Researchers at MIT have developed an artificial intelligence tool to improve the efficiency of robotic warehouses. The researchers, drawing from their work in moderating traffic congestion, created a deep-learning model that can process information about the warehouse, including robot paths, tasks, and obstacles, to optimize warehouse functionality. By grouping the robots into smaller units, the AI could perform robot decongestion at four times the speed of a strong random search method. This technique has potential uses in other complex planning tasks, such as computer chip design and pipe routing in large buildings.
Managing a warehouse full of up to 800 robots moving goods around – similar to cars navigating through a crowded city center – can be complex. Traditional algorithms struggle to keep up with the fast pace of these operations. The new system, however, can encode and process the trajectories, origins, destinations, and inter-robot relationships in real-time, according to the research team.
The researchers liken the process of a robotic warehouse to a game of “Tetris.” When a customer order is received, the robot retrieves the item and delivers it to a human operator for packaging. This is repeated hundreds of times, and collisions can occur if any robot paths overlap. To solve this, the AI system was designed to focus on the most congested areas for replanning purposes, effectively utilizing time.
Furthermore, the system can analyze smaller groups of robots simultaneously. For instance, it can divide an 800-robot warehouse into groups of 40, which helps to predict which group could best optimize overall operations if their trajectories were coordinated efficiently.
The researchers’ approach allowed for less repetition and greater efficiency by encoding constraints for each subproblem only once. Their new technique was able to decongest a warehouse up to four times faster than non-learning-based methods, solving the issue at 3.5 times the speed. The team now aims to simplify the neural model into rule-based insights for easier implementation.
This work was supported by Amazon and the MIT Amazon Science Hub and will be presented at the International Conference on Learning Representations.
