Recursive Self Improvement for Embodied Agents—A Survey of Foundations Mechanisms Harnesses and Evaluation

Embodied foundation models have expanded robot capabilities, but autonomous execution does not automatically provide autonomous improvement. Recursive self-improvement (RSI) concerns retained changes that can influence subsequent improvement, including the stronger case in which an agent revises the procedures that generate and evaluate future updates. This survey connects large-model learning and reasoning, agent memory and skills, and RSI theory. We compare general agent harnesses with embodied harnesses and organize the literature along four axes: improvement targets, update mechanisms, decision authority and evidence conditions. The method review covers reflection and knowledge revision, executable programs and harness evolution, adaptive experience acquisition, parameter learning, world-model revision, and cross-layer selection and meta-improvement. We analyze how large models guide interventions across these components and distinguish execution correction, persistent capability improvement and revision of the improver. Long-horizon tasks and physical priors are examined as complementary considerations for memory, execution, world modeling and evaluation within this broader framework. We compare representative methods and review benchmarks, datasets, development partitions and resource accounting. Existing evidence supports several forms of retained improvement, while comparable evidence for improved improvement procedures remains limited. Open problems include causal failure attribution, reliable feedback, safe inheritance, competence preservation, model validity, transferable improvers and principled stopping decisions.