An LLM-Driven Closed-Loop Autonomous Learning Framework for Robots Facing Uncovered Tasks in Open Environments

Hong Su's new framework allows autonomous robots to learn uncovered tasks in open environments without relying on repeated LLM interactions. The system uses an LLM as a high-level reasoner to analyze tasks, select candidate models, plan data collection, and organize execution or observation strategies. Robots learn from both self-execution and observing others, then train and adjust in quasi-real-time, storing validated results in a local method library for future reuse. This closed-loop process reduces LLM calls from 1.0 to 0.2 per task and cuts average execution time from 7.78s to 6.78s in repeated-task experiments.

The framework marks a shift from dependency on cloud-based LLMs to self-contained learning. By converting both execution-derived and observation-derived experience into reusable local capability, robots can handle novel tasks more efficiently. The approach is particularly valuable for applications in dynamic environments like disaster response, exploration, or manufacturing where predefined methods fall short. With reduced latency and lower operational costs, this framework could accelerate deployment of truly autonomous robots that adapt on the fly without constant human or cloud intervention.