VLANeXt: Recipes for Building Strong VLA Models

Xiao-Ming Wu; Bin Fan; Kang Liao; Jian-Jian Jiang; Runze Yang; Yihang Luo; Zhonghua Wu; Wei-Shi Zheng

cs.CVFeb 20, 2026

VLANeXt: Recipes for Building Strong VLA Models

Authors: Xiao-Ming Wu, Bin Fan, Kang Liao, Jian-Jian Jiang, Runze Yang, Yihang Luo, Zhonghua Wu, Wei-Shi Zheng, +1 more

Organizations: 1S-Lab, Nanyang Technological University · 2ACE Robotics · 3Sun Yat-sen University · 4Shanghai Jiao Tong University · 5SenseTime Research

Abstract

Following the rise of large foundation models, Vision-Language-Action models (VLAs) emerged, leveraging strong visual and language understanding from Vision-Language Models for general-purpose policy learning. Yet, the current VLA landscape remains fragmented and exploratory. Although many groups have proposed their own VLA models, inconsistencies in training protocols and evaluation settings make it difficult to identify which design choices truly matter. To bring structure to this evolving space, we reexamine the VLA design space under a unified framework and evaluation setup. Starting from a simple VLA baseline similar to RT-2, which is the origin of VLA, we systematically dissect design choices along three dimensions: foundational components, perception essentials, and action modelling perspectives. From this study, we distill 12 key findings that together form a practical recipe for building strong VLA models. The outcome of this exploration is a simple yet effective model, VLANeXt. It outperforms the state-of-the-art methods on the LIBERO and LIBERO-plus benchmarks and demonstrates strong performance in real-world experiments. We release a unified and easy-to-use codebase to reproduce our findings, explore the design space, and develop new VLA variants on top of a shared foundation. The codebase is available at https://github.com/DravenALG/VLANeXt.

Figures & tables

Figure 1 : Performance comparison on the LIBERO and LIBERO-plus benchmarks . We compare VLANeXt with representative VLA baselines across model scales. Despite its smaller model size, VLANeXt achieves higher success rates than prior methods on both standard task performance (LIBERO) and robustness/generalization (LIBERO-plus), demonstrating the effectiveness of the design recipe distilled in this work.

Figure 2 : Ablation trajectory across the VLA design space (spatial suite) . We progressively evolve a baseline VLA through changes in foundational components, perception, and action modeling. Results are reported on LIBERO initially, and on LIBERO-plus once LIBERO performance saturates, providing a more sensitive test of robustness and generalization. The trajectory culminates in the final VLANeXt model (2.5B) vs. OpenVLA-OFT (7B).

Figure 3 : Design choices for the policy module.

Figure 4 : Design choices for the VLM-Policy connection.

Foundational Components
	LIBERO	LIBERO-plus
Model	Original	Camera	Robot	Language	Light	Background	Noise	Layout	Total
RT-2 like Baseline
Baseline	19.8	-	-	-	-	-	-	-	$<$ 5.0
Policy Module Design
Baseline	19.8	-	-	-	-	-	-	-	$<$ 5.0

Table 1 : Ablation across the VLA design space on LIBERO, LIBERO-plus (spatial suite). Each block varies in one design aspect. LIBERO-plus evaluates robustness under diverse perturbations. Performance improves steadily as effective design choices are incorporated.

Figure 5 : Design choices for proprioception conditioning.

Figure 6 : Augmenting action prediction with an auxiliary world modeling objective.

Figure 7 : VLANeXt architecture . Multi-view visual inputs, language instructions, and proprioception are tokenized and processed by a multimodal LLM, with meta queries enabling soft interaction with the policy module. Action chunks are predicted using flow matching and further regularized by a frequency-domain objective.

Baseline Direct Policy Models
Model	S	O	G	L	Avg
Diffusion Policy	78.3	92.5	68.3	50.5	72.4
Octo	78.9	85.7	84.6	51.5	75.1
MDT	78.5	87.5	73.5	64.8	76.4
Baseline VLA Models
TraceVLA	84.6	85.2	75.1	54.1	74.8

Table 2 : LIBERO benchmark performance. The results are shown in success rate (%). S, O, G, L: Spatial, Object, Goal, and Long suites, respectively. We color the best and second best results.

Baseline VLA Models
Model	Suite	Camera	Robot	Language	Light	Background	Noise	Layout	Total
OpenVLA	Average	0.8	3.5	23.0	8.1	34.8	15.2	28.5	15.6
WorldVLA	Average	0.1	27.9	41.6	43.7	17.1	10.9	38.0	25.0
NORA	Average	2.2	37.0	65.1	45.7	58.6	12.8	62.1	39.0
UniVLA	Average	1.8	46.2	69.6	69.0	81.0	21.2	31.9	42.9
$\pi_{0}$	Average	13.8	6.0	58.8	85.0	81.4	79.0	68.9	53.6

Table 3 : LIBERO-plus benchmark performance. All the evaluated methods are trained using the LIBERO dataset, and are directly tested in the LIBERO-plus benchmark with unseen perturbations. The results are shown in success rate (%). We color the best and second best results (in average). The complete per-suite results of the listed methods can also be found in ( Fei et al., 2025b ) .

Figure 8 : Our single-arm and bimanual arm tasks for the real-world experiments.

Baseline VLA Models
	Single Arm		Bimanual Arm
Model	Clean	Drawer	Clean	Lifting
OpenVLA-OFT	7/20	7/20	5/20	9/20
$\pi_{0}$	10/20	8/20	10/20	13/20
Ours
VLANeXt	14/20	11/20	11/20	15/20

Table 4 : Real-world evaluation results. Results are shown with (success count/total count). We color the best and second best .

Latent Action Models
Model	Backbone	Loss	LIBERO
Model	Backbone	Loss	Spatial	Object	Goal	Long	Avg
VLANeXt	Qwen3-VL-2B	flow match. (noise)	99.0	99.2	96.6	94.8	97.4
VLANeXt-LAM	Qwen3-VL-2B	LAM (vicreg)	95.4	99.6	97.0	95.4	96.9
VLANeXt-LAM	Qwen3-VL-2B	LAM (vae)	98.0	97.8	95.2	96.2	96.8
VLANeXt-LAM	Qwen3-VL-2B	LAM (vq-vae)	98.0	99.8	95.6	97.0	97.6

Table 5: Performance of the additional baselines on the LIBERO Benchmark.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9 : Qualitative experiments of our method in real-world tasks.

Figure 10 : Qualitative experiments of our method in the four suites of the LIBERO benchmark.

Figure 11 : Qualitative experiments of our method in the 7 types of perturbations in the same task in the LIBERO-plus benchmark.

Model	LIBERO					LIBERO-Plus
Model	Spatial	Object	Goal	Long	Avg	Spatial	Object	Goal	Long	Avg
VLANeXt $-$ freq. $-$ prop. $-$ mv. $-$ soft	90.0	31.2	93.6	83.4	74.6	53.7	38.2	36.7	50.9	44.9
VLANeXt $-$ freq. $-$ prop. $-$ mv.	91.8	61.8	94.8	90.2	84.7	56.2	41.4	47.3	54.0	49.7
VLANeXt $-$ freq. $-$ prop.	97.6	98.8	96.0	93.8	96.6	80.5	74.4	69.0	68.1	73.0
VLANeXt $-$ freq.	98.0	99.4	95.6	95.0	97.0	87.7	77.9	71.3	71.4	77.1
VLANeXt	99.0	99.2	96.6	94.8	97.4	93.1	86.5	76.2	79.7	83.9

Appendix

Table 6 : Results on LIBERO and LIBERO-Plus Benchmarks. We color the best and second best results. Freq. means the frequency-domain loss from the time-series forecasting perspective. Prop. means the input of proprioception to the VLM. Mv. means multi-view inputs. And soft means soft connection.

Model	Real-world clean table
VLANeXt w/o. action chunk	1/10
VLANeXt w/o. multiview	6/10
VLANeXt w/o. soft connection	5/10
VLANeXt w/o. proprioception	4/10
VLANeXt	7/10

Appendix

Table 7 : Ablation Study on Real-World Clean Table. We color the best .

Optimization
Hyperparameter	Value
Optimizer	AdamW
Learning Rate	$1.0\times 10^{-4}$
Batch Size	256
Training Steps	10,000
Warmup Steps	500

Appendix

Table 8 : Hyperparameters for VLANeXt training on LIBERO and LIBERO-plus benchmark, four suites.

Explore similar work

StarVLA- $α$ : Reducing Complexity in Vision-Language-Action Systems
Apr 13, 2026Jinhui Ye, Ning Gao, Senqiao Yang +7Diffusion-Based Vision-Language-Actions Scalable Robot Learning
HKUST · XJTU · CUHK +3
vla-eval: A Unified Evaluation Harness for Vision-Language-Action Models
Mar 14, 2026Suhwan Choi, Yunsung Lee, Yubeen Park +4Preprocessing
Allen Institute for AI (AI2) · Seoul National University
VLA Foundry: A Unified Framework for Training Vision-Language-Action Models
Apr 21, 2026Jean Mercat, Sedrick Keh, Kushal Arora +5Vision-Language-Action Framework Qwen3
Toyota Research Institute

Apr 13, 2026, cs.ROJ/K move · Enter open · S save

StarVLA- $α$ : Reducing Complexity in Vision-Language-Action Systems

Jinhui Ye, Ning Gao, Senqiao Yang, Jinliang Zheng+6

HKUST · XJTU · CUHK+3

Diffusion-Based Vision-Language-Actions Scalable Robot Learning

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models have recently emerged as a promising paradigm for building general-purpose robotic agents. However, the VLA landscape remains highly fragmented and complex: as existing approaches vary substantially in architectures, training data, embodiment configurations, and benchmark-specific engineering. In this work, we introduce StarVLA-

α

, a simple yet strong baseline designed to study VLA design choices under controlled conditions. StarVLA-

α

deliberately minimizes architectural and pipeline complexity to reduce experimental confounders and enable systematic analysis. Specifically, we re-evaluate several key design axes, including action modeling strategies, robot-specific pretraining, and interface engineering. Across unified multi-benchmark training on LIBERO, SimplerEnv, RoboTwin, and RoboCasa, the same simple baseline remains highly competitive, indicating that a strong VLM backbone combined with minimal design is already sufficient to achieve strong performance without relying on additional architectural complexity or engineering tricks. Notably, our single generalist model outperforms

π_{0.5}

by 20% on the public real-world RoboChallenge benchmark. We expect StarVLA-

α

to serve as a solid starting point for future research in the VLA regime. Code will be released at https://github.com/starVLA/starVLA.

Abstract

Explore similar work

StarVLA-ααα: Reducing Complexity in Vision-Language-Action Systems

vla-eval: A Unified Evaluation Harness for Vision-Language-Action Models

VLA Foundry: A Unified Framework for Training Vision-Language-Action Models

StarVLA- $α$ : Reducing Complexity in Vision-Language-Action Systems