In a vast robotic warehouse, hundreds of robots move swiftly across the floor, picking up items and delivering them to human workers for packing and shipping. These types of warehouses are becoming increasingly common in various industries, such as e-commerce and automotive production.
However, managing 800 robots efficiently and preventing collisions is a challenging task. Even the most advanced path-finding algorithms struggle to keep up with the fast pace of e-commerce and manufacturing.
A team of MIT researchers, who specialize in using AI to address traffic congestion, applied their expertise to solve this problem. They developed a deep-learning model that considers key information about the warehouse, robots, planned paths, tasks, and obstacles to predict the best areas to alleviate congestion and enhance overall efficiency.
Their approach involves dividing the robots into smaller groups and using traditional algorithms to coordinate these groups, allowing for faster decongestion. This method proved to be nearly four times faster than a random search method.
Besides optimizing warehouse operations, this deep learning technique could also be applied to other complex planning tasks, such as computer chip design or pipe routing in large buildings.
Cathy Wu, the Gilbert W. Winslow Career Development Assistant Professor in Civil and Environmental Engineering, explained, “We created a new neural network architecture that can handle real-time operations in large warehouses. It efficiently processes information about hundreds of robots, their paths, origins, destinations, and interactions with other robots.”
Wu collaborated with Zhongxia Yan, a graduate student in electrical engineering and computer science, on this project. Their work will be presented at the International Conference on Learning Representations.
From a top-down view, a robotic e-commerce warehouse resembles a fast-paced game of Tetris. When an order comes in, a robot retrieves the requested item and delivers it to a human operator for packing. With hundreds of robots working simultaneously, collisions can occur if their paths intersect.
Traditional algorithms prevent crashes by adjusting the trajectory of one robot while replanning a new path for the other. However, with numerous robots and potential collisions, the complexity grows exponentially.
Given the need for rapid replanning, the MIT researchers use machine learning to focus on congested areas where travel time can be reduced the most.
The neural network architecture developed by Wu and Yan considers smaller groups of robots and predicts which group can improve the overall solution the most. An iterative process is used to decongest each group, leading to significant improvements in efficiency.
The neural network can efficiently reason about groups of robots by capturing intricate relationships between individual robots, even if their initial paths are far apart but intersect during their journeys.
The researchers’ approach minimizes computation by encoding constraints only once, rather than repeating the process for each subgroup. This results in faster decongestion and improved overall efficiency.
The team tested their technique in various simulated environments, including warehouse setups, random obstacles, and maze-like settings. Their learning-based approach decongested the warehouse up to four times faster than non-learning-based methods, even with the additional computational overhead of running the neural network.
In the future, the researchers aim to derive simple, rule-based insights from the neural model to make implementation and maintenance easier in actual warehouse settings. This research was supported by Amazon and the MIT Amazon Science Hub.
