MIT researchers have developed a deep-learning model to improve the efficiency of warehouse robots. The team used a neural network architecture to encode features including the robots’ paths, tasks, and obstacles in the warehouse. This enabled the model to predict where congestion was most likely to occur and take measures to counteract it.
The groundbreaking method works by splitting the robots into smaller groups, which makes managing and decongesting them more manageable. When compared to a robust random search approach, this method improved decongestion rates fourfold.
Beyond making warehouse operations more streamlined, this deep learning approach could also be applied to other intricate planning tasks, such as designing computer chips or laying out pipes in large buildings.
The system drew real-time operational parallels from busy city traffic to the busy life inside large warehouses, with hundreds of robots needing to navigate efficiently without colliding with each other.
According to Cathy Wu, the Gilbert W. Winslow Career Development Assistant Professor in Civil and Environmental Engineering (CEE), their neural network architecture can efficiently code hundreds of robots and their trajectories, sources, destinations, and interactions with other robots.
From a bird’s eye view, a robotic e-commerce warehouse resembles a fast-paced game of “Tetris”. Hundreds of robots simultaneously fetching and delivering items can lead to paths crossing and potential collisions. The MIT team made time a crucial factor during replanning, utilizing machine learning to focus on the most congested areas in the warehouse.
The researchers outperformed robust non-learning-based approaches in multiple simulated environments, including warehouses and maze-like building interiors, managing to decongest the warehouse up to four times faster.
Their future plans include deriving simple rules from their neural model to build rule-based methods, which can be easily implemented and maintained in real warehouse settings. This research was supported by Amazon and the MIT Amazon Science Hub.
