In crowded robotic warehouses, managing hundreds of robots to navigate efficiently and avoid collisions is a complex challenge. This has prompted a group of MIT researchers to create a deep-learning model to tackle the problem. Using AI, the team has developed an innovative model that encapsulates details about the warehouse environment – the robots, planned routes, tasks, and hurdles and employs this to predict the best areas of the warehouse to manage robot traffic to enhance overall productivity.
This model allows the robots to be broken down into smaller groups, enabling easier and more rapid coordination through existing algorithms. The result is a system that manages traffic four times faster than other robust random search methods. Besides optimising warehouse functions, the potential for this deep learning solution could spill over into other intricate planning tasks, such as computer chip design or pipe routing in large structures.
The model is particularly efficient as it allows for real-time operations. It is capable of encoding hundreds of robots, along with their trajectories, origins, destinations and interactions with other robots in an efficient manner. The algorithms for managing these robots are agile and must be very quick, with robots being replanned approximately every 100 milliseconds.
Maintaining speed is therefore crucial during the planning process. To keep pace with the requirement of managing each robot ten times every second, MIT’s solution employs machine learning to focus the planning on the most critical areas of congestion, reducing total travel time for the robots.
The team assembled a neural network that works with smaller groups of robots simultaneously. For example, in a warehouse populated by 800 robots, the network might split the floor into smaller sections of 40 robots each. It then predicts which group offers the most potential to improve overall efficiency if a search-based solution were to coordinate the trajectories of the robots in that group.
The researchers tested their solution in several simulated environments, mirroring warehouses, mazes, and areas with random obstacles. Results showed the learning-based approach was up to four times more efficient than the non-learning based approaches. Even when the additional computational cost of running the neural network was factored in, results were still 3.5 times faster.
The ultimate goal of the researchers is to draw simple rule-based insights from the model to make it easier to interpret and implement, as well as maintain in actual warehouse environments. The highlighting strength of this approach is its novel architecture where convolution and attention mechanisms interact effectively and smoothly.
The research was sponsored by Amazon and the MIT Amazon Science Hub.
