This paper proposes a novel bidirectional approach that integrates discrete commands and continuous actions for efficient decision-making in adversarial situations, such as strategic confrontations, in swarm robotics. Existing task and motion planning methods decouple decision-making into two layers, but their unidirectional structure fails to capture inter-layer interdependencies, limiting adaptability in dynamic environments. The proposed bidirectional approach, based on hierarchical reinforcement learning, effectively maps commands to task assignments and actions to path planning, utilizing cross-training techniques to enhance learning across the hierarchical framework. Furthermore, it introduces a trajectory prediction model that links abstract task representations to feasible planning goals. Experimental results demonstrate that the proposed approach outperforms existing methods, achieving a match-winning rate of over 80% and a decision-making time of less than 0.01 seconds. Demonstration through large-scale experiments and real-world robot experiments further highlights the generalizability and practicality of the proposed approach.
