A new framework called RoboSkill lets robots learn from their own task attempts, then reuse what worked instead of starting over each time.
Researchers built RoboSkill around a three-step loop: the agent explores a task to gather information, executes it while adjusting to feedback, then updates a library of reusable skills from the resulting records. The next time a similar task comes up, it pulls from that library rather than reasoning from scratch, uses touch sensing alongside vision to cut uncertainty during physical contact, and stores some skills as runnable code instead of plain-text instructions. On the LIBERO-10 benchmark, across four different agents, the loop raised first-attempt success rates by 12.5 to 25.0 percentage points and cut average runtime by 7.6 to 72.4 percent. On real robots, success rates rose 8.3 percentage points, and runtime for successful trials dropped by at least 14.4 percent.
The real cost problem in embodied AI is not the task itself, it is the re-thinking. General-purpose multimodal agents are decent at zero-shot robotic tasks, but they pay for that flexibility by reasoning through the physical world from scratch on every run. A system that remembers what already worked and skips redundant exploration is the more useful upgrade for anyone trying to deploy robots that handle repeat tasks, not one-off demos.
That 7.6 to 72.4 percent runtime range is wide enough to suggest skill reuse pays off unevenly across agents and tasks, and the paper's real-robot numbers are a small sample next to the simulated benchmark.