Enhancing Goal Inference via Correction Timing

A team from Georgia Tech and Oregon State University has published groundbreaking research showing that analyzing when humans correct robots—not just how they correct them—dramatically improves goal inference. Presented at the 25th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2026), the paper 'Enhancing Goal Inference via Correction Timing' demonstrates that the timing of human interventions contains valuable signals about task objectives that traditional correction analysis misses.

Traditional robot learning from corrections treats interventions either as new demonstrations or preference signals, focusing on the corrected trajectory itself. This research reveals that the decision moment to intervene—influenced by factors like task progress, motion legibility, and alignment with expectations—provides critical contextual information. The team tested three applications: identifying problematic robot motion features that trigger corrections, inferring correction goals faster using timing and initial direction, and learning more precise task constraints.

The results show significant improvements in the first two applications, with goal inference becoming 40% more accurate when incorporating timing data. This means robots can understand human intent faster and with fewer corrections, making human-robot collaboration more efficient. The research represents a paradigm shift in how robots interpret human feedback, moving beyond just analyzing what was corrected to understanding why the correction happened when it did. This has immediate implications for manufacturing, healthcare robotics, and domestic assistance robots where seamless human-robot interaction is critical.