Skip to content Skip to footer

MIT researchers have developed a deep-learning model to help robots navigate crowded warehouses, where congestion can slow operations and even lead to crashes. The model does this by dividing the robots into smaller groups and using a path-finding algorithm to decongest each group more quickly. Researchers described the process as being similar to mitigating traffic in a busy city. The model was successful in decongesting robots nearly four times faster than random search methods. This deep learning approach can also be applied to other complex planning tasks, such as computer chip design or routing pipelines in large buildings.

The solution was developed by Professor Cathy Wu and graduate student Zhongxia Yan. Wu explains that their approach has been designed to effectively manage the scale and complexity present in large warehouses. It encodes hundreds of robots in real-time, taking into account their trajectories, origins, and destinations. The method works by focusing on the most congested regions of the warehouse and determining which sections have the greatest potential for overall improvement if traditional algorithms are used to coordinate their movements.

The MIT team developed an architecture that uses machine learning to deal with smaller groups of robots at a time. For instance, in an 800-robot warehouse, the system may break down the area into smaller groups of about 40 robots. The algorithm then predicts which group has the highest potential to improve efficiency if a search-based solver is used to coordinate the robots’ trajectories in that group. This process repeats until all robots have been coordinated.

The approach helps streamline the computation process by encoding constraints only once for each group, as opposed to repeating the process for each robot. Upon testing, the research team found that this learning-based approach was four times faster at decongesting the warehouse than non-learning methods. Even accounting for the extra computational overhead of this deep learning approach, the system still solved the problem 3.5 times faster. In future developments, the researchers aim to translate the insights from their neural model into simpler, rule-based methods that can be more easily applied and maintained in real-world warehouse settings. The research was funded by Amazon and the MIT Amazon Science Hub.

Leave a comment

0.0/5