Fudan's RoboSkill lets robot agents save learned skills as text and code

A team at Fudan University's Institute of Trustworthy Embodied Intelligence proposed RoboSkill, a framework where robot agents save effective methods as written guidance plus executable code and reuse them later. On LIBERO-10, first-attempt success rose by 12.5–25 points and simulation time fell by 7.6–72.4%.

• Best agent succeeded on 91.7% of simple and 80% of hard real-robot tasks
• Adding tactile data cut average simulation time from 25.35 to 15.93 minutes
• Skill reuse reduced one agent's simulation time from 94.5 to 35.9 minutes
• Text-plus-code skills raised first-attempt success from 15% to 67.5% for one agent