Imagine robots self-organizing into large groups to assist people with physically demanding tasks, leveraging their core capabilities in perception, manipulation, and navigation to interact with the physical world. To successfully complete a given mission as a team, these robots must make their own decisions to allocate and carry out tasks defined in the mission and cooperate with one another when the task requires it. We realize this capability in multirobot systems through a learning-based paradigm, designing computational models to train a large number of robots to work as a team. Achieving a high level of autonomy in distributed information processing, decision-making, and collaborative manipulation is crucial. To enable a high level of cooperation among a large number of robots, the project aims to build computational models based on deep reinforcement learning (DRL). These models are designed to enhance the robots' ability to cooperate in carrying out multiple tasks using their perception, manipulation, and navigation skills. Additionally, the project's goal is to implement these models on a multi-robot platform and validate the effectiveness of the models through lab experiments.