Rethinking the Practicality of Vision-Language-Action Model: A Comprehensive Benchmark and an Improved Baseline
Wenxuan Song, Jiayi Chen, Xiaoquan Sun, Huashuo Lei, Yikai Qin, wei zhao, Pengxiang Ding, Han Zhao, Tongxin Wang, Pengxu Hou, Zhide Zhong, Haodong Yan, Donglin Wang, Jun Ma, Haoang Li
AI summary
Problem
Existing vision-language-action models are hindered by excessive parameter scales, prohibitive pre-training costs, and limited applicability to diverse embodiments like mobile manipulation.
Approach
The authors introduce CEBench, a cross-embodiment benchmark spanning simulation and real-world environments with domain randomization, and propose LLaVA-VLA, a lightweight baseline that uses multi-view image concatenation, proprioceptive tokenization, action chunking, and a two-stage training paradigm to unify navigation and manipulation without costly pre-training.
Key results
- CEBench benchmark spanning single-arm, bimanual, and mobile manipulation across simulation and real-world settings with domain randomization
- LLaVA-VLA (0.5B) matches or exceeds 7B models on manipulation tasks while requiring less than 10% of the parameters
- Multi-view image concatenation and proprioceptive tokenization significantly boost performance over standard projection methods
- A two-stage post-training and fine-tuning curriculum eliminates the need for large-scale cross-embodiment pre-training
Why it matters
Enables practical deployment of VLA models on resource-constrained mobile robots by demonstrating that lightweight, pretraining-free architectures can achieve state-of-the-art performance across diverse embodiments.
Abstract
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 *Wenxuan Song, Jiayi Chen and Xiaoquan Sun contributed equally to this work. Corresponding author: Haoang Li (haoangli@hkust-gz.edu.cn). 1The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. 2Huazhong University of Science and Technology, Wuhan, China. 3Westlake University, Hangzhou, China. 4Zhejiang University, Hangzhou, China. 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. 2026 IEEE International Conference on Robotics and Automation (ICRA 2026) June 1-5, 2026. Vienna, Austria 979-8-3315-8160-2/26/$31.00 ©2026 IEEE 9114