In an enormous robotic warehouse, hundreds of robots zip back and forth, picking up items and delivering them to human workers for packing and shipping. This is becoming an increasingly common scene in various industries, from e-commerce to automotive manufacturing. However, managing these large numbers of robots, ensuring they reach their destinations effectively, and avoiding collisions amongst them is a complex task. Even advanced path-finding algorithms stumble in coping with the frenetic pace of e-commerce or manufacturing industries.
A group of researchers from MIT has decided to tackle this issue using AI. Taking inspiration from their experience using AI to address traffic congestion in cities, they developed a deep-learning model specifically for warehouse environments. This model is able to encode valuable data about the warehouse –including information about the robots, their intended routes, tasks, and any obstacles– and use this data to calculate the best areas to address congestion, thereby improving the overall efficiency of operations.
The researchers’ approach involves dividing the warehouse robots into smaller groups, which can be decongested more quickly with traditional robot-coordinating algorithms. The result is that their method can decongest the robots almost four times faster than a forced random search method. Moreover, this deep learning approach could be applied to other complex scheduling tasks, such as computer chip design or pipe routing in large buildings.
The researchers built a neural network capable of encoding vast amounts of data such as robot trajectories, origins, destinations, and relations with other robots, while efficiently reusing computations across robot groups. The network subdivides the warehouse robots into smaller units, predicting which group offers the best potential for improving the overall efficiency if a search-based solver was employed to coordinate their routes.
Significantly, the network is capable of formulating complex relationships that exist within individual robots in each group. That is to say, even if two robots start far apart from each other, the network can predict if their paths will intersect. Once it identifies the most congested group, this algorithm selects the next most congested group and so on.
Furthermore, the model only needs to enter constraints once, rather than repeating this process for each subgroup. This streamlining of computations significantly improves the model’s overall efficiency. The effectiveness of this approach was tested in various simulated environments, including those that replicated warehouses, random obstacles, and even maze-like settings that resemble building interiors.
The results were impressive, showing that the model could decongest the warehouse up to four times faster than strong, non-learning-based methods. The researchers hope in the future to be able to derive straightforward, rule-based insights from their neural model to make it more interpretable and easier to implement in actual warehouse settings.
The research, supported by Amazon and the MIT Amazon Science Hub, will be presented at the International Conference on Learning Representations. Andrea Lodi, a professor at Cornell Tech who was not involved in the research, praised the model for its efficiency, ability to handle the spatial-temporal component of constructed paths without problem-specific feature engineering, and success in improving both the quality and speed of large neighborhood search solutions.
