USTC's SERA Edits VLM Planner Harness Between Rollouts, Hits 51.4% on RoboMME Long-Horizon Manipulation Tasks

SERA (Self-Evolving Robotic Agent), from USTC, SJTU, Shanghai AI Lab and MBZUAI, treats a VLM planner as a robotic agent inside an editable harness that tracks task context and uses execution feedback to improve subsequent planning. The planner proposes and evaluates harness revisions from environment rollouts, with the selected configuration frozen before held-out tests. Paired with a fixed VLA executor, SERA reaches 51.38% overall success on RoboMME, outperforming history-blind VLA baselines on long-horizon, non-Markovian manipulation.