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

Khuyen Pham; Tran Nguyen Le; Edoardo Fazzari; Cesare Stefanini

cs.ROOct 16, 2026

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

Authors: Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari, Cesare Stefanini

Organizations: MBZUAI · Technical University of Denmark

Abstract

Vision-language-action (VLA) models have shown strong promise for generalist robotic manipulation, but their inference cost remains a major bottleneck for real-time control. This limitation is especially pronounced in flow-matching VLAs such as $π_{0.5}$ , where the action expert is evaluated repeatedly during iterative denoising. We introduce MambaAct, an efficient Mamba action expert for flow-matching VLA policies. MambaAct retains the pretrained visual-language prefix encoder of $π_{0.5}$ , while replacing its transformer-based action expert with a compact Mamba-based action expert. Within this expert, Mamba serves as the action mixer that models dependencies across the noisy continuous-action sequence. To preserve visual-language conditioning without repeated token-level cross-attention, MambaAct uses a lightweight Context Compressor to map Vision-language model (VLM) prefix tokens into a compact conditioning signal that gates and modulates each Mamba action block. MambaAct uses a $10\times$ smaller action expert and reduces flow-matching denoising latency by $64.8\%$ on LIBERO and $58.8\%$ in real-world deployment, while improving mean LIBERO success by $1.08$ percentage points across three independent fine-tuning seeds and maintaining gains on LIBERO-Plus and real-world benchmarks. Our code and model weights are publicly available at https://github.com/giakhuyendihoc/MambaAct.

Figures & tables

Fig. 1: Latency decomposition of

\pi_{0.5}

and SmolVLA on LIBERO benchmark highlighting action generation as the dominant source of latency.

Fig. 2: MambaAct architecture. MambaAct retains and jointly fine-tunes the pretrained $\pi_{0.5}$ visual-language backbone while replacing its Transformer action expert with a compact Mamba-3 expert. The VLM prefix is compressed into visual, language, and learned-query global context tokens $\mathbf{C}$ and pooled into $\mathbf{z}$ , which block-specific FiLM MLPs map to scale, shift, and residual-gate parameters. The noisy action chunk $\mathbf{X}_{t}$ is projected into the initial action-token sequence $\mathbf{H}^{(0)}$ , while the embedded flow time $\mathbf{c}_{t}$ is injected through block-specific projections. The $L=4$ gated Mamba-3 blocks transform the full action-token sequence into $\mathbf{H}^{(L)}$ , and a final RMSNorm-linear head predicts the flow velocity $\hat{\mathbf{V}}_{\theta}$ . Repeated Euler updates use this velocity to obtain the denoised action chunk $\mathbf{X}_{0}$ .

Fig. 3: Real-world benchmark setups: (left) tube placing and (right) nut pouring.

LIBERO success rate (%)
Model	Spatial	Object	Goal	L-10	Avg.
$\pi_{0.5}$	$98.00\pm 0.43$	$98.33\pm 0.66$	$97.73\pm 0.84$	$92.80\pm 0.43$	$96.72\pm 0.14$
MambaAct	$\mathbf{98.73\pm 0.34}$	$\mathbf{99.60\pm 0.43}$	$\mathbf{98.60\pm 0.28}$	$\mathbf{94.27\pm 1.52}$	$\mathbf{97.80\pm 0.43}$

TABLE I: LIBERO performance and efficiency results. Success rates are mean

\pm

standard deviation over three independent fine-tuning seeds

\{17,27,42\}

, with

50

trials per task using evaluation seed

7

. Latency is measured on the seed-

42

checkpoints using a single H100.

Controlled baseline
Method	Baseline	Spatial	Object	Goal	L-10	Avg.	Speedup
$\pi_{0.5}$ [ 2 ]	–	$98.0$	$98.3$	$97.7$	$92.8$	$96.7$	$1.00\times$
Action-expert replacement & conditioning
MambaAct (Ours)	$\pi_{0.5}$	$\underline{98.7}$	$\underline{99.6}$	$\underline{98.6}$	$94.3$	$\underline{97.8}$	$\underline{2.12\times}$
Denoising-step compression
SnapFlow [ 7 ]	$\pi_{0.5}$	$\mathbf{99.0}$	$\mathbf{100.0}$	$\mathbf{99.0}$	$\mathbf{97.0}$	$\mathbf{98.8}$	$\mathbf{3.30\times}$

TABLE II: LIBERO comparison with efficient VLA methods. Speedup is measured relative to the listed baseline. Bold and underlined values denote the highest and second-highest values listed, respectively. For

\pi_{0.5}

and MambaAct, success rates are means over three independent fine-tuning seeds

\{17,27,42\}

; other results are from the cited works.

Method	Camera	Robot	Language	Light	Backgr.	Noise	Layout	Total
OpenVLA	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
OpenVLA-OFT	56.4	31.9	79.5	88.7	93.3	75.8	74.2	69.6
OpenVLA-OFT w	10.4	38.7	70.5	76.8	93.6	49.9	69.9	55.8
OpenVLA-OFT m	55.6	21.7	81.0	92.7	91.0	78.6	68.7	67.9
NORA	2.2	37.0	65.1	45.7	58.6	12.8	62.1	39.0
WorldVLA	0.1	27.9	41.6	43.7	17.1	10.9	38.0	25.0

TABLE III: Zero-shot robustness on LIBERO-Plus. All models are trained on standard LIBERO and evaluated without LIBERO-Plus fine-tuning. Published results are from prior work [ 22 ] ;

\pi_{0.5}

and MambaAct use our matched training and evaluation protocol.

Task	Model	Model Size (B)	Train (h)	GFLOPs
Nut pouring	Baseline $\pi_{0.5}$	$3.62$	$14.3$	$5001.2$
Nut pouring	MambaAct	$2.96$ $\downarrow 18.2\%$	$14.0$	$4488.8$ $\downarrow 10.2\%$
Tube placing	Baseline $\pi_{0.5}$	$3.62$	$14.3$	$5001.2$
Tube placing	MambaAct	$2.96$ $\downarrow 18.2\%$	$14.0$	$4488.8$ $\downarrow 10.2\%$

TABLE IV: Real-world efficiency and success rates for nut pouring and tube placing, with inference run on an RTX5090. E2E latency is the mean end-to-end policy-query latency and includes prefix encoding, action denoising, input/output transformations, CPU-GPU transfers, serialization, WebSocket communication, and runtime scheduling.

Mixer	Conditioning	Spatial	Object	Goal	L-10	Avg	Size (B)	Train (h)	Latency (ms)
Mamba	Gate	99.2	99.8	98.8	95.8	98.4	0.030	13.78	107.61
Mamba	Cross-Attn	98.6	99.4	94.6	94.2	96.7	0.017	13.83	119.66
Transformer	Gate	82.8	90.6	86.8	82.6	85.7	0.030	13.69	74.22
Transformer	Cross-Attn	75.2	82.8	75.8	68.6	75.6	0.018	13.72	85.60

TABLE V: Action-mixer and conditioning-interface ablation on LIBERO. All models use the same Context Compressor. Size reports action-expert core size.

Model	Spatial	Object	Goal	L-10	Avg.	Train (h)	Prefix	Denoise	E2E
							Latency (ms)
MambaAct	$99.2$	$99.8$	$98.8$	$95.8$	$98.4$	$13.78$	$46.61$	$56.40$	$107.6$
MambaAct L12	$98.0$	$99.8$	$98.0$	$94.0$	$97.5$	$10.81$	$41.15$	$56.06$	$102.91$

TABLE VI: MambaAct composed with VLM-depth reduction on LIBERO. Both variants use fine-tuning seed

42

and are evaluated with

50

trials per task using evaluation seed

7

. Latency is measured on a single H100.

Model efficiency and latency
Model	Params (B)	Train (h)	GFLOPs/ query	Latency (ms)
				Prefix	Denoise	E2E
MambaAct	$2.96$	$14.00$	$4488.84$	$41.15$	$31.11$	$83.26$
MambaAct L12	$2.30$	$13.55$	$3209.70$	$32.05$	$31.03$	$74.19$

TABLE VII: MambaAct composed with VLM-depth reduction on real-world tasks. Latency is measured on an RTX 5090 for the tube task; nut pouring shows similar latency.

	LIBERO success rate (%)					Latency (ms)			Params
Model	Spatial	Object	Goal	L-10	Avg.	Prefix	Denoise	E2E	Action expert (M)
Released SmolVLA	$76.6$	$93.0$	$85.4$	$70.2$	$81.30$	$25.53$	$163.15$	$188.69$	$98.25$
SmolVLA-MambaAct	$82.6$	$92.4$	$87.0$	$77.8$	$\mathbf{84.95}$	$30.69$	$59.89$	$\mathbf{90.61}$	$64.31$

TABLE VIII: Mamba action-expert transfer to SmolVLA on LIBERO. Both models are evaluated with

50

trials per task and evaluation seed

7

(

2{,}000

episodes total), while retaining SmolVLA’s native action horizon

H=50

, execution horizon

10

, and

10

flow-matching denoising steps. Latency is measured on a single H100.

Efficient VLA Model Inference

Efficient vision-language-action (VLA) model inference reduces robot-control latency, memory use, and energy while preserving task performance.

See all 175 papers

Mamba-based Selective State Space Modeling Improves the Accuracy-Complexity Tradeoff of SmolVLA Vision-Language-Action Experts
Aug 10, 2026Farida Mohsen, Thowayba Elkaffash, Mohammad Reza Chalak Qazani +3Vision-Language-Action Models Efficient VLA Model Inference
College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar · College of Engineering, Qatar University, Doha, Qatar · College of Science and Engineering, James Cook University, Townsville, QLD, 4814, Australia
LaMP: Learning Vision-Language-Action Policy with 3D Scene Flow as Latent Motion Prior
Mar 26, 2026Xinkai Wang, Chenyi Wang, Yifu Xu +7Language-Conditioned Robot Manipulation Vision-Language-Action Models
Southeast University, China · Shanghai Innovation Institute, China · Shanghai Jiao Tong University, China +1
PAMAE: Phase-Aware-MoE Action Experts Towards Reliable Flow-Matching Vision-Language-Action Policies
Jun 25, 2026Jiayu Yang, Tao Yang, Xiang Chang +3Flow Matching Language-Conditioned Robot Manipulation
Department of Artificial Intelligence, School of Informatics, Xiamen University, Xiamen, China · Department of Computer Science, Aberystwyth University, Aberystwyth, U.K.

Aug 10, 2026J/K move · Enter open · S save

Mamba-based Selective State Space Modeling Improves the Accuracy-Complexity Tradeoff of SmolVLA Vision-Language-Action Experts

Farida Mohsen, Thowayba Elkaffash, Mohammad Reza Chalak Qazani, Mohamed Mabrok+2

College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar · College of Engineering, Qatar University, Doha, Qatar · College of Science and Engineering, James Cook University, Townsville, QLD, 4814, Australia

Vision-Language-Action Models Efficient VLA Model Inference Efficient VLA Models Selective SSMs

Open paper AskPDF arXiv

Vision-language-action (VLA) models face a crucial tradeoff between their task success rate and the policy-call frequency. Executing a single action per inference (

N=1

) enables accurate robot control but comes at the cost of huge compute time overheads, making real-time implementation infeasible. On the other hand, executing longer action horizons before replanning (

N\gg1

) reduces compute complexity, but inevitably degrades the system's success rate. In order to improve the VLA accuracy-complexity tradeoff, this paper investigates Mamba's selective state-space modeling as an alternative to causal self-attention within the action expert of the popular SmolVLA model, widely used as a reference model for its highly accurate yet low complexity nature. We evaluate both the Mamba- and Transformer-based experts on the widely-adopted LIBERO benchmark suites across three execution horizons

N\!\in\!\{1,25,50\}

, respectively corresponding to high, moderate and low compute complexities. Our results remarkably show that the advantage of the Mamba expert increases with the execution horizon, indicating significant success retention under long execution horizons

N = 50

and

N = 25

. When

N = 50

actions are executed before replanning (i.e., corresponding to feasible real-time deployment), the Mamba expert outperforms the Transformer baseline by

7.8\%

. In addition, when

N = 25

actions are executed before replanning, our Mamba expert outperforms the Transformer baseline by

3.7\%

. Finally, under per-action replanning (

N=1

), our Mamba variant matches the Transformer-based mean success rate while significantly reducing the overall model parameter complexity by

24\%

thanks to Mamba's compute-efficient nature.

Abstract

Efficient VLA Model Inference

Explore similar work

Mamba-based Selective State Space Modeling Improves the Accuracy-Complexity Tradeoff of SmolVLA Vision-Language-Action Experts

LaMP: Learning Vision-Language-Action Policy with 3D Scene Flow as Latent Motion Prior

PAMAE: Phase-Aware-MoE Action Experts Towards Reliable Flow-Matching Vision-Language-Action Policies