ART framework cuts robot data needs, boosts success by 20% with tool-use

Vision-language-action (VLA) models have advanced robot control, but they typically require enormous datasets and struggle to generalize beyond training conditions. In a new CVPR 2026 Findings paper, Yi Ding and colleagues propose ART (Agentic Robot with Tool-use), a tool-injection framework that evolves any VLA model into an agent capable of leveraging off-the-shelf tool modules. Rather than predicting a continuous action space end-to-end, ART selects and applies tools for low-level vision, high-level affordance, and embodiment enhancement. This dramatically shrinks the action solution space, improving cross-task generalizability while reducing data dependency—a critical step for real-world deployment.

To validate the approach, the team built a 30K dataset of tool-use trajectories and action demonstrations, far smaller than those used by existing baselines. They also designed a training regimen for long-horizon tool-use reasoning in challenging environments. Experiments show ART achieves a 20% higher success rate than mainstream VLA baselines on both simulation and real-world tasks, including pick-and-place in darkness at novel viewpoints. The agentic design also enables more efficient training, lightweight deployment, and easy integration of new tools. These results suggest that combining tool-use with VLA models is a practical path toward robust, adaptable, and extensible robot agents in complex environments.