MixVLA: Adaptive Mixing of Non-Invariant Information for Generalizable Vision-Language-Action Models

Pingrui Zhang; Yu Zhang; Pengyuan Wu; Bin Wang; Haoming Song; Xianqiang Gao; ZhaxiZhuoma; Zhigang Wang

cs.ROOct 2, 2026

MixVLA: Adaptive Mixing of Non-Invariant Information for Generalizable Vision-Language-Action Models

Authors: Pingrui Zhang, Yu Zhang, Pengyuan Wu, Bin Wang, Haoming Song, Xianqiang Gao, ZhaxiZhuoma, Zhigang Wang,

Organizations: Fudan University · Shanghai AI Laboratory · Harbin Institute of Technology, Shenzhen, China · Northwestern Polytechnical University ·

Abstract

Vision-Language-Action (VLA) models have achieved remarkable advances in robotic manipulation, yet their zero-shot generalization under out-of-distribution (OOD) conditions remains limited. These models often entangle task-relevant invariant structure with environment-specific non-invariant factors, causing policies to rely on spurious appearance cues during action prediction. In this work, we propose MixVLA, a model-agnostic training framework that improves the generalization of VLA models without requiring additional OOD data or architectural modifications. The key component of MixVLA is Adaptive Mixing of Non-Invariant Information (AMI). AMI stochastically mixes non-invariant representations to regularize distribution-specific variability while preserving complementary predictive cues. The mixed non-invariant features are then fused with invariant representations for final action prediction, resulting in improved robustness without sacrificing policy expressiveness. Extensive experiments across challenging manipulation settings, including LIBERO, LIBERO-Plus, the RoboTwin perturbation suite, and real-world tasks, demonstrate that MixVLA improves overall zero-shot robustness while retaining strong in-domain performance.

Figures & tables

Figure 1: Left: Data scaling improves robustness by collecting massive and diverse demonstrations, but incurs substantial labor and financial costs. Middle: Some works improve generalization with task-specific modules (e.g., camera rotation or geometric awareness), which limits general applicability across VLA architectures. Right: MixVLA is a model-agnostic training framework that explicitly regularizes and mixes non-invariant signals in the latent representation space, enabling robust zero-shot action prediction without additional data or architectural modifications.

Figure 2: Overview of MixVLA. Left (Decomposition): A standard VLA policy is trained with the base objective to obtain $W_{\text{Normal}}$ , while an Invariant VLA is trained with an information bottleneck objective to obtain $W_{\text{Inv}}$ . The Non-Invariant VLA is then derived via weight arithmetic. Right (Regularization): Non-invariant embeddings are randomly mixed within the mini-batch using AMI . The mixed representation $\tilde{NIe}$ is concatenated with the invariant embedding $Ie$ and passed to the policy head for training.

Figure 3: Experimental Setup. Our experiments cover zero-shot generalization (Q1), representation analysis (Q2), cross-policy and real-world applicability (Q3), and training efficiency with Self-MixVLA (Q4).

Method	Camera	Robot	Language	Light	Background	Noise	Layout	Total
OpenVLA ( Kim et al., 2024 )	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
WorldVLA ( Cen et al., 2025 )	0.1	27.9	41.6	43.7	17.1	10.9	38.0	25.0
NORA ( Hung et al., 2025 )	2.2	37.0	65.1	45.7	58.6	12.8	62.1	39.0
UniVLA ( Bu et al., 2025 )	1.8	46.2	69.6	69.0	81.0	21.2	31.9	42.9
$\pi_{0}$ ( Black et al., 2024 )	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6
$\pi_{0}$ -Fast ( Pertsch et al., 2025 )	65.1	21.6	61.0	73.2	73.2	74.4	68.8	61.6

Table 1: Performance on the LIBERO-Plus benchmark (out-of-distribution evaluation).

Model	Spatial	Object	Goal	Long	Avg
OpenVLA ( Kim et al., 2024 )	84.7	88.4	79.2	53.7	76.5
WorldVLA ( Cen et al., 2025 )	85.6	89.0	82.6	59.0	79.1
NORA ( Hung et al., 2025 )	92.2	95.4	89.4	74.6	87.9
UniVLA ( Bu et al., 2025 )	96.5	96.8	95.6	92.0	95.2
$\pi_{0}$ ( Black et al., 2024 )	90.0	86.0	95.0	73.0	86.0
$\pi_{0}$ -Fast ( Pertsch et al., 2025 )	96.4	96.8	88.6	60.2	85.5

Table 2: Comparison on the LIBERO benchmark (in-domain setting).

Method	C2C	C2R	Avg.
Fast-WAM ( Yuan et al., 2026 )	77.8	1.9	39.9
X-VLA ( Zheng et al., 2026 )	68.0	20.9	44.5
X-WAM ( Guo et al., 2026 )	70.0	25.8	47.9
Spatial Forcing ( Li et al., 2026 )	77.2	26.7	52.0
$\pi_{0.5}$ ( Intelligence et al., 2025 )	70.7	46.0	58.4
4D-WAM ( Yang et al., 2026 )	81.5	41.8	61.6

Table 3: Clean-to-clean and clean-to-randomized evaluation on RoboTwin 2.0.

IB	AMI	LIBERO-Plus								RoboTwin
IB	AMI	Camera	Robot	Language	Lighting	Background	Noise	Layout	Total	C2C	C2R	Avg.
		51.2	30.7	74.2	87.8	92.6	76.0	73.7	67.5	73.8	53.8	63.8
✓		48.8	58.6	72.7	91.3	93.0	78.8	74.1	72.2	75.8	59.6	67.7
✓	✓	49.6	63.3	82.6	96.1	93.9	82.3	76.8	76.2	75.6	62.4	69.0

Table 4: Ablation study of IB and AMI on LIBERO-Plus and RoboTwin 2.0. Experiments use OpenVLA-OFT on LIBERO-Plus and

\pi_{0.5}

on RoboTwin 2.0.

Panel of Figure 4 — Figure 4: Qualitative analysis under distribution shifts. (a) Grad-CAM visualizations along each policy’s own rollout. (b) t-SNE projections of mean-pooled token features from eight tasks. Colors indicate tasks; circles and triangles denote clean and randomized observations, respectively.

Figure 5: Real-world robustness to lighting changes.

Method	C2C	C2R	Avg.
Self-MixVLA	73.0	60.8	66.9
MixVLA	75.6	62.4	69.0

Table 5: Single-run variant on RoboTwin 2.0.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Ablation of mixing probability.

Figure 7: Visualization of manipulation trajectories on the LIBERO benchmark. Each row presents a sequence of key frames from a representative task execution, illustrating the manipulation progress from the initial state to task completion. Different rows correspond to different manipulation tasks and scene configurations, demonstrating the policy behavior under the standard in-domain setting.

	Hard	Soft	Camera	Robot	Language	Light	Background	Noise	Layout	Total
✓		49.0	61.7	82.3	95.5	93.9	82.1	76.8	75.7
		✓	49.6	63.3	82.6	96.1	93.2	82.3	76.1	76.0

Appendix

Table 6: Comparison of hard mixing and soft mixing on LIBERO-Plus. We compare two variants of Adaptive Mixing of Non-Invariant Information (AMI): hard mixing , which performs direct interpolation between non-invariant embeddings, and soft mixing , which mixes feature statistics through adaptive normalization. Results are reported as task success rates (%) under different perturbation types.

Figure 8: Three-camera observations for real-world manipulation. Representative views from one wrist-mounted camera and two side-view cameras used in our Franka experiments.

Figure 9: Visualization of manipulation trajectories under LIBERO-Plus perturbations. Each row presents a sequence of key frames from a representative task execution in the LIBERO-Plus benchmark. The scenes contain various observation perturbations such as changes in background textures, lighting conditions, and visual noise. The visualization illustrates how the policy performs manipulation under out-of-distribution environments from the initial state to task completion.

Vision-Language-Action Models

Vision-language-action models turn visual observations and natural-language instructions into actions for robots and other embodied agents.

See all 1,190 papers

AC-VLA: Robust Out-of-Distribution Action Execution via Compositional Learning
Jul 17, 2026Xiaojiang Peng, Kai Peng, Jie Lu +3Robot Skill Learning Generalization in Robotic Manipulation
Shenzhen Technology University · Tongji University · Great Bay University +1
$M^2$ -VLA: Boosting Vision-Language Models for Generalizable Manipulation via Layer Mixture and Meta-Skills
Apr 27, 2026Siyao Xiao, Yuhong Zhang, Zhifang Liu +9Language-Conditioned Robot Manipulation VLM Adaptation
Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China. · PengCheng Laboratory, Shenzhen, China. · College of Intelligent Robotics and Advanced Manufacturing, Fudan University, Shanghai, China. +1
Scaling by Diversified Experience for Vision-Language-Action Models
Jun 8, 2026Leiyu Wang, Zhaofengnian Wang, Xueqi Li +3VLM Robustness Vision-Language-Action Models
Shanghai Jiao Tong University · Shanghai Innovation Institute · Tongji University +3

Jul 17, 2026J/K move · Enter open · S save

AC-VLA: Robust Out-of-Distribution Action Execution via Compositional Learning

Xiaojiang Peng, Kai Peng, Jie Lu, Zheng Lian+2

Shenzhen Technology University · Tongji University · Great Bay University+1

Robot Skill Learning Generalization in Robotic Manipulation Robotic Manipulation Vision-Language-Action Models Robustness of VLA Models

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models excel at end-to-end robotic manipulation but struggle with out-of-distribution (OOD) generalization when familiar sub-tasks are recombined in unseen configurations. We identify two mutually reinforcing failure modes: trajectory overfitting, where models overfit to holistic trajectory patterns rather than compositional sub-skill semantics; and perceptual shortcut, where action tokens over-rely on wrist-view textures at the expense of global spatial grounding. To address both, we introduce AC-VLA, a plug-and-play Action Compositional learning framework comprising two architecture-agnostic components: (i) a compositional learning module that uses an LLM-driven instruction decomposer and a proprioceptive trajectory aligner to generate dense sub-task supervision, followed by mixed training on complete demonstrations and decomposed data to endow the model with compositional generalization; and (ii) a state-conditioned asymmetric masking strategy that suppresses wrist-view inputs during closed-gripper phases, enforcing global semantic grounding. All components are architectural modification-free and directly integrable into any VLA backbone. Instantiated on

π_{0.5}

and evaluated on LIBERO and LIBERO-OOD benchmarks, AC-VLA achieves a ~28% absolute improvement on compositional OOD tasks while maintaining near-perfect in-distribution performance.

Abstract

Vision-Language-Action Models

Explore similar work

AC-VLA: Robust Out-of-Distribution Action Execution via Compositional Learning

M2M^2M2-VLA: Boosting Vision-Language Models for Generalizable Manipulation via Layer Mixture and Meta-Skills

Scaling by Diversified Experience for Vision-Language-Action Models

$M^2$ -VLA: Boosting Vision-Language Models for Generalizable Manipulation via Layer Mixture and Meta-Skills