In this paper, we propose Interaction-Merged Motion Planning (IMMP), a novel method that effectively utilizes datasets from various source domains for motion planning of autonomous robots. While existing domain adaptation or ensemble learning methods suffer from domain imbalance, catastrophic forgetting, and high computational cost, IMMP adapts to the target domain by utilizing parameter checkpoints trained in different domains. It consists of two steps: a pre-merging step that captures agent actions and interactions, and a merging step that constructs an adaptive model that efficiently transfers various interactions to the target domain. Experimental results on various benchmarks and models show that IMMP outperforms existing methods.
