is particularly suitable for real-time control applications that require low latency and high task success rates, such as robotic manipulation and automation systems.
- Model Optimization: Developers can leverage MambaAct's architectural improvements to enhance the inference efficiency of existing VLA models.
- Cross-Domain Migration: The architectural enhancements in MambaAct can be migrated to other domains that require efficient sequence processing, such as natural language processing and video analysis.
It is recommended that developers pay attention to the memory usage of MambaAct and fine-tune it according to specific application scenarios to fully leverage its performance advantages.
— END —Source: Hugging Face Daily Papers (2026-10-16)
Tags: #Hugging Face #MambaAct #VLA Models #Real-time Control #Model Optimization
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments