Vlaser

Abstract

While significant research has focused on developing embodied reasoning capabilities using Vision-Language Models (VLMs) or integrating advanced VLMs into Vision-Language-Action (VLA) models for end-to-end robot control, few studies directly address the critical gap between upstream VLM-based reasoning and downstream VLA policy learning. In this work, we take an initial step toward bridging embodied reasoning with VLA policy learning by introducing Vlaser -- a Vision-Language-Action Model with synergistic embodied reasoning capability, which is a foundational vision-language model designed to integrate high-level reasoning with low-level control for embodied agents. Built upon the high-quality Vlaser-6M dataset, Vlaser achieves state-of-the-art performance across a range of embodied reasoning benchmarks—including spatial reasoning, embodied grounding, embodied QA, and task planning. Furthermore, we systematically examine how different VLM initializations affect supervised VLA fine-tuning, offering novel insights into mitigating the domain shift between internet-scale pre-training data and embodied-specific policy learning data. Based on these insights, our approach achieves state-of-the-art results on the WidowX benchmark and competitive performance on the Google Robot benchmark. The code, model and data are available at https://github.com/OpenGVLab/Vlaser/.

Overview

Vlaser contains the following key contributions:

We propose Vlaser-6M dataset, which comprises multi-task embodied reasoning data sources including embodied QA, grounding, spaital intelligence, planning as well as in-domain simulation-sourced data.
Vlaser-8B achieves state-of-the-art embodied reasoning capabilities compared to other embodied reasoning VLMs.
The pretrained Vlaser VLM significantly accelerates convergence in downstream VLA policy learning.
Vlaser VLA achieves state-of-the-art results on the WidowX benchmark and competitive performance on the Google Robot benchmark on downstream closed-loop simulation evaluation.

Vlaser Architecture

Vlaser adopts a two-stage training recipe, designed to optimize both embodied reasoning and end-to-end robot control. It includes a VLM pretraining followed by a VLA finetuning:

(1) VLM Pretraining. The multimodal pretraining process for Vlaser is conducted by supervised fine-tuning (SFT) on Vlaser-6M dataset based on InternVL3 to empower VLM with stronger embodied reasoning capability.
(2) VLA Finetuning. Vlaser-VLA is finetuned by incorporating a flow-matching-based action expert on robot-specific datasets.

Performance

Embodied Reasoning Capability

Closed-loop Evaluation on WidowX

Closed-loop Evaluation on Google Robot

Citation


  @article{luo2025visual,
    title={Visual Embodied Brain: Let Multimodal Large Language Models See, Think, and Control in Spaces},
    author={Luo, Gen and Yang, Ganlin and Gong, Ziyang and Chen, Guanzhou and Duan, Haonan and Cui, Erfei and Tong, Ronglei and Hou, Zhi and Zhang, Tianyi and Chen, Zhe and others},
    journal={arXiv preprint arXiv:2506.00123},
    year={2025}
  }

Acknowledgement

This website is adapted from Nerfies, licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning