Cleaning up a robot dialogue dataset tripled its performance.

Teaching industrial robots to talk with human workers is difficult when training data has errors. An earlier dataset called IRWOZ helped robots hold conversations, but its initial version had lots of noisy dialogue states and messy speech, which harmed performance. To fix this, researchers created IRWOZ 2.0. They used large language modelslarge language modelsAI systems trained on vast amounts of text that can generate and understand human language., specifically Mistral and Claude-3.5, to generate better text. They also added human corrections and removed typos automatically to clean everything up. The updated dataset now covers four industrial areas: Assembly, Delivery, Position, and Relocation. In testing, tracking the state of the conversation worked much better. A model called GPT-2 saw its BLEU-4 language score jump from 0.1651 on the old data up to 0.5604 on IRWOZ 2.0.

Why it matters

This helps factory workers collaborate more safely and efficiently with robots. When robots understand commands better, tasks get done with fewer mistakes.

Questions answered