VLANeXt: A Recipe for Building Robust VLA Models

Conference: ICML 2026
arXiv: 2602.18532
Code: https://github.com/DravenALG/VLANeXt
Area: Embodied AI / VLA / Robotic Learning
Keywords: Vision-Language-Action models, Robotic learning, VLA design space, Multimodal fusion, Instruction-conditioned control

TL;DR

This paper systematically explores the VLA model design space, distilling 12 key design principles through 500+ controlled experiments to build the efficient and powerful VLANeXt model. It surpasses SOTA on the LIBERO benchmark and validates the effectiveness of these design principles in real-world robotic tasks.

Background & Motivation

Background: VLA models leverage pre-trained VLMs to provide visual and language understanding for general robotic policy learning. Numerous VLA models have been proposed (RT-2, OpenVLA, π series, etc.), but their training protocols and evaluation settings vary significantly.

Limitations of Prior Work: The VLA field remains in a "primordial soup" stage—ideas are abundant but lack systematicity. Different methods adopt different VLM backbones, architectural designs, and loss functions, making fair comparison difficult.

Key Challenge: How to systematically compare VLA design choices under a unified framework to distinguish which designs are truly effective?

Goal: Revisit the VLA design space under a unified framework and evaluation setup to discover reproducible and generalizable design recipes.

Key Insight: Starting from RT-2, the study evolves across three dimensions: base components, perception elements, and action modeling. This systematic ablation path clearly demonstrates the contribution of each design choice.

Core Idea: Gradually optimize design choices through large-scale controlled experiments (>500) under a unified evaluation protocol, integrating fragmented VLA methodologies into 12 actionable design principles.

Method

Overall Architecture

Pipeline: Multimodal input (multi-view RGB + proprioception + language instructions) → Multimodal LLM → Soft-link to policy module → Action chunk prediction + frequency-domain auxiliary objective. The core feature is the introduction of a learnable query buffer between the VLM and the policy module to achieve a smooth transition between representation spaces.

Key Designs

1. Soft-linking Policy Module and VLM:

   • Function: Establishes a soft information flow between the VLM text representation space and the policy module action prediction space.
   • Mechanism: Compared to the "text token reuse" (tight coupling) of RT-2 and the "complete decoupling" of MetaQuery, soft-linking adopts hierarchical connections with an inserted learnable query buffer. Each VLM layer output interacts with policy module queries through cross-attention, followed by conditioning on timestep information via adaLN.
   • Design Motivation: Resolves the underfitting of hard links and the information loss of complete decoupling. Soft-linking achieves optimal performance on LIBERO-plus (56.2%), a 2.5% improvement over loose coupling.

2. Multi-view + VLM-side Proprioception Fusion:

   • Function: Integrates robot proprioception and multi-view observations, fused at the VLM level.
   • Mechanism: Multi-view inputs are processed by the multimodal LLM image encoder; proprioception is converted into tokens via linear projection and input to the VLM alongside visual tokens. Proprioception should be injected at the VLM level rather than the policy module level.
   • Design Motivation: The alignment between state information provided by proprioception and visual instructions is higher at the VLM level (98.0% vs. 96.2%). Multi-view observations provide complementary geometric cues (91.8% → 97.6%).

3. Flow Matching + Frequency-domain Auxiliary Loss:

   • Function: Treats action chunk prediction (8 steps) as a continuous time series, using flow matching as the primary loss and frequency-domain MSE as the auxiliary objective.
   • Mechanism: The primary loss uses flow matching to model continuous action distributions. The frequency-domain auxiliary loss converts actions via DCT and assigns higher weights to low-frequency components \(L_{\text{freq}} = \text{MSE}(\text{DCT}(\hat{a}), \text{DCT}(a))\), where weights \(w(\text{freq}) \propto 1/(\text{freq}+1)\).
   • Design Motivation: Regression loss is surpassed by flow matching in high-performance regimes. Frequency-domain regularization prevents overfitting to trajectory jitter, reaching 99.0% performance (+1% compared to regression) without increasing training overhead.

Key Experimental Results

Main Results

Method	LIBERO (%)	LIBERO-plus (%)	Model Size
OpenVLA	76.5	15.6	7B
OpenVLA-OFT	97.1	69.6	7B
π₀	86.0	53.6	11B
π₀-Fast	85.5	61.6	7B
NORA	87.9	39.0	N/A
UniVLA	95.2	42.9	N/A
FLOWER	96.9	Unreported	N/A
VLANeXt	97.4	83.9	2.5B

VLANeXt surpasses all baselines with a 2.5B model size (approximately 1/3 the size of OpenVLA-OFT).

Ablation Study

Design Dimension	Configuration	LIBERO (%)	LIBERO-plus (%)
Base Components	Single-layer policy head + Text token reuse	19.8	<5.0
	Separate policy head (2 layers)	30.2	16.6
	Large policy module (12 layers)	64.4	34.0
	+ Action chunking (chunk=8)	74.6	43.4
	+ Flow matching loss	80.0	45.0
	+ Qwen3-VL-2B backbone	90.0	53.7
	+ Soft-link	91.8	56.2
Perception Elements	+ Multi-view	97.6	80.5
	+ VLM-side proprioception	98.0	87.7
Action Modeling	+ Frequency-domain auxiliary loss	99.0	93.1

Key Findings

• Large policy modules provide the greatest contribution (+33.8%).
• Strong VLM backbones (+10.0%) are superior to merely increasing parameters.
• Perception elements (multi-view + proprioception) add a cumulative +13.0%.
• Frequency-domain loss is simple yet effective, with negligible computational overhead.
• Video history is not beneficial—adding temporal history actually causes performance drops (91.8% → 85.0%).
• Proprioception placement is sensitive—VLM-level injection is far superior to policy-level (98.0% vs. 96.2%).

Highlights & Insights

• Systematic Design Exploration: 500+ controlled experiments decompose the VLA design space under a unified framework. The "recipe" mindset offers more methodological value to the community than isolated innovations.
• Deep Insights into Multimodal Fusion: Proprioception should be fused on the VLM side rather than the policy side, as multi-view observations provide geometric compensation.
• Time-Series Concept Transfer: Migrating frequency-domain regularization from time-series forecasting to action generation is simple and elegant, significantly improving robustness to disturbances in LIBERO-plus.
• Efficiency-Performance Balance: The 2.5B VLANeXt is significantly smaller than the 7B OpenVLA-OFT yet achieves superior performance.
• Open-Source Contribution: Releasing a unified, lightweight framework lowers the barrier to entry for VLA research.

Limitations & Future Work

• Evaluation is limited to LIBERO/LIBERO-plus simulation benchmarks, with a small sample size for real-world robot validation.
• Dataset characteristics—LIBERO primarily involves manipulation tasks, lacking diverse scenarios like navigation.
• Computational efficiency—inference latency and VRAM usage for the 2.5B model were not reported in detail.
• Improvements: Transferring designs across embodiments and tasks; adaptive fusion strategies; integration with online learning; in-depth analysis of the frequency-domain loss mechanism.

Related Work & Insights

• vs RT-2/OpenVLA: VLANeXt outperforms OpenVLA-OFT 7B at a 2.5B scale through optimized design details.
• vs π series: π is tightly coupled at 11B; VLANeXt’s soft-link is more lightweight with better performance.
• vs World Model methods (WorldVLA): Auxiliary tasks add +2% but triple training time; VLANeXt replaces these with frequency-domain loss to achieve a better efficiency-performance balance.

Rating

• Novelty: ⭐⭐⭐⭐ Systematically explores the VLA design space; contribution to methodology is significant.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 500+ controlled trials + 2 simulation benchmarks + real robots + exhaustive ablations.
• Writing Quality: ⭐⭐⭐⭐ Clear framework; smooth evolution of design choices.
• Value: ⭐⭐⭐⭐⭐ Sets a precedent for shifting VLA research from fragmented exploration to systematic design.

Related Papers

• [ICML 2026] Any3D-VLA: Enhancing VLA Robustness via Diverse Point Clouds
• [ICML 2026] VLA-Arena: An Open-Source Framework for Evaluating Vision-Language-Action Models
• [ICML 2026] Self-Captioning Multimodal Interaction Tuning: Amplifying Exploitable Redundancies for Robust Vision Language Models
• [CVPR 2026] Dynamic Token Reweighting for Robust Vision-Language Models
• [ICLR 2026] Directional Embedding Smoothing for Robust Vision Language Models