Robo-Blocks: LLM-powered scaffolding helps novices program social robots

Programming social robots traditionally requires expertise in planning, interaction design, and coding—a steep barrier for novices. While large language models (LLMs) can generate code from natural language, they risk obscuring critical programming elements and fostering over-reliance rather than skill development. Researchers at the University of Wisconsin-Madison addressed this challenge with Robo-Blocks, a block-based programming environment that combines the accessibility of visual programming with LLM-powered generative scaffolding. Unlike typical code-generation tools that output black-box scripts, Robo-Blocks guides users through structured narratives—step-by-step story-like sequences that connect high-level intents (e.g., “make the robot greet visitors”) to concrete, executable behaviors. This approach ensures users understand the logic behind each block, preserving designer intent while accelerating development.

Through a Research through Design (RtD) process, the team deployed Robo-Blocks with novice robot programmers and analyzed how they interacted with the generative scaffolding. They discovered distinct user personas and usage patterns, revealing that different learners benefited from varying levels of scaffolding—some preferred more narrative guidance, while others experimented freely before consulting the LLM. The findings provide actionable design insights for future LLM-powered end-user programming tools, particularly how to balance automation with learning. The paper, presented at the 2026 ACM CHI conference, underscores that generative AI can empower novices without sacrificing comprehension, paving the way for more inclusive social robotics development.