MambaAct Brings Compact Mamba Experts to Faster VLA Control

Introduction

Vision-language-action (VLA) models give robots a way to connect instructions and visual observations with continuous manipulation. Their practical bottleneck, however, is often not understanding the task but producing actions quickly enough. This is particularly evident in flow-matching policies: the action expert is evaluated repeatedly as the noisy action trajectory is denoised, so its cost is multiplied across inference steps.

MambaAct addresses this bottleneck by changing the part of the policy that is called most frequently. Rather than rebuilding the complete VLA, it retains the pretrained vision-language prefix encoder from π0.5 and replaces the Transformer-based action expert with a compact Mamba-based design.

What changes

• Mamba for action mixing: The new expert uses Mamba to model dependencies across the noisy continuous-action sequence. This provides a sequence-mixing mechanism tailored to the action side of the policy while reducing the size of the expert.
• Compressed conditioning: Repeated token-level cross-attention can be expensive during iterative denoising. MambaAct introduces a lightweight Context Compressor that maps the visual-language prefix tokens into a compact conditioning representation.
• Block-wise modulation: Instead of sending the full prefix sequence through every action block, the compressed signal gates and modulates the Mamba blocks. The action module therefore receives task context without repeatedly processing all visual-language tokens.
• Targeted reuse of pretrained components: The visual-language encoder remains intact, allowing the method to focus its efficiency changes on the repeatedly executed action-generation path.

Reported results

The project reports an action expert roughly 10 times smaller than the original one. On LIBERO, flow-matching denoising latency falls by 64.8%. In a real-robot deployment, the reported reduction is 58.8%. Efficiency does not come at the expense of the stated task results: across three independent fine-tuning seeds, mean LIBERO success improves by 1.08 percentage points, with gains also maintained on LIBERO-Plus and real-world benchmarks.

The result is important because a modest reduction in the cost of one action-expert call can become a substantial end-to-end improvement when that call is repeated throughout denoising. It also suggests that visual-language conditioning does not always require full token-level attention at every stage of action generation.

Why it matters

MambaAct separates two responsibilities that are often tightly coupled in VLA architectures. The prefix encoder supplies a rich representation of the scene and instruction, while the action expert consumes a compact summary and focuses on temporal action structure. This division may make flow-matching policies more suitable for robots with limited compute or strict control-rate requirements.

The available material does not yet establish how the method behaves across hardware configurations, power budgets, denoising schedules, or a broader set of efficient action architectures. Transfer across robot embodiments and task distributions also requires further validation. The project has released its code, while model-weight availability follows the project’s stated release schedule.

Source: Hugging Face Daily Papers