Yifei Yao
Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents
An embodied agent learns from execution experience by refining text memory across training episodes, then uses frozen memory at evaluation without updating model weights.
SimplEvo builds task memory from a demonstration and refines working memory across training episodes. The agent records lessons about task procedure, completion checks, and recovery from its own executions. Both memories are frozen for evaluation, with no model weight updates or post-trained robot policies.
Experiments on EBench and RoboDojo examine task performance, memory design, scene generalization, and transfer across multimodal models. The preprint is awaiting arXiv release; the paper link will be added when available.