Event

FLEX language-guided representation editing method accepted to CoRL 2026

1 reports · 1 sources

Audrey Lee et al. propose FLEX (Flexible Feature Editing Pipeline), which uses a language model to iteratively correct a robot's internal representation of task features based on natural-language corrections through three operations: Add (add missing features), Generalize (generalize overly specific features), and Mask (mask irrelevant features); the paper has been accepted to CoRL 2026 (the 10th Conference on Robot Learning, Austin). The method was evaluated in 2D navigation scenarios, 7-DoF Franka tabletop manipulation tasks, and a real-world study using natural-language corrections and real-robot demonstrations; the authors report that comparisons with single-operation and scaffold-free baselines show that all three operations are necessary in interactive scenarios.