arxiv DoReMi: Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment