Evolve Vision-Language-Action Model into an Agent with On-the-fly Tool-use

AI 导读

本文提出ART框架，将VLA模型与实时工具使用结合，提升机器人任务的泛化能力与数据效率，实验显示成功率提高20%。

This paper introduces ART, integrating VLA models with real-time tool use to enhance robotic task generalization and data efficiency, achieving 20% higher success rates in simulations and real-world tasks.

重点速览

• ART框架通过工具使用减少动作空间复杂性，提升任务泛化能力 ART reduces action space complexity via tool use, enhancing task generalization
• 实验显示ART在模拟和现实任务中成功率比主流方法高20% Experiments show 20% higher success rates than baselines in simulations and real-world tasks
• 模块化工具利用实现高效训练、轻量部署和新工具扩展 Modular tool utilization enables efficient training, lightweight deployment, and scalable tool integration
• 数据集规模仅为基线方法的1/10，降低数据依赖 Dataset size is 1/10 of baselines, reducing data dependency

一句话：ART框架通过模块化工具使用显著提升机器人任务的效率与适应性 / ART's modular tool use significantly enhances robotic task efficiency and adaptability

推荐理由 该研究为机器人系统在复杂环境中的实际部署提供了高效且可扩展的解决方案，具有重要应用价值