Multi-agent systems learn to handle new tasks without retraining
SpecMAGIC: Specification-Guided Multi-Task Generalization in Multi-Agent Reinforcement Learning
Artificial Intelligence
Summary
Multi-agent teams often need to tackle many different tasks that require working together in specific ways. The authors create a way to describe these tasks clearly using a special language, and then break them down into smaller parts. Their system, SpecMAGIC, plans high-level task steps for the team while letting each agent handle its own part independently. This approach lets the team handle new, unseen tasks right away without needing to learn from scratch again. Tests show it works better than past methods, especially as team size grows.
What this means in practice
- •For robotics system designers: Create robot teams that instantly adapt to new complex tasks by reusing learned subtasks without retraining.
- •For warehouse automation engineers: Deploy multi-robot fleets that coordinate flexibly across varying task sequences to improve operational efficiency.
Authors
Zijian Guo, İlker Işık, H. M. Sabbir Ahmad, Wenchao Li
Abstract
Multi-task generalization is an important yet underexplored problem in multi-agent reinforcement learning (MARL). Many real-world applications require multi-agent systems to generalize across complex, temporally extended tasks with diverse inter-agent coordination requirements. We introduce a specification language that combines temporal-logic task specifications with operators for expressing such requirements. We propose SpecMAGIC, a hierarchical framework that leverages the automaton structure of task specifications for subtask decomposition. A centralized high-level planner identifies and allocates compatible subtask sequences across agents, while decentralized low-level policies execute individual subtasks. This enables zero-shot generalization to new tasks by composing reusable subtask policies according to new specifications, without retraining. We establish soundness and probabilistic completeness of the high-level planner. Experiments on multi-agent navigation tasks show that SpecMAGIC substantially outperforms existing approaches on unseen task specifications and maintains strong performance as the number of agents increases.