SimVLA: A Simple VLA Baseline for Robotic Manipulation

Yuankai Luo; Woping Chen; Tong Liang; Baiqiao Wang; Zhenguo Li

cs.ROFeb 20, 2026

SimVLA: A Simple VLA Baseline for Robotic Manipulation

Authors: Yuankai Luo, Woping Chen, Tong Liang, Baiqiao Wang, Zhenguo Li

Organizations: Frontier Robotics

Abstract

Vision-Language-Action (VLA) models have emerged as a promising paradigm for general-purpose robotic manipulation, leveraging large-scale pre-training to achieve strong performance. The field has rapidly evolved with additional spatial priors and diverse architectural innovations. However, these advancements are often accompanied by varying training recipes and implementation details, which can make it challenging to disentangle the precise source of empirical gains. In this work, we introduce SimVLA, a streamlined baseline designed to establish a transparent reference point for VLA research. By strictly decoupling perception from control, using a standard vision-language backbone and a lightweight action head, and standardizing critical training dynamics, we demonstrate that a minimal design can achieve state-of-the-art performance. Despite having only 0.5B parameters, SimVLA outperforms multi-billion-parameter models on standard simulation benchmarks without robot pretraining. SimVLA also reaches on-par real-robot performance compared to pi0.5. Our results establish SimVLA as a robust, reproducible baseline that enables clear attribution of empirical gains to future architectural innovations. Website: https://frontierrobo.github.io/SimVLA

Figures & tables

Figure 1: Out-of-box real-robot task examples. We deploy SimVLA without any additional fine-tuning on our held-out scenes and evaluate it on a set of multi-stage tasks that require both dexterous manipulation and semantic understanding.

Model	Backbone	LIBERO Avg	VRAM (GB)
OpenVLA-OFT	7B	97.1	62.0 GB
$\pi_{0.5}$	3B	96.9	51.3 GB
VLA-Adapter	0.5B	97.3	24.7 GB
SimVLA (Ours)	0.5B	98.6	9.3 GB

Table 1: Performance and efficiency summary. We report LIBERO success (%) and peak training VRAM at Batch=8 (GB) under a matched evaluation setup.

Figure 2: SimVLA overview. SimVLA is a minimal baseline: a VLM encoder produces fused vision-language tokens once per control step, and a lightweight action transformer performs flow-matching denoising to generate a continuous action chunk.

Large Models ( $\geq$ 4B)
Model	B	Spatial	Object	Goal	Long	Avg
UniVLA ( Bu et al., 2025 )	9	96.5	96.8	95.6	92.0	95.2
FlowVLA ( Zhong et al., 2025 )	8.5	93.2	95.0	91.6	72.6	88.1
UnifiedVLA ( Wang et al., 2025c )	8.5	95.4	98.8	93.6	94.0	95.5
OpenVLA ( Kim et al., 2024 )	7	84.7	88.4	79.2	53.7	76.5
OpenVLA-OFT ( Kim et al., 2025 )	7	97.6	98.4	97.9	94.5	97.1

Table 2: Comparison on the LIBERO benchmark. We report the success rate (%) on the official test episodes for each suite (Spatial/Object/Goal/Long), and the overall average across the four suites. Bold ∗ denotes the best performance, and Bold denotes the second-best. “B” indicates the backbone scale in billions.

Task Suite	Method	Ori	Obj	Pos	Sem	Task
Spatial	OpenVLA ( Kim et al., 2024 )	98.0	97.0	0.0	97.0	0.0
	$\pi_{0.5}$ ( Black et al., 2025 )	98.0	97.0	20.0	97.0	1.0
	SimVLA	99.0	98.0	29.0	98.0	0.0
Object	OpenVLA ( Kim et al., 2024 )	99.0	98.0	0.0	98.0	0.0
	$\pi_{0.5}$ ( Black et al., 2025 )	98.0	98.0	17.0	96.0	1.0
	SimVLA	100.0	85.0	1.0	100.0	4.0

Table 3: Robustness evaluation on the LIBERO-PRO benchmark. We report the success rate (%) across five perturbation dimensions: Original (Ori), Object (Obj), Position (Pos), Semantic (Sem), and Task (Task). Bold indicates the best performance.

Large Models ( $\geq$ 4B)
Method	Spoon	Carrot	Stack	Eggplant	Avg
RT-1-X ( O’Neill et al., 2024 )	0.0	4.2	0.0	0.0	6.3
Octo-Base ( Team et al., 2024 )	12.5	8.3	0.0	43.1	31.3
OpenVLA ( Kim et al., 2024 )	0.0	0.0	0.0	4.1	7.8
OpenVLA-OFT ( Kim et al., 2025 )	12.5	4.2	8.3	37.5	39.6
DD-VLA ( Liang et al., 2025b )	29.2	29.2	20.8	70.8	54.2

Table 4: Comparison on WidowX robot tasks ; success rates (%).

Large Models ( $\geq$ 4B)
Model (Variant Aggregation)	Pick	Move	Open	Avg
RT-1-X ( O’Neill et al., 2024 )	49.0	32.3	29.4	36.9
RT-2-X ( O’Neill et al., 2024 )	82.3	79.2	35.3	65.6
Octo-Base ( Team et al., 2024 )	0.6	3.1	1.1	1.6
OpenVLA ( Kim et al., 2024 )	54.5	47.7	17.7	40.0
OpenVLA-OFT ( Kim et al., 2025 )	65.3	59.0	12.2	45.5

Table 5: Comparison on Google Robot tasks; success rates (%).

Data & representation (default: $H{=}10$ , shuffling on, normalization on)
Knob	Value	Spatial	Object	Goal	Long	Avg
SimVLA	Default settings	99.4	99.8	98.6	96.4	98.6
Action chunk horizon $H$	$H=20$	99.2	89.6	92.4	88.4	92.4
Action chunk horizon $H$	$H=30$	95.4	93.8	80.6	79.2	87.3
Data shuffling	off	6.2	0.0	13.6	0.0	9.9
Action normalization	off	22.6	3.2	23.2	0.0	12.3

Table 6: Ablations on LIBERO. Each row corresponds to one ablation setting with the remaining knobs fixed to the default configuration.

Figure 3: Real-robot zero-shot results on Galaxea R1 Lite.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Libero	WidowX	Google Robot
Batch size (per GPU)	64	80	80
Global batch size	64 $\times$ 4 = 256	80 $\times$ 4 = 320	80 $\times$ 4 = 320
Action head	Large (1024,24,16)	Large (1024,24,16)	Large (1024,24,16)
Action chunk horizon $H$	10	30	30
Image resize	128 $\times$ 128	224 $\times$ 224	224 $\times$ 224
Action normalization	on	on	on

Appendix

Table 7: Hyperparameters for simulation datasets training.

Configuration	Value
Batch size (per GPU)	32
Global batch size	32 $\times$ 64=2048
Action head	Large (1024,24,16)
Action chunk horizon $H$	30
Image resize	224 $\times$ 224
Action normalization	on

Appendix

Table 8: Hyperparameters for Galaxea-500h training.

Explore similar work

Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline
Feb 26, 2026Wenxuan Song, Jiayi Chen, Xiaoquan Sun +12Vision-Language-Action Models Efficient VLA Models
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · Huazhong University of Science and Technology, Wuhan, China. · Westlake University, Hangzhou, China. +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
SimVLA: Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation
Oct 8, 2026Kyoungin Baik, Youngwoon LeeSim-to-Real Transfer Vision-Language-Action Models
Yonsei University · Seoul National University

Feb 26, 2026J/K move · Enter open · S save

Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline

Wenxuan Song, Jiayi Chen, Xiaoquan Sun, Huashuo Lei+11

The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · Huazhong University of Science and Technology, Wuhan, China. · Westlake University, Hangzhou, China.+1

Vision-Language-Action Models Efficient VLA Models Mobile Manipulation

Open paper AskPDF arXiv

Vision-Language-Action (VLA) models have emerged as a generalist robotic agent. However, existing VLAs are hindered by excessive parameter scales, prohibitive pre-training requirements, and limited applicability to diverse embodiments. To improve the practicality of VLAs, we propose a comprehensive benchmark and an improved baseline. First, we propose CEBench, a new benchmark spanning diverse embodiments in both simulation and the real world with consideration of domain randomization. We collect 14.4k simulated trajectories and 1.6k real-world expert-curated trajectories to support training on CEBench. Second, using CEBench as our testbed, we study three critical aspects of VLAs' practicality and offer several key findings. Informed by these findings, we introduce LLaVA-VLA, a lightweight yet powerful VLA designed for practical deployment on consumer-grade GPUs. Architecturally, it integrates a compact VLM backbone with multi-view perception, proprioceptive tokenization, and action chunking. To eliminate reliance on costly pre-training, LLaVA-VLA adopts a two-stage training paradigm including post-training and fine-tuning. Furthermore, LLaVA-VLA extends the action space to unify navigation and manipulation. Experiments across embodiments demonstrate the capabilities of generalization and versatility of LLaVA-VLA , while real-world mobile manipulation experiments establish it as the first end-to-end VLA model for mobile manipulation. We will open-source all datasets, codes, and checkpoints upon acceptance to foster reproducibility and future research.

Abstract

Explore similar work

Rethinking the Practicality of Vision-language-action Model: A Comprehensive Benchmark and An Improved Baseline

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

SimVLA: Zero-Shot Sim-to-Real VLA Learning for Mobile Manipulation

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