As automated warehouses become increasingly popular in various industries, ensuring the efficiency and safety of hundreds of robots navigating such spaces is a significant challenge. Notably, even top-performing algorithms struggle to keep pace with the demands of e-commerce or manufacturing. To address this, a group of MIT researchers, who use AI to mitigate traffic congestion, have applied their insights to this issue.
The team has built a deep-learning model that encapsulates crucial warehouse information, like the robots, their intended paths, jobs, and hurdles, and forecasts the optimal areas to clear congestion to boost overall productivity. Their strategy groups the warehouse robots, facilitating swifter decongestion of smaller robot groups using conventional coordination algorithms. This approach decongests the robots nearly four times faster than a robust random search method.
This deep learning strategy could also be used for other complicated planning tasks, like the design of computer chips or routing pipes in large buildings. The researchers’ method involves a new neural network architecture that can encode hundreds of robots efficiently concerning their journeys, origins, destinations, and relationships with other robots.
The setup is similar to playing “Tetris” at a fast pace: hundreds of robots, when an order appears, have to pick the appropriate shelf, retrieve the item, and deliver it to a human operator for packing. The routing process needs to be very speedy, with robots being replanned about every 100 milliseconds. Therefore, machine learning is used to focus replanning on the most congested areas, effectively reducing total travel time for the robots.
The researchers’ approach increased efficiency by segmenting the robots into smaller groups. For instance, a warehouse with 800 robots might have the network divide the warehouse floor into smaller 40-robot groups. The network then predicts which group has the most potential to improve the overall solution if a search-based solver were used. The algorithm repeats this process iteratively.
The neural network can efficiently assess groups of robots as it captures intricate relationships between individual machines. The method reduces redundant computation by encoding constraints only once per iteration, instead of for each subproblem. Their solution decongests warehouses up to four times faster than robust, non-learning methods, and even with the added computational burden of operating the neural network, it still solves the problem 3.5 times faster.
Moving forward, the researchers aim to derive simpler, rule-based insights from their model, which could be easier to implement and maintain in real robotic warehouse settings. The research was funded by Amazon and the MIT Amazon Science Hub.