pi0.5 행동 모듈을 10분의 1 크기 맘바로 바꾸자 응답 시간 53% 단축, 성공률은 소폭 상승

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 \pi_{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 \pi_{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.

I Introduction

Recent flow-matching VLAs such as \pi_{0} [1] and \pi_{0.5} [2] achieve strong manipulation performance, but their iterative action-generation process introduces substantial inference cost. At every policy query, the visual-language prefix can be computed once, whereas the action expert must be evaluated repeatedly across multiple flow-matching denoising steps. As a result, action generation can become the dominant source of latency even when the visual-language backbone itself is relatively compact.

Recent efforts to improve VLA efficiency have largely reduced this cost by decreasing the number of sampling steps, compressing or pruning pretrained networks, reducing visual-token processing, or redesigning the policy around smaller architectures. While effective, these approaches primarily accelerate inference by reducing how often existing components are evaluated or by simplifying the visual-language pathway and overall model. In contrast, the architecture of the repeatedly evaluated action expert itself remains comparatively underexplored as an efficiency axis. This motivates us to ask whether flow-matching inference can instead be accelerated by replacing the action-generation module while preserving the pretrained visual-language representations and standard flow-matching formulation.

We address this question with MambaAct, a compact Mamba [3] action expert for flow-matching VLA policies. MambaAct retains the pretrained \pi_{0.5} visual-language encoder and replaces its large Transformer-based action expert with a lightweight Mamba sequence mixer. Mamba is attractive in this setting because its selective state-space formulation provides a compact inductive bias for modeling ordered, strongly correlated action sequences. As we show empirically, this design enables aggressive action-expert compression while preserving policy performance. To efficiently condition the newly initialized expert on the pretrained VLM, we introduce a Context Compressor and gated FiLM [4] interface that transform the visual-language prefix into a compact representation reused throughout flow-matching denoising. The resulting expert contains only 29.7M parameters, compared with 311.5M for the original \pi_{0.5} action expert.

Our main contributions are:

• We introduce MambaAct, establishing action-expert replacement as a complementary efficiency axis for flow-matching VLAs. MambaAct replaces the repeatedly evaluated Transformer action expert with a compact Mamba expert while retaining the pretrained visual-language backbone and standard flow-matching formulation, without additional pretraining or distillation.
• We show that Mamba enables aggressive action-expert compression: MambaAct uses over 10\times fewer action-expert parameters than \pi_{0.5} while improving mean LIBERO [5] success by 1.08 percentage points across three fine-tuning seeds.
• Through controlled ablations, we examine the contributions of the Mamba action mixer and conditioning interface to the efficiency-performance trade-off. We further demonstrate that MambaAct composes with VLM-depth reduction and transfers to SmolVLA [6], indicating that action-expert replacement generalizes across flow-matching VLA backbones.

II Related Work

II-A Efficient Vision-Language-Action Models

The growing computational cost of VLA models has motivated a range of approaches that reduce inference cost at different stages of the policy. One direction focuses on accelerating iterative action generation. SnapFlow [7], for example, uses progressive self-distillation to compress iterative flow matching into a single action-expert evaluation, thereby reducing the number of sampling steps required at inference time. Complementary approaches instead reduce the cost of each network evaluation by simplifying the underlying architecture. Shallow-\pi [8] distills shallower visual-language and action transformers, while CLP [9] identifies and removes redundant layers from pretrained VLA models.

Another line of work targets the visual-language pathway, where long visual-token sequences can dominate computation. General VLM pruning methods such as FastV [10] and DivPrune [11] preserve early visual processing while removing redundant tokens before later transformer layers, and have subsequently been evaluated in VLA policies [9]. VLA-specific methods further adapt this idea to robotic control: LightVLA [12], ADP [13], and SpecPrune-VLA [14] selectively reduce visual tokens while retaining task-relevant information for action prediction. Beyond pruning individual components, several works pursue efficiency through compact architectures or coordinated reductions across multiple computational axes. SmolVLA [6] reduces the overall model scale, whereas CogVLA [15] combines instruction-conditioned visual aggregation, token sparsification, and a redesigned action decoder. Similarly, EfficientVLA [16] combines token and layer reduction with feature reuse across action-generation steps.

MambaAct targets a different design point: it retains the pretrained \pi_{0.5} backbone and standard flow-matching formulation, but replaces the transformer action expert with a compact Mamba-based expert trained jointly during task fine-tuning, without teacher models, distillation objectives, or additional pretraining.

II-B Mamba in Vision-Language-Action Models

Mamba [3] replaces pairwise self-attention with selective state-space recurrence, providing a compact mechanism for modeling ordered sequences. This makes it attractive for repeatedly evaluated action experts operating under tight parameter budgets. Mamba has been explored across sequence modeling and computer vision [17], with recent extensions to robotic perception and control [18, 19, 20]. Within VLA modeling, RoboMamba [18] uses Mamba primarily as the multimodal visual-language reasoning backbone, targeting efficient perception and reasoning rather than dense continuous action generation.

Mamba has also been explored for continuous control. Mamba Policy [19] applies Mamba to continuous-action diffusion, demonstrating selective state-space models as compact action-denoising networks. However, it is not designed as an action expert conditioned on representations from a pretrained vision-language backbone. More recently, SUREFlow [20] develops a lightweight Mamba-based multimodal policy with uncertainty-aware residual flow matching, jointly predicting action velocity and uncertainty to selectively refine unreliable actions. Unlike MambaAct, SUREFlow constructs a compact end-to-end multimodal policy rather than preserving a pretrained foundation VLM and replacing only its action expert.

MambaAct targets this complementary setting: accelerating the action-generation pathway of a pretrained flow-matching VLA while preserving its learned visual-language representations. It retains the \pi_{0.5} VLM and standard flow-matching objective, while replacing only the repeatedly evaluated Transformer action expert with a compact Mamba expert conditioned through compressed VLM context and gated FiLM modulation. We further show that this action-expert replacement transfers to SmolVLA.

III Methodology

Throughout this section, N denotes the number of visual-language prefix tokens, H the action horizon, d the action-expert hidden width, and d_{a} the padded action dimension. \mathbf{H}^{(l)}\in\mathbb{R}^{H\times d} denotes the action-token sequence after block l. Superscript (l) indexes action-expert blocks, whereas \ell denotes the language instruction. Batch dimensions are omitted for clarity.

III-A Preliminaries

Flow-matching VLAs and denoising latency bottleneck. The \pi family of models, including \pi_{0} and \pi_{0.5}, formulates continuous action generation through flow matching.

Let \mathbf{A}_{0}\in\mathbb{R}^{H\times d_{a}} be a clean padded action chunk. Training samples

\displaystyle u \\ \displaystyle\sim\operatorname{Beta}(1.5,1.0), \\ \displaystyle t \\ \displaystyle=0.001+0.999u, \\ \displaystyle\boldsymbol{\epsilon} \\ \displaystyle\sim\mathcal{N}(\mathbf{0},\mathbf{I}), \\ \displaystyle\mathbf{X}_{t} \\ \displaystyle=t\boldsymbol{\epsilon}+(1-t)\mathbf{A}_{0},

and the action expert predicts \hat{\mathbf{V}}_{\theta}(\mathbf{X}_{t},t,\mathbf{o}).

At inference, we initialize \mathbf{X}_{1}\sim\mathcal{N}(\mathbf{0},\mathbf{I}), where \mathbf{I} denotes the identity covariance matrix and \mathbf{X}_{t}\in\mathbb{R}^{H\times d_{a}} is the noisy action chunk at flow time t. We then integrate the learned flow backward from t=1 to t=0 using N_{\mathrm{denoise}} Euler denoising steps. With uniform step size \Delta t=1/N_{\mathrm{denoise}}, each step updates

\mathbf{X}_{t-\Delta t}=\mathbf{X}_{t}-\Delta t\,\hat{\mathbf{V}}_{\theta}(\mathbf{X}_{t},t,\mathbf{o}).

We use N_{\mathrm{denoise}}=10, a common choice in flow-matching VLA inference [1, 2, 6]. Although the visual-language prefix is computed once per policy query, the action expert is evaluated at every denoising step.

The action generation cost therefore scales approximately linearly with the number of denoising steps, making the repeatedly evaluated action expert a major inference bottleneck. Figure 1 confirms this: 10-step denoising accounts for 70.1\% of end-to-end latency in \pi_{0.5} and 86.4\% in SmolVLA. Thus, iterative action generation remains the dominant cost even with a much smaller visual-language backbone, motivating the quest for a more efficient expert whose per-step savings accumulate across denoising.

Mamba sequence mixers. We use the SISO Mamba3 formulation [3] along the action-token dimension; its complete selective scan, native gating, skip connection, and internal mixer projection are denoted by \operatorname{Mamba3}(\cdot).

III-B MambaAct

Figure 2 gives an overview of MambaAct. MambaAct retains and jointly fine-tunes the pretrained \pi_{0.5} visual-language backbone, while replacing its 311.5M-parameter Transformer action expert with a 29.7M-parameter Mamba expert. Given \mathbf{o}=\{\mathbf{I}^{w},\mathbf{I}^{e},\ell\}, where \mathbf{I}^{w} and \mathbf{I}^{e} denote the wrist and third-person images, respectively, and \ell is the language instruction, the backbone produces

\mathbf{P}=f_{\mathrm{VLM}}(\mathbf{o})\in\mathbb{R}^{N\times d_{\mathrm{vlm}}}.

Context compression

We split the VLM prefix into visual and language tokens, \mathbf{P}=[\mathbf{P}_{v};\mathbf{P}_{\ell}], where \mathbf{P}_{v} and \mathbf{P}_{\ell} denote the visual and language portions of the prefix, respectively. We then project all prefix tokens from the VLM hidden width to the action-expert width using

\widetilde{\mathbf{P}}=\mathbf{P}\mathbf{W}_{c}+\mathbf{b}_{c}\in\mathbb{R}^{N\times d}.

where \mathbf{W}_{c} and \mathbf{b}_{c} are the learned projection matrix and bias, respectively. Visual tokens are compressed into K_{v} contiguous segments:

\mathbf{C}_{v,i}=\frac{\sum_{n\in\mathcal{B}^{v}_{i}}m^{v}_{n}\widetilde{\mathbf{P}}_{v,n}}{\max(1,\sum_{n\in\mathcal{B}^{v}_{i}}m^{v}_{n})},\qquad i=1,\ldots,K_{v},

where \mathcal{B}^{v}_{i} denotes the set of visual-token indices assigned to the i-th contiguous segment, \widetilde{\mathbf{P}}_{v,n} is the projected representation of the n-th visual token, and m^{v}_{n}\in\{0,1\} is its validity mask. Language tokens are pooled analogously into \mathbf{C}_{\ell}. In addition, K_{g} learned queries \mathbf{Q}_{g}\in\mathbb{R}^{K_{g}\times d} extract global summaries, \mathbf{C}_{g}=\operatorname{MHA}(\mathbf{Q}_{g},\widetilde{\mathbf{P}},\widetilde{\mathbf{P}}). The complete context and condition vectors are

\begin{gathered}\mathbf{C}=[\mathbf{C}_{v}+\mathbf{e}_{v};\mathbf{C}_{\ell}+\mathbf{e}_{\ell};\mathbf{C}_{g}+\mathbf{e}_{g}],\\
\mathbf{C}\leftarrow\mathbf{C}+\operatorname{MLP}_{c}\left(\operatorname{RMSNorm}(\mathbf{C})\right),\\
\mathbf{z}=\frac{\sum_{k=1}^{K}m_{k}\mathbf{C}_{k}}{\max\left(1,\sum_{k=1}^{K}m_{k}\right)}\in\mathbb{R}^{d},\qquad K=K_{v}+K_{\ell}+K_{g}.\end{gathered}

Here, m_{k} is the context-validity mask, while \mathbf{e}_{v}, \mathbf{e}_{\ell}, and \mathbf{e}_{g} are modality embeddings for visual, language, and global context tokens, respectively.

Action and flow-time inputs

Executable actions are padded to d_{a}=32. The noisy chunk is projected into exactly H action tokens, with no additional robot-state token:

\mathbf{H}^{(0)}=\mathbf{X}_{t}\mathbf{W}_{\mathrm{in}}+\mathbf{b}_{\mathrm{in}}\in\mathbb{R}^{H\times d}.

Here, \mathbf{H}^{(0)} is the initial H-token action representation, \mathbf{W}_{\mathrm{in}} is the learned input projection, and \mathbf{b}_{\mathrm{in}} its bias.

Flow time is encoded separately as \mathbf{c}_{t}=\operatorname{MLP}_{t}(\phi(t)).

Gated Mamba action expert

The expert contains L=4 blocks. Each \mathbf{H}^{(l)}\in\mathbb{R}^{H\times d} denotes the action-token sequence after block l. For l=0,\ldots,L-1,

\displaystyle\mathbf{q}^{(l)}_{t} \\ \displaystyle=\mathbf{W}^{(l)}_{t}\operatorname{SiLU}(\mathbf{c}_{t})+\mathbf{b}^{(l)}_{t}, \\ \displaystyle\widetilde{\mathbf{H}}^{(l)} \\ \displaystyle=\mathbf{H}^{(l)}+\mathbf{1}_{H}(\mathbf{q}^{(l)}_{t})^{\top}, \\ \displaystyle(\boldsymbol{\gamma}^{(l)},\boldsymbol{\beta}^{(l)},\boldsymbol{\rho}^{(l)}) \\ \displaystyle=\operatorname{MLP}^{(l)}_{\mathrm{gate}}\!\left(\operatorname{RMSNorm}(\mathbf{z})\right), \\ \displaystyle\mathbf{U}^{(l)} \\ \displaystyle=\operatorname{Mamba3}^{(l)}\!\left(\operatorname{RMSNorm}(\widetilde{\mathbf{H}}^{(l)})\right), \\ \displaystyle\overline{\mathbf{U}}^{(l)} \\ \displaystyle=(1+\boldsymbol{\gamma}^{(l)})\odot\mathbf{U}^{(l)}+\boldsymbol{\beta}^{(l)}, \\ \displaystyle\mathbf{R}^{(l)} \\ \displaystyle=\widetilde{\mathbf{H}}^{(l)}+\sigma(\boldsymbol{\rho}^{(l)})\odot\overline{\mathbf{U}}^{(l)}, \\ \displaystyle\mathbf{H}^{(l+1)} \\ \displaystyle=\mathbf{R}^{(l)}+\operatorname{FFN}^{(l)}\!\left(\operatorname{RMSNorm}(\mathbf{R}^{(l)})\right).

Here, \mathbf{1}_{H}\in\mathbb{R}^{H} is an all-ones vector that broadcasts \mathbf{q}^{(l)}_{t} across the action horizon. \mathbf{U}^{(l)} denotes the Mamba output, \overline{\mathbf{U}}^{(l)} its FiLM-modulated representation, and \mathbf{R}^{(l)} the gated residual state. \boldsymbol{\gamma}^{(l)}, \boldsymbol{\beta}^{(l)}, and \boldsymbol{\rho}^{(l)} are the block-specific FiLM scale, shift, and residual-gate logits. We zero-initialize the time projections and FiLM scale/shift outputs, and initialize the gate logits to \log(0.2/0.8). This gives identity FiLM modulation and an initial residual multiplier of 0.2, attenuating the newly initialized Mamba branch while keeping it active. The value 0.2 is an initialization hyperparameter rather than a tuned optimum, and the gates are subsequently learned during fine-tuning. FiLM parameters are reused across denoising steps, while time conditioning and action tokens are recomputed at each step.

Velocity prediction and training

After the final block, the velocity head and target are

\displaystyle\hat{\mathbf{V}}_{\theta}(\mathbf{X}_{t},t,\mathbf{o}) \\ \displaystyle=\operatorname{RMSNorm}(\mathbf{H}^{(L)})\mathbf{W}_{\mathrm{out}}+\mathbf{b}_{\mathrm{out}}, \\ \displaystyle\mathbf{V}^{\star} \\ \displaystyle=\boldsymbol{\epsilon}-\mathbf{A}_{0},

where \mathbf{A}_{0}\in\mathbb{R}^{H\times d_{a}} denotes the clean ground-truth action chunk before noise is added, and \boldsymbol{\epsilon} is the sampled Gaussian noise used to construct \mathbf{X}_{t}. We minimize the mean flow-matching error

\mathcal{L}_{\mathrm{FM}}=\mathbb{E}\left[\frac{1}{Hd_{a}}\left\|\hat{\mathbf{V}}_{\theta}-\mathbf{V}^{\star}\right\|_{F}^{2}\right].

III-C Experimental Setup

We evaluate whether replacing the transformer-based action expert with a lightweight Mamba action expert can preserve or improve policy performance while reducing inference cost. We use LIBERO to measure simulated task performance and latency, LIBERO-Plus to assess zero-shot robustness, and two real-world UR5 tasks - tube placing and nut pouring - to evaluate physical manipulation performance and deployment efficiency.

Base policy and model configuration

We use OpenPI \pi_{0.5} [2] as both the base model and the controlled baseline. MambaAct retains the pretrained PaliGemma VLM backbone [21] including SigLIP vision encoder and Gemma visual-language prefix encoder, while replacing its transformer-based action expert with a lightweight Mamba action expert.

All \pi_{0.5}-based variants are initialized from the same base checkpoint and fine-tuned using a single-stage full-model recipe, in which the vision encoder, executed VLM layers, context-compression module, and action expert are optimized jointly. The MambaAct model executes all 18 VLM layers and uses a four-block Mamba action expert with width 512, state size 64, expansion factor 1, and head dimension 64. The Context Compressor produces K=64 context tokens, comprising 44 visual tokens, 16 language tokens, and 4 learned-query global summary tokens.

Training protocol

All \pi_{0.5}-based policies are trained for 30{,}000 optimizer steps with batch size 16 and gradient accumulation 1. For the main LIBERO comparison, we independently fine-tune MambaAct and the matched \pi_{0.5} baseline with training seeds \{17,27,42\}. All other \pi_{0.5}-based experiments use seed 42 unless otherwise stated. We fine-tune the full VLM backbone together with the full action expert, without freezing either component. We use AdamW with gradient clipping at 1.0, weight decay 10^{-10}, and a cosine learning-rate schedule from 5\times 10^{-5} to 5\times 10^{-6} without warmup. Training uses bfloat16 precision, gradient checkpointing, eight CPU workers, and a single NVIDIA H100 GPU.

Simulation benchmarks

We evaluate each policy on the standard LIBERO benchmark [5], including the Spatial, Object, Goal, and LIBERO-10 (later denoted as L-10) suites. Each suite contains ten tasks. We run 50 trials per task using evaluation seed 7 and report the success rate for each suite together with the unweighted average across all four suites. LIBERO inference uses image resolution 224\times 224, action horizon H=10, ten flow-matching denoising steps, and replanning every five environment steps. For the main LIBERO comparison, we report mean and standard deviation across these three fine-tuning seeds.

However, standard LIBERO performance may mask overfitting to benchmark-specific visual and environmental conditions, as policies can degrade under small perturbations. We therefore evaluate MambaAct and the matched \pi_{0.5} baseline on LIBERO-Plus [22], which contains 10{,}030 test-only task instances across seven perturbation categories: CAMERA, ROBOT, LANGUAGE, LIGHT, BACKGROUND, NOISE, and LAYOUT. Following the official protocol, we run one rollout per task instance using a fixed evaluation seed and report category-level and overall success rates.

Real-world benchmarks

Our setup uses two UR5 robots (Fig. 3): the left robot performs manipulation and carries a wrist-mounted RealSense camera, while the right robot remains fixed and carries a second RealSense camera providing the third-person view. We use third-person and wrist-mounted RGB observations at 30 fps and 640\times 480, resized to 224\times 224 with aspect-ratio-preserving padding. Object positions lie on a 10{\times}10 cm grid, with blue locations used for training and in-distribution (ID) evaluation and red locations reserved for out-of-distribution (OOD) testing. For each task, we collect 112 demonstrations from 28 training configurations, with four episodes per configuration, and evaluate on 28 ID and 52 OOD trials.

We consider two real-world tasks: Tube Placing and Nut Pouring. In the former, the robot transfers a tube from a source holder to a target holder; training uses source positions A1-A2, target positions B1-B7, and two tube slots, while OOD evaluation introduces unseen source positions A3-A4 and target positions B8-B10. In the latter, the robot first places the black container on the target, then grasps the blue container and pours its contents into the black container. Training uses blue-container positions A1-A2, black-container positions B1-B7, and targets C1-C2, while OOD evaluation uses unseen blue-container positions A3-A4 and black-container positions B8-B10.

For real-world evaluation, we use a normalized success rate with partial credit for meaningful intermediate completion. For tube placing, a trial receives 1 for successful placement, 0.5 when the tube is grasped but not inserted into the target holder, and 0 when grasping fails. For nut pouring, a trial receives 1 for successful pouring, 0.5 when the black container is successfully placed but pouring fails, and 0 when the preceding grasp-and-place stage fails.

Real-world inference

For physical deployment, the policies predict delta-joint action chunks with horizon H=50 from the visual-language observation. At each policy query, the robot executes the first 30 predicted actions at a control frequency of 30\,\mathrm{Hz} before replanning. Real-world inference runs on an NVIDIA RTX 5090 GPU.

Latency measurement

We report Prefix, Denoise, and E2E (end-to-end) latency. Prefix covers visual-language encoding and context construction, Denoise covers all ten flow-matching steps, and E2E includes the full policy query with I/O, CPU-GPU transfer, serialization, communication, and scheduling overhead. Simulation and real-world latency are measured on a single H100 and RTX 5090, respectively.

IV Results

IV-A Main Results

IV-A1 LIBERO

Table I highlights the performance and efficiency gains of MambaAct on LIBERO. Across three independent fine-tuning seeds, MambaAct improves average success from 96.72\pm 0.14\% to 97.80\pm 0.43\% (+1.08 percentage points), while reducing the action-expert core from 311.5M to 29.7M parameters (\downarrow 90.5\%). Replacing the original Gemma-style action pathway also removes its associated action-side parameter tensors, reducing the total model size from 3.62B to 2.96B parameters after introducing the Mamba expert and conditioning modules. On the seed-42 checkpoints used for latency measurement, the Context Compressor adds modest prefix overhead (40.76 to 46.61 ms), while denoising latency decreases by 64.8\%, yielding a 52.9\% reduction in E2E latency.

Table II compares MambaAct with recent efficient VLA methods on LIBERO. Given heterogeneous hardware and implementation settings, we report speedup relative to each method’s own baseline. MambaAct achieves 97.80\pm 0.43\% average success across three independent fine-tuning seeds, the second-highest average listed. Its best run (seed 42) reaches 99.2\%, 99.8\%, 98.8\%, and 95.8\% on Spatial, Object, Goal, and L-10, respectively, placing within the top two on every suite and averaging 98.4\%, close to SnapFlow’s reported 98.8\%. SnapFlow [7], however, evaluates 10 trials per task and does not report variability across fine-tuning seeds, whereas we evaluate three independently fine-tuned checkpoints with 50 trials per task each. Moreover, SnapFlow retains the full \pi_{0.5} architecture while compressing its denoising trajectory to one step; in contrast, MambaAct reduces the action-expert core from 311.5M to 29.7M parameters while retaining standard flow-matching inference. SnapFlow also uses progressive self-distillation, whereas MambaAct is trained within standard downstream fine-tuning without a separate distillation stage. Beyond simulation, we additionally validate MambaAct on two real-world UR5 manipulation tasks, and release our code and model weights to support reproducibility.

IV-A2 LIBERO-Plus

Table III shows that MambaAct achieves the highest overall score at 79.7\%, slightly outperforming the matched \pi_{0.5} baseline at 78.7\%. The largest gain is observed under variations in the robot’s initial pose, where MambaAct improves from 70.9\% to 77.6\%, while MambaAct also achieves the highest listed lighting robustness score at 95.2\% and matches the strongest layout result at 86.4\%. The main degradation occurs under observation noise, where performance drops from 77.4\% to 73.5\%, while camera robustness is essentially unchanged. Overall, the results indicate that the strong standard LIBERO performance of MambaAct does not come at the cost of substantially reduced robustness to perturbations.

IV-A3 Real-world benchmarks

Results are shown in Table IV. Both models achieve 100\% ID success, while the main difference appears under OOD evaluation. Compared with the baseline \pi_{0.5}, MambaAct improves OOD success from 86.5\% to 95.2\% on nut pouring and from 76.9\% to 82.7\% on tube placing, while also reducing model size, GFLOPs, and training time. Most notably, denoising latency decreases by 58.8\%, yielding roughly 33\% lower end-to-end latency on both tasks. Overall, MambaAct improves real-world OOD robustness while substantially accelerating action generation.

IV-B Ablation Studies

We conduct three ablation studies to assess the source of gain, composability, and transferability of MambaAct. Unless otherwise stated, all ablations use a fine-tuning seed of 42 and an evaluation seed of 7.

Ablation 1: Architectural source of gain. We isolate the contributions of the action mixer and conditioning interface using a 2\times 2 comparison of Mamba and Transformer mixers with gated and cross-attention conditioning (Table V).

Under gated conditioning, Mamba achieves 98.4\% average success, substantially outperforming the size-matched Transformer at 85.7\%. The same trend holds with cross-attention (96.7\% vs. 75.6\%), suggesting that the Mamba action mixer is important for preserving performance under aggressive action-expert compression. We hypothesize that selective state-space mixing provides a useful inductive bias for flow-matching action generation: Mamba propagates information along the ordered, strongly correlated action-token sequence through a selective recurrent state rather than unrestricted pairwise token interactions. This may allow a compact expert to model action-token dynamics more effectively, consistent with prior results on Mamba-based diffusion policies [19].

Gated conditioning also consistently outperforms cross-attention: with Mamba, it improves average success from 96.7\% to 98.4\% while reducing latency from 119.66 to 107.61 ms; with Transformer, it improves success from 75.6\% to 85.7\%. These results suggest that repeated token-level attention to the compressed context is not necessary for strong action generation in the evaluated setting. However, the gated models use larger action experts (30M vs. 17-18M parameters), so the improvement cannot be attributed to conditioning alone.

One possible explanation is the optimization setting of MambaAct. The action expert and conditioning interface are initialized from scratch, while the VLM starts from pretrained \pi_{0.5} weights and is jointly fine-tuned. We hypothesize that gated FiLM provides a constrained interface for aligning these components: feature-wise scale and shift adapt the action representation to the visual-language condition, while the residual gate controls the magnitude of the conditioned update. This interpretation is consistent with prior findings on FiLM [4] and adaptive modulation in diffusion transformers [24]. Overall, the ablation suggests that Mamba provides a stronger performance-capacity trade-off under aggressive action-expert compression. Gated variants further improve the observed efficiency-performance trade-off, although the accuracy contribution of the conditioning interface remains partially confounded by expert size.

Ablation 2: Composability. Here, we study whether MambaAct can compose with complementary prefix-side VLA acceleration. Efficient VLA systems can be accelerated along multiple axes, including reducing denoising steps, pruning or compressing visual tokens, truncating VLM depth, and replacing the action expert. Because MambaAct primarily reduces computation in the repeatedly evaluated action pathway, its acceleration mechanism is distinct from methods that reduce computation in the visual-language prefix. We test this composability using VLM-depth reduction as a representative prefix-side acceleration. As a controlled efficiency–accuracy trade-off, we combine MambaAct with VLM-depth reduction by executing only the first 12 VLM layers instead of all 18 layers before passing the resulting context representation to the Mamba action expert.

Tables VI and VII show that MambaAct remains effective when combined with prefix-side acceleration. On LIBERO, the L12 variant reduces training time from 13.78 to 10.81 hours and lowers E2E latency from 107.6 to 102.91 ms, while maintaining 97.5\% average success compared with 98.4\% for the full model. This suggests that the Mamba action expert can operate effectively on intermediate VLM representations, rather than requiring the full-depth visual-language prefix.

The same overall trend appears in real-world evaluation. MambaAct L12 reduces the deployed model size from 2.96B to 2.30B parameters and lowers RTX5090 inference latency from 83 ms to 74 ms. Despite this additional compression, it preserves 100\% ID success on both tasks and maintains comparable OOD performance, reaching 95.0\% final success on nut pouring and 90.0\% on tube placing.

Ablation 3: Transferability. We test whether the Mamba action expert transfers beyond \pi_{0.5} by replacing SmolVLA’s Transformer action expert with a 4-block Mamba expert. This reduces the action-expert size from 98.25M to 64.31M parameters, a 34.5\% reduction.

Table VIII shows that MambaAct transfers effectively to SmolVLA. The published SmolVLA score of 87.3\% in Table II uses the paper’s 100k-step training setting, whereas the released LIBERO checkpoint was trained for 25k steps with batch size 32. For a fair comparison, we therefore re-evaluate the released checkpoint under our protocol of 50 trials per task with evaluation seed 7, obtaining 81.30\% average success. Under the same 25k-step, batch-32, seed-1000 training budget, SmolVLA-MambaAct reaches 84.95\% (+3.65 percentage points), while reducing denoising latency by 63.3\% and E2E latency by 52.0\%. The largest gains occur on LIBERO-10 (70.2\%\rightarrow 77.8\%) and SPATIAL (76.6\%\rightarrow 82.6\%).

These results indicate that the Mamba action-expert replacement transfers beyond the \pi-family, reducing action-expert size while improving average task success on a second flow-matching VLA backbone.

In summary, the ablations support three observations. First, Mamba substantially outperforms size-matched Transformer mixers under both evaluated conditioning interfaces, suggesting that the action mixer itself is important for preserving performance under aggressive expert compression. Second, gated conditioning provides an efficient interface between the pretrained VLM and the newly initialized action expert, although its accuracy advantage over cross-attention is partially confounded by model capacity. Third, the resulting Mamba action expert remains effective under VLM-depth reduction and transfers to a second flow-matching VLA architecture. Together, these results are consistent with benefits from selective action sequence modeling and lightweight visual-language conditioning beyond compression alone.

V Conclusion

We present and publicly release MambaAct, a lightweight Mamba-based action expert tailored for flow-matching vision-language-action models. Rather than altering the pretrained vision-language backbone or truncating denoising steps, MambaAct directly optimizes the frequently evaluated action-generation pathway. By substituting the Transformer action expert in \pi_{0.5} with a Mamba mixer, MambaAct enables aggressive compression of the repeatedly evaluated action pathway, reducing the core action-expert parameter count from 311.5M to 29.7M (\sim 10\times) while substantially lowering denoising latency. Crucially, this compression improves mean LIBERO performance across three independent fine-tuning seeds, while retaining strong robustness on LIBERO-Plus and on real-world robotic tasks. Furthermore, MambaAct seamlessly integrates with VLM-depth reduction and generalizes effectively when transferred to SmolVLA.

Despite these strengths, several limitations remain. First, our real-world experiments are currently restricted to two UR5 tasks under primary spatial distribution shifts. In addition, while MambaAct composes with VLM-depth reduction, its compatibility with other efficiency techniques, such as visual-token pruning and denoising-step compression, remains to be evaluated. We also leave deeper analysis of the gated visual-language conditioning and selective Mamba state dynamics to future work.

References

1. [1] K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. (2024) \pi_{0}: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164.
2. [2] K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, M. Y. Galliker, et al. (2025) \pi_{0.5}: A vision-language-action model with open-world generalization. In 9th Annual Conference on Robot Learning,
3. [3] A. Lahoti, K. Li, B. Chen, C. Wang, A. Bick, J. Z. Kolter, T. Dao, and A. Gu (2026) Mamba-3: improved sequence modeling using state space principles. In The Fourteenth Int. Conf. on Learning Representations,
4. [4] E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville (2018) Film: visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32.
5. [5] B. Liu, Y. Zhu, C. Gao, Y. Feng, Q. Liu, Y. Zhu, and P. Stone (2023) Libero: benchmarking knowledge transfer for lifelong robot learning. Adv. Neural Inf. Process. Syst. 36, pp. 44776–44791.
6. [6] M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. (2025) Smolvla: a vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844.
7. [7] W. Luan, J. Li, W. Zhao, W. Zhang, T. Wu, and R. Ma (2026) Snapflow: one-step action generation for flow-matching vlas via progressive self-distillation. arXiv preprint arXiv:2604.05656.
8. [8] B. Jeon, Y. Choi, and T. Kim (2026) Shallow-\pi: knowledge distillation for flow-based vlas. arXiv preprint arXiv:2601.20262.
9. [9] G. Nguyen, T. Ho, T. Ha, K. Vo, P. L. Møller, Q. T. Nguyen, L. Dinh, T. Dam, V. Duong, T. M. Luu, et al. (2026) Finetuning vision-language-action models requires fewer layers than you think. arXiv preprint arXiv:2606.20246.
10. [10] L. Chen, H. Zhao, T. Liu, S. Bai, J. Lin, C. Zhou, and B. Chang (2024) An image is worth 1/2 tokens after layer 2: plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pp. 19–35.
11. [11] S. R. Alvar, G. Singh, M. Akbari, and Y. Zhang (2025) Divprune: diversity-based visual token pruning for large multimodal models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 9392–9401.
12. [12] T. Jiang, X. Jiang, Y. Ma, X. Wen, B. Li, K. Zhan, P. Jia, Y. Liu, S. Sun, and X. Lang (2025) The better you learn, the smarter you prune: towards efficient vision-language-action models via differentiable token pruning. arXiv preprint arXiv:2509.12594.
13. [13] X. Pei, Y. Chen, S. Xu, Y. Wang, Y. Shi, and C. Xu (2025) Action-aware dynamic pruning for efficient vision-language-action manipulation. arXiv preprint arXiv:2509.22093.
14. [14] H. Wang, J. Xu, Y. Xiang, J. Pan, Y. Zhou, Y. Li, and G. Dai (2025) Specprune-vla: accelerating vision-language-action models via action-aware self-speculative pruning. arXiv preprint arXiv:2509.05614.
15. [15] W. Li, R. Zhang, R. Shao, J. He, and L. Nie (2026) CogVLA: cognition-aligned vision-language-action models via instruction-driven routing & sparsification. Adv. Neural Inf. Process. Syst. 38, pp. 137646–137675.
16. [16] Y. Yang, Y. Wang, Z. Wen, L. Zhongwei, C. Zou, Z. Zhang, C. Wen, and L. Zhang (2026) Efficientvla: training-free acceleration and compression for vision-language-action models. Adv. Neural Inf. Process. Syst. 38, pp. 40891–40914.
17. [17] E. Fazzari, D. Romano, F. Falchi, and C. Stefanini (2025) Selective state models are what you need for animal action recognition. Ecological Informatics 85, pp. 102955. External Links: ISSN 1574-9541, Document
18. [18] J. Liu, M. Liu, Z. Wang, P. An, X. Li, K. Zhou, S. Yang, R. Zhang, Y. Guo, and S. Zhang (2024) Robomamba: efficient vision-language-action model for robotic reasoning and manipulation. Adv. Neural Inf. Process. Syst. 37, pp. 40085–40110.
19. [19] J. Cao, Q. Zhang, J. Sun, J. Wang, H. Cheng, Y. Li, J. Ma, K. Wu, Z. Xu, Y. Shao, et al. (2025) Mamba policy: towards efficient 3d diffusion policy with hybrid selective state models. In 2025 IEEE/RSJ Int. Conf. on Intel. Robots and Syst. (IROS), pp. 11359–11366.
20. [20] M. T. Islam, S. N. Peddapalli, S. Lee, and S. Ahn (2026) SUREFlow: state-space uncertainty-aware residual flow matching for robust robot manipulation. arXiv preprint arXiv:2607.10504.
21. [21] L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, D. Salz, M. Neumann, I. Alabdulmohsin, M. Tschannen, E. Bugliarello, et al. (2024) Paligemma: a versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726.
22. [22] S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, et al. (2025) Libero-plus: in-depth robustness analysis of vision-language-action models. arXiv preprint arXiv:2510.13626.
23. [23] M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. (2024) Openvla: an open-source vision-language-action model. arXiv preprint arXiv:2406.09246.
24. [24] W. Peebles and S. Xie (2023) Scalable diffusion models with transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4172–4182.