In order to improve efficiency in large-scale robotic warehouses, a team of researchers from the Massachusetts Institute of Technology (MIT) have developed a deep learning model which assists in navigating robots to decongest warehouse floors. The way this model works is by splitting the hundreds of robots into smaller, manageable groups which are easier for traditional algorithms to coordinate.
The warehouse logistics problem, similar to the intricacies of city traffic management, lies in the fundamental challenge of keeping hundreds of robots from colliding with each other, as they zip back and forth in the warehouse grabbing items for human workers to pack and ship. The complexity is such that even the most advanced path-finding algorithms find it difficult to maintain efficiency, a problem which the researchers from MIT managed to address.
The team developed a neural network architecture that encoded the necessary information about the warehouse, the robots, their planned paths, tasks and obstacles. This model then helped to predict which areas of the warehouse could be decongested to enhance the overall efficiency of operations.
An important factor of this model is its ability to comprehend the relationships between individual robots. For instance, two robots that are initially positioned far apart may still have intersecting paths during their trips. Another key feature that sets this model apart is its capacity to encode constraints only once, eliminating the need for the system to repeat the encoding process for each sub-problem.
This system was tested on simulated environments and was found to decongest the warehouse up to four times faster than existing approaches. The key to this approach was the interaction between convolution and attention mechanisms within the model, allowing for the management of constructed paths without the need for problem-specific feature engineering.
Despite potential difficulties in interpretation from this neural model, the researchers aim to derive simple, rule-based methods that can be implemented and maintained in actual warehouse settings. This research was supported by Amazon and the MIT Amazon Science Hub.
