<p>Microrobotic swarms, through their collective and reconfigurable behaviours, can flexibly adapt to various targeted delivery and navigation tasks. However, achieving autonomous navigation and obstacle avoidance in unknown environments remains a challenge. Here we propose a reinforcement-learning-based control strategy for swarm navigation under partial observation. To bridge the gap between simulation and reality, we introduce a generalized sim-to-real transfer method, enabling the simulation-trained swarm to effectively explore in unknown environments. Our model combines temporally extended attention with multi-level domain randomization of the environment, perception, actuation and dynamics, allowing the policy to use current sensory inputs and historical context to select magnetic actuation commands. Our model efficiently performs autonomous navigation and obstacle avoidance through benchmarking against human operators in simulated environments. The proposed strategy enables swarm navigation and dynamic obstacle avoidance, cargo transportation, moving target tracking and recovery from temporary vision loss, and hovering by its partial sensing and decision-making. Analysis of action sequences and attention scores reveals that the swarm makes task-priority-based decisions, optimizing trajectories towards the target in unknown environments.</p>