Continual Robot Learning
Overview
The field studies how robots can keep learning from new tasks and real-world experience without losing skills they already use. Common approaches include replaying stored or generated experiences, adding modular adapters or skill experts, reusing skill libraries, and adapting policies through online reinforcement learning or interaction feedback. Researchers evaluate these methods in simulated manipulation and locomotion benchmarks such as LIBERO, Meta-World, and CALVIN, as well as on physical robots facing new tasks, terrain, or dynamics. A central challenge is balancing fast adaptation with reliable retention: replay and modularity can help, but forgetting remains, and safe validation of deployment-time updates is still an open concern.
Latest papers
89 papers24 papers in the last four weeks, up 85% on the four weeks before. 0.1% of all new papers.