MambaAct: Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models

MambaAct. Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models. By Khuyen Pham and three others, from MBZUAI and the Technical University of Denmark. Published on the sixteenth of October. Its code repository is worked out: about two hundred source files, with a readme and a licence.

Vision-language-action models look promising for generalist robotic manipulation, but inference cost is a major bottleneck for real-time control. The problem is sharpest in flow-matching models such as pi zero point five. There the action expert is evaluated repeatedly during iterative denoising, so whatever it costs once, it costs many times over for every action. A robot that has to act in real time pays that repeated cost on every control step, and the part of the model that runs most often becomes the part that decides how fast the whole policy can respond.

MambaAct is an efficient action expert built on Mamba, for flow-matching policies of this kind. It keeps the pretrained visual-language prefix encoder of pi zero point five as it is, and replaces only the transformer-based action expert with a compact one based on Mamba. Inside that expert, Mamba acts as the action mixer, modeling dependencies across the noisy sequence of continuous actions.

That leaves the question of conditioning. The action expert still has to know what the vision-language model saw, and the authors want that without repeated cross-attention at the level of individual tokens. Their answer is a lightweight Context Compressor. It maps the prefix tokens from the vision-language model into a compact conditioning signal, and that signal gates and modulates each Mamba action block.

The resulting action expert is ten times smaller. Flow-matching denoising latency drops by sixty-four point eight percent on LIBERO. In real-world deployment the drop is fifty-eight point eight percent.

Speed does not come at the cost of success here. Mean success on LIBERO improves by one point zero eight percentage points, averaged across three independent fine-tuning seeds, and the gains hold on LIBERO-Plus and on real-world benchmarks. Code and model weights are publicly available.

The design choice here is narrow on purpose. The pretrained encoder is left untouched, and the part that runs again at every denoising step is the part that gets made small, which is where a saving multiplies. Conditioning survives the swap through a compact signal rather than repeated attention, so the smaller expert still acts on what the robot sees and is told.

MambaAct. Efficient Mamba Action Experts for Flow-Matching Vision-Language-Action Models. By Khuyen Pham and three others, from MBZUAI and the Technical University of Denmark. Published on the sixteenth of October. This is the long version: the whole paper rather than the abstract.

In flow-matching robot policies, producing the motion is where the time goes. A model like pi zero point five reads the camera images and the instruction once per query, then hands the result to a separate network called the action expert. That expert starts from random noise and refines it into a chunk of actions over ten denoising steps, running once per step. The authors measured the split. On LIBERO, the denoising steps account for seventy point one percent of end-to-end latency. In SmolVLA, whose vision-language backbone is far smaller, the share rises to eighty-six point four percent. Shrinking the part that sees does not fix the part that acts. Earlier efficiency work mostly cut steps, pruned layers or tokens, or shrank the whole model. The expert's own architecture, the authors argue, has been left largely unexplored.

Two pieces of background. The first is flow matching. In training, a clean action chunk is blended with Gaussian noise at a random time, and the expert learns to predict the velocity that points from the clean chunk toward the noise. At inference the policy starts from pure noise and takes ten equal Euler steps backward along that predicted velocity, arriving at an action chunk. The visual-language prefix is computed once. The expert runs ten times.

The second is Mamba. A Transformer mixes a sequence by letting every token attend to every other token. Mamba instead carries a selective state along the sequence, updating it in order. The authors' case for it is inductive bias: an action chunk is ordered and strongly correlated, so passing information along that order may suit it. They offer that as a hypothesis, not a measurement.

MambaAct keeps the pretrained pi zero point five backbone, a SigLIP vision encoder and a Gemma prefix encoder, fine-tuned jointly. What it discards is the Transformer action expert of three hundred and eleven point five million parameters. In its place goes a Mamba expert of twenty-nine point seven million, trained from scratch during ordinary task fine-tuning, with no teacher, no distillation and no extra pretraining.

Rather than letting the expert attend to every prefix token at every step, a Context Compressor shrinks the prefix. Visual tokens are split into contiguous segments and averaged within each, ignoring padded positions. Language tokens are pooled the same way, and a few learned queries use attention to extract global summaries. The result is sixty-four context tokens: forty-four visual, sixteen language and four global. Those are averaged once more into a single condition vector, computed once per query.

The expert itself is four gated Mamba blocks of width five hundred and twelve. The noisy chunk is projected into one token per action step, with no separate robot-state token. In each block, an embedding of the flow time is added to every action token, the sequence passes through Mamba, and then the condition vector acts through FiLM. A small network turns it into a feature-wise scale and shift for the Mamba output, plus a gate deciding how much of that output joins the residual stream. Scale and shift start at zero, so the modulation begins as the identity, and the gate starts the new Mamba branch at a fifth of its output: damped but active. Because scale, shift and gate depend only on the condition vector, they are reused across denoising steps. Only the time signal and the action tokens are recomputed.

On LIBERO, across three independent fine-tuning seeds, average success rises from ninety-six point seven two percent for the matched baseline to ninety-seven point eight. Denoising latency falls by sixty-four point eight percent and end-to-end latency by fifty-two point nine. Zero-shot on LIBERO-Plus the overall score edges up by one point, with the clearest gain when the robot's initial pose is perturbed and a drop under observation noise. On two real UR5 tasks, tube placing and nut pouring, both models succeed every time in distribution. Out of distribution, nut pouring rises from eighty-six point five to ninety-five point two percent.

The comparison that most directly tests the thesis is a two by two ablation: Mamba or a size-matched Transformer as the mixer, with gated or cross-attention conditioning. With gating, Mamba averages ninety-eight point four percent and the Transformer eighty-five point seven. With cross-attention, the gap is ninety-six point seven against seventy-five point six. At this tiny size, the mixer decides whether the policy still works. The idea also transfers. In SmolVLA, swapping in a Mamba expert cuts the expert by thirty-four point five percent and lifts average LIBERO success by three point six five points under a matched training budget.

The paper closes with a short limitations paragraph. Real-world testing covers two UR5 tasks, under mainly spatial distribution shifts. Composition has been shown only with backbone-depth reduction, not with visual-token pruning or denoising-step compression. Deeper analysis of the gating and of Mamba's state dynamics is left for later.

The evidence has further edges. Only the main LIBERO comparison uses three seeds. The ablations, the real-world runs and the other experiments use a single fine-tuning seed, and the ablation's ninety-eight point four is that seed's best run, above the three-seed mean. In the comparison with other efficient methods, SnapFlow, which compresses denoising to one step, reports both higher average success and a larger speedup. Finally, the ablation table shows something the text does not dwell on: at matched size, the Transformer experts have lower end-to-end latency than the Mamba ones. The speed comes from making the expert small. Mamba is what keeps a small expert accurate.

The idea to carry away is that in an iterative generator, the part that runs repeatedly deserves its own design budget. MambaAct leaves the expensive pretrained perception alone, compresses its output once, and spends little per denoising step. Mamba's role in that bargain is not raw speed. It is that an ordered, recurrent mixer stays accurate at a size where a Transformer of equal weight falls well behind.