Navigating hundreds of robots in a warehouse without causing accidents is a growing challenge for many industries, from e-commerce to automotive production. To address this issue, a team of researchers at the Massachusetts Institute of Technology (MIT) has developed an AI-based approach that increases efficiency and reduces congestion.
The team devised a deep learning model that captures key information about the warehouse, such as the planned paths of the robots and the tasks they are set to accomplish. This information is then used to identify the best area of the warehouse to decongest in order to increase overall operational efficiency.
The researchers’ technique separates the warehouse robots into smaller groups, making it easier and quicker to decongest these robots using traditional algorithms. This approach has proven nearly four times faster at resolving congestion than a strong random search method.
According to Cathy Wu, who led the research, this new neural network architecture is especially suited for real-time operations. It can process and track the movements of hundreds of robots, encoding their trajectories, origins, and destinations along with their relationships with other robots.
The researchers liken the activity in the warehouse to a game of “Tetris” at high speed. As customer orders come in, robots are dispatched to retrieve items and bring them to human operators for packing. With hundreds of robots in motion, potential collisions are a constant risk.
Traditionally, search-based algorithms prevent collisions by keeping one robot on course while altering the path of another. However, this approach becomes increasingly complex with more robots to coordinate. Wu and her team’s method focuses replanning on areas with the most congestion, saving time.
The neural network can work efficiently because it is able to process the multiple complex relationships between individual robots. It achieves this by encoding constraints only once for each group of robots, which reduces redundancy and streamlines computation.
In tests, this deep-learning approach decongested warehouses up to four times faster than non-learning-based methods, even with the additional computing resources required to run the deep learning model.
In the future, researchers aim to develop simpler, rule-based methodologies that are easier to implement and maintain in warehouse settings. Their work was supported by Amazon and the MIT Amazon Science Hub. The team’s findings will be presented at the International Conference on Learning Representations.
