A team at Fudan University's Institute of Trustworthy Embodied Intelligence has proposed RoboSkill, a framework that lets general-purpose robot agents keep what they learn on the job. Instead of collecting new demonstration data for every task, the agent explores a task, then saves the method that worked as a skill made of written guidance plus the code it actually ran. The paper, titled Explore, Execute, Evolve: A Skill Acquisition and Reuse Loop for Embodied Agents, was posted to arXiv, and the code is on GitHub.
The work starts from a shift in robotics. Vision-language-action models usually need fresh data from each deployment site, and robot data is costly, since people must teleoperate machines, reset scenes and repeat demonstrations. The team found that today's multimodal agents can instead observe, reason and write control code on the spot. In the LIBERO-10 simulation benchmark, with a four-hour budget and repeated attempts, the best agents reached 100 percent success. On a real robot, the strongest agent succeeded on 91.7 percent of simple tasks and 80 percent of hard ones.
The catch is speed. Agents loop through observing, reasoning, acting and checking, and the reasoning between short arm movements can take minutes. Average simulation runs took as long as 94.5 minutes, and successful real-robot tasks took 20 to 27 minutes.
RoboSkill attacks this in two ways. First, it adds touch. Pressure, force and torque readings tell the agent whether the gripper has made contact, so it does not need to keep checking images. With touch added, one agent's average simulation time fell from 25.35 minutes to 15.93, and a real test-tube retrieval task dropped from 33.63 minutes to 20.
Second, it accumulates skills. After a task, the agent reviews what happened and writes up the effective method, then keeps updating it with later results. A method for probing table height by contact, for example, can be reused in a new scene even though the height must be measured again. Skill reuse cut one agent's average simulation time from 94.5 minutes to 35.9.
Saving code alongside text mattered most. Moving from text-only skills to text plus code reduced average time for all four agents tested. For one, time fell from 67.1 to 39.9 minutes, and first-attempt success rose from 15 percent to 67.5 percent. Overall, the team reports first-round success gains of 12.5 to 25 points on LIBERO-10, simulation time savings of 7.6 to 72.4 percent, and an 8.3-point gain in real-robot success. The researchers say skills also transfer across tasks and across different agent models.
SEE ALSO: 30,000 Hours of Tactile Data Fills Embodied Intelligence Gap · Amap Open-Sources Lambda-0