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

TL;DR

Vlaser addresses the gap between upstream embodied reasoning in Vision-Language Models and downstream Vision-Language-Action policy learning by introducing a two-component embodied VLM with a dedicated data engine and a two-stage training pipeline. Using the Vlaser-6M dataset, it achieves state-of-the-art results across 12 embodied-reasoning benchmarks and demonstrates that in-domain simulated data more effectively accelerates VLA fine-tuning than out-of-domain data, while revealing a domain gap that limits transfer to real robots. The architecture couples an InternVL3-based VLM backbone with a flow-matching action-expert for low-level control, enabling robust open-loop reasoning and closed-loop manipulation, and provides practical guidance on data streams for VLA transfer. The work offers open-source resources to support reproducibility and future research in embodied AI, with implications for real-world robotic control and generalization across tasks and embodiments.

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.

Figures (7)

• Figure 1: Overall framework, capabilities, and evaluation of Vlaser.Top-left: Composition of the Vlaser-6M dataset, featuring multi-task embodied data—including QA, grounding, spatial reasoning, and planning—along with in-domain simulation-sourced pairs. Top-right: A LiDAR visualization illustrating the state-of-the-art embodied reasoning capability of the Vlaser VLM. Bottom-left: The pre-trained Vlaser VLM significantly accelerates convergence in downstream Vision-Language Action model (VLA) policy learning on WidowX platform bridgedata. Bottom-right: Successful closed-loop operation of an agent powered by Vlaser within the SimplerEnv benchmark li24simpler.
• Figure 2: An illustration of Vlaser architecture. Vlaser includes two components and corresponding training phases: 1) the Multimodal Pretraining is for embodied reasoning enhancement based on the corresponding data engine; 2) VLA training is performed on the action expert module, which handles low-level control based on flow matching action generation.
• Figure 3: An illustration of Vlaser-6M data engine for in-domain general QA sample in SimplerEnv.
• Figure 4: An illustration of Vlaser-6M data engine for in-domain embodied grounding QA sample in SimplerEnv.
• Figure 5: An illustration of Vlaser-6M data engine for in-domain spatial reasoning QA sample in SimplerEnv.