Review of Decentralized Reinforcement Learning for Multi-Agent Robotic Task Scheduling

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Multi-robot task scheduling is fundamental to managing multiagent robotic systems (MARS) in unpredictable environments such as logistics, disaster management and autonomous exploration.Centralized task allocation scales poorly and adapts slowly to change, and reinforcement learning (RL) offers an alternative in which robots learn scheduling strategies through interaction.This review analyses 75 studies on multiagent task allocation with RL, with particular attention to PettingZoo's Pistonball environment as a multi-agent RL (MARL) benchmark.It combines a bibliometric profile of the corpus, a side-by-side comparison of representative studies, a qualitative comparison of learning paradigms and a graded appraisal of the evidence behind five recurring claims.Fully decentralized learning remains difficult because concurrent policy updates produce non-stationarity, whereas centralized training with decentralized execution (CTDE) has proved more practicable; hardware-validated schedulers are the weakestsupported area.The review identifies the need for systematic comparison of CTDE-based algorithms for vision-based multirobot task scheduling under identical conditions in the Pistonball environment.