A team of researchers from Tsinghua University, in collaboration with scholars from the National University of Singapore, has unveiled the ParallelWorld framework for embodied AI. This innovative framework integrates multi-step rehearsals into high-level strategic planning, empowering agents to meticulously explore various potential action pathways prior to taking any definitive action. By comparing subsequent observations and task progression across these diverse options, agents can make well-informed decisions. ParallelWorld adopts a rolling planning mechanism that seamlessly blends rehearsal and execution phases, facilitating adaptive adjustments in search breadth and depth. Consequently, it significantly elevates the success rate of pertinent manipulation tasks, improving from 56% to an impressive 83%. In the realm of most embodied spatial intelligence tasks, ParallelWorld surpasses the capabilities of sequential active exploration, delivering superior success rates across eight distinct robotic manipulation scenarios when compared to baseline models. Nonetheless, the framework exhibits limitations in tasks demanding precise geometric reasoning and intricate physical contact dynamics. Additionally, the adaptive search process necessitates approximately 5.8 times longer than direct execution. Future research endeavors will focus on addressing challenges such as enhancing the reliability of physical predictions and mitigating reasoning costs.