In the growing field of warehouse automation, managing hundreds of robots zipping through a large warehouse is a logistical challenge. Delivery paths, potential collisions and congestion all pose significant issues, making the task a complex problem that even the best algorithms find hard to manage. To solve this, a team of MIT researchers has developed a deep-learning model designed to improve overall efficiency by predicting ideal areas for congestion reduction.
The new system sees warehouse robots divided into smaller groups, allowing for faster, more efficient decongestion using traditional robot coordination methods. This approach allows for advanced streamlining within warehouse operations, potentially providing similar benefits for complex planning tasks in other fields, such as computer chip design or extensive pipe routing.
According to the researchers, the new system can efficiently encode large amounts of information, tracking the movements and tasks of hundreds of robots simultaneously. The warehouse environment operates online, with a continual, rapid cycle of robot replanning to avoid potential collisions. As such, the congestion algorithm had to be designed to quickly identify the most action-needed areas to ensure efficient operations.
The researchers’ system divides large numbers of robots into smaller groups. Once split, the algorithm identifies the groups with the most potential to enhance the overall solution if applying a trajectory coordination approach.
One of the ways the new neural network improves efficiency is by capturing the complex relationships between individual robots. For example, the system takes into account the possibility that paths may cross during their journeys. By streamlining computation and encoding constraints just once, repetition for each subproblem can be avoided.
Tests showed the learning-based approach decongested warehouses up to four times faster than non-learning-based techniques. Further research is planned to simplify the network’s decisions and extract rule-based insights from the neural model. Simpler rule-based methods could potentially be easier to develop and maintain in real-world robot warehouse settings.
The development of the new system is a significant step forward in warehouse automation, potentially leading to faster, more efficient operations within the industry. The researchers’ deep-learning approach is not only impressive in terms of its ability to improve existing large neighborhood search methods in regards to quality and speed, but it also generalizes wonderfully to unseen cases. The work was supported by Amazon and the MIT Amazon Science Hub.
