![[2602.18224] SimVLA: A Simple VLA Baseline for Robotic Manipulation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.18224] SimVLA: A Simple VLA Baseline for Robotic Manipulation
Summary
The paper introduces SimVLA, a streamlined Vision-Language-Action baseline for robotic manipulation, achieving state-of-the-art performance with minimal design.
Why It Matters
SimVLA addresses the complexities in VLA models by providing a clear, reproducible baseline that enhances understanding of empirical gains in robotic manipulation. This is crucial for future research and development in the field, as it simplifies comparisons and fosters innovation.
Key Takeaways
- SimVLA achieves superior performance with only 0.5B parameters.
- The model decouples perception from control, simplifying design.
- Standardized training dynamics allow for clear attribution of performance gains.
- It outperforms larger models without requiring robot pretraining.
- SimVLA serves as a reproducible baseline for future VLA research.
Computer Science > Robotics arXiv:2602.18224 (cs) [Submitted on 20 Feb 2026] Title:SimVLA: A Simple VLA Baseline for Robotic Manipulation Authors:Yuankai Luo, Woping Chen, Tong Liang, Baiqiao Wang, Zhenguo Li View a PDF of the paper titled SimVLA: A Simple VLA Baseline for Robotic Manipulation, by Yuankai Luo and 4 other authors View PDF HTML (experimental) 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 t...
