Agentic Language-Grounded Adaptive Robotic Assembly

Nicholas Cote, Jaimyn Drake, Sachin Chitta

Keywords: Robotics, Large Language Models, Robotic Assembly, Adaptive Robotics

TL;DR: Using a team of VLA and LLM, we show that a robot can learn what a manufacturing process looks like and adapt when parts and processes vary on the factory floor.

Abstract: High-mix, low-volume manufacturing requires systems that can adapt to changing parts and processes. We present an agentic, language-grounded approach for robotic assembly that employs a team of Large Language Models and Vision-Language-Action Models to adapt to parts of similar kind and perform corrective actions during assembly. Given a high-precision wheel-on-axle insertion task, we demonstrate that our agentic approach outperforms a single-model and generalizes to out-of-distribution parts and grasps without changing the nominal assembly process.

Submission Number: 8