Researchers introduce LEXP, a system that helps a robot reuse lessons from earlier attempts at manipulation tasks. It combines task planning with a module that interprets execution evidence, saves lessons with conditions for when they apply, and feeds relevant guidance into later actions; the model weights and action interface stay fixed.
In the researchers’ tests across 12 real-world tasks, LEXP succeeded in 23 of 36 trials (63.9%), compared with 19 of 36 (52.8%) using memory alone and 2 of 36 (5.6%) with the base ChatGPT-6-Astra model. These are the paper’s reported results, not an independent evaluation.
Why it matters
For robotics researchers, the results suggest that retaining task-specific lessons can improve a fixed model’s performance across repeated physical attempts; how well the approach transfers beyond these 12 tasks remains unknown.
