Coding a Classical Robot Controller in the Age of Coding Agents (discourse.openrobotics.org)

In the AI for Industry Challenge, the team MacCody explored a unique approach to robotics by implementing classical control methods augmented with coding agents to tackle a complex task involving the UR5e robot. The challenge revolved around inserting fiber connectors into a randomized board using various sensory inputs, where many would typically default to machine learning models due to the complexity of the task. Their experiment aimed to determine if a human-coding agent collaboration could efficiently develop a robust classical controller without extensive training data. Their findings, while not placing them in the top rankings, highlighted that coding agents significantly enhanced the speed of experimentation, allowing for rapid iterations, debugging, and tool development across various facets of the project. The significance of this work lies in the demonstration that coding agents can effectively support robotics tasks traditionally thought to require learned policies. The project underscored the potential of adapting classical control methodologies, further developing a staged visuotactile controller that evolved through multiple development phases. Key technical insights included leveraging geometric techniques for perception instead of relying solely on learned models, leading to a nuanced understanding of interactions between visual cues, contact mechanics, and recovery logic in complex environments. This work illustrates a shift towards reconsidering the use of explicit programming in robotics, potentially influencing how future challenges in AI and robotics are approached.