ROS 2 + VLA: Can robots learn to work with vines dynamically?

The Italian Users Group of ROS (Robot Operating System) is exploring a provocative question: Can robots transcend scripted automation and learn to perform complex, context-dependent tasks—like vine pruning in agriculture—using ROS 2 and Vision-Language-Action (VLA) models?

This isn’t just about automating a repetitive motion. It’s about enabling robots to perceive their environment (e.g., identifying vine structures, assessing grapevine health), interpret human-defined goals (e.g., “prune this branch”), and dynamically learn how to act—rather than following a rigid, pre-programmed sequence. The proposed architecture integrates real-time 3D perception, imitation learning, and manipulation within ROS 2, forming a closed-loop system: Observe the scene, Understand context, Act with learned strategies, Verify the outcome, and Learn from feedback to improve future actions. While no off-the-shelf solution exists yet, the discussion is sparking interest across robotics professionals, particularly those working with embodied AI and robotic manipulation.

This line of inquiry aligns with broader trends in robotics, where ROS 2 serves as the infrastructure backbone, and VLA models provide the cognitive layer. Projects like CRISP (Closing the Gap Between ROS 2 and Robot Learning) and end-to-end imitation learning frameworks for robots such as SO-101 are already pushing the boundaries of what ROS 2 can do beyond traditional automation. The agricultural use case—vine pruning—is just one example of how adaptive, learning-based robotics could transform industries where variability and context are critical.