A large warehouse is a logistical headache. You need to be able to find everything quickly, transfer it to where it’s needed, and avoid unnecessary jams and delays. Increasingly, warehouse work is becoming robotized. Big warehouses can have hundreds of robots navigating the aisles, each with its own goal.

Coordinating this many robots in real-time is no easy task, and every single crash can have devastating cascading consequences.

That’s where a new AI comes in.

“We devised a new neural network architecture that is actually suitable for real-time operations at the scale and complexity of these warehouses. It can encode hundreds of robots in terms of their trajectories, origins, destinations, and relationships with other robots, and it can do this in an efficient manner that reuses computation across groups of robots,” says Cathy Wu from MIT, senior author of the study.

Traffic in a warehouse

The idea was to start with city car traffic. Wu and lead author Zhongxia Yan analyzed robot activity in a warehouse and realized that it’s a lot like cars trying to move around a crowded city center. So they applied principles from urban driving.

Specifically, they split the robots into smaller groups that can be handled more efficiently. Instead of trying to manage 800 robots, robots were split into groups of around 40 and managed as groups instead. Here’s why that matters.

Every time an order comes into the warehouse, a robot moves to the requested area to grab the requested item. It then delivers it to another human or robot that packages the item. But with hundreds of robots doing this at the same time, the risk of collision is huge.

Traditionally, one robot’s trajectory would be kept the same while the others would change to ensure no collisions happen. But that gets extremely complex very fast. Meanwhile, the new AI starts by encoding information about the robots, paths, and obstacles, and then looks at which paths decongest the traffic inside the robot groups the most.

Furthermore, the traditional approach is also very computationally intensive. The orders for robots are readjusted every 100 ms — that means a robot can receive different instructions 10 times per second. All this needs to be computed and managed so that collisions don’t happen. For every iteration, traditional algorithms constantly consider all the robots and their position and objectives.

AI optimised warehouse

In the new approach, the researchers optimized a warehouse with 800 robots. The AI considers the spatial relationships of the 800 robots just once per decision. Then, the algorithm looks at how a group of 40 robots can be decongested as a whole, without affecting the other 760 robots negatively. By identifying which groups are the most effective groups to decongest, the algorithm decongests the warehouse up to four times faster than strong, non-learning-based approaches. It also does it 3.5 times faster than conventional computing.

“This approach is based on a novel architecture where convolution and attention mechanisms interact effectively and efficiently. Impressively, this leads to being able to take into account the spatiotemporal component of the constructed paths without the need of problem-specific feature engineering. The results are outstanding: Not only is it possible to improve on state-of-the-art large neighborhood search methods in terms of quality of the solution and speed, but the model generalizes to unseen cases wonderfully,” says Andrea Lodi, the Andrew H. and Ann R. Tisch Professor at Cornell Tech, and who was not involved with this research.

The approach was tested in several real-life environments, as well as complex maze-like settings, and was successful at decongesting warehouses much faster.

The only drawback is that sometimes it wasn’t clear why the algorithm was making certain decisions. So, for the future, the team wants to obtain more insights from the model. This will not only help them further improve the algorithm, but also implement rules that may be necessary in actual warehouse settings.

The study was presented at the ICLR conference.