The field of robot learning has long envisioned generalist agents capable of robust interaction in
unstructured real-world environments. While the emerging paradigm of large-scale imitation learning
has granted robots broad initial capabilities, these systems frequently exhibit brittleness upon
deployment. A critical bottleneck lies in the reliance on static, pre-collected demonstration datasets,
as this data is often suboptimal, coverage is narrow, and deployment conditions are out of distribution.
This thesis posits that higher reliability requires moving beyond static demonstration datasets to the
effective operationalization of online experience, the agent’s direct interaction with the world, as an
informative signal for self-improvement and adaptation
