Written by Adam Zeve
Inside a giant autonomous warehouse, hundreds of robots roam the aisles, collecting and delivering goods and fulfilling a steady stream of customer orders. In this congested environment, even a small traffic jam or a small collision can snowball into massive speed reductions.
To avoid this avalanche of inefficiencies, researchers at MIT and tech company Symbotic have developed a new way to automatically keep robot swarms running smoothly. Their method learns which robot should go first at each moment based on the traffic situation and adapts to prioritize the robots that are likely to get stuck. In this way, the system can proactively reroute robots to avoid bottlenecks.
This hybrid system uses deep reinforcement learning, a powerful artificial intelligence technique for solving complex problems, to determine which robots should be prioritized. Fast and reliable planning algorithms then provide instructions to the robot, allowing it to quickly respond to ever-changing conditions.
In simulations inspired by the layout of a real e-commerce warehouse, this new approach increased throughput by approximately 25% compared to other methods. Importantly, the system can quickly adapt to new environments with different numbers of robots and warehouse layouts.
“Many decision-making problems exist in manufacturing and logistics because companies rely on algorithms designed by human experts. But we found that with the power of deep reinforcement learning, we can achieve superhuman performance. This is a very promising approach, because even a 2-3% increase in throughput can have a big impact in these huge warehouses,” said MIT Institute for Information and Decision Systems (LIDS). said Han Zheng, a graduate student and lead author of a paper on this new approach.
Zheng’s paper was also joined by LIDS postdoctoral researcher Yening Ma. Brandon Araki and Jingkai Chen of Symbotic. and lead author Cathy Wu, Class of 1954 Career Development Associate Professor in Civil and Environmental Engineering (CEE) and the Institute for Data, Systems and Society (IDSS) at MIT, and a member of LIDS. This research today Artificial Intelligence Research Journal.
route change robot
Coordinating hundreds of robots simultaneously in an e-commerce warehouse is no easy task.
The problem is particularly complex because warehouses are dynamic environments, and robots continue to receive new tasks even after completing their goals. You must change direction quickly when entering and exiting the warehouse floor.
Companies often leverage algorithms created by human experts to determine when and where robots should move to maximize the number of packages they can handle.
However, when congestion or collisions occur, companies may be forced to shut down the entire warehouse for hours to manually resolve the issue.
“In this situation, we cannot accurately predict the future. All we know is what will happen in the future in terms of incoming packages and distribution of future orders. Planning systems need to adapt to these changes as warehouse operations progress,” says Zheng.
MIT researchers achieved this adaptability using machine learning. They started by designing a neural network model to observe the warehouse environment and prioritize robots. They use deep reinforcement learning to train this model. This is a trial-and-error method in which the model learns how to control the robot in a simulation that mimics a real warehouse. The model is rewarded for making decisions that improve overall throughput while avoiding conflicts.
Over time, the neural network learns how to efficiently coordinate many robots.
“By interacting with a simulation inspired by the layout of a real warehouse, our system can receive feedback and use it to make decisions more intelligently. The trained neural network can adapt to warehouses with different layouts,” Zheng explains.
It is designed to understand the long-term constraints and obstacles in each robot’s path, while also taking into account the dynamic interactions between robots as they move through the warehouse.
The model predicts current and future robot interactions to plan to avoid congestion before it occurs.
After the neural network determines which robots to prioritize, the system employs proven planning algorithms to instruct each robot how to move from one point to another. This efficient algorithm allows the robot to quickly respond to changing warehouse environments.
This combination of methods is important.
“This hybrid approach is based on my group’s research on how to achieve the best of both machine learning and traditional optimization techniques. Pure machine learning techniques still struggle to solve complex optimization problems, and it takes a lot of time and effort for human experts to design effective techniques. However, when used properly, expert-designed techniques can greatly simplify machine learning tasks,” said Wu.
overcome complexity
After the researchers trained the neural network, they tested the system in a different simulated warehouse than the one they saw during training. Industrial simulations are too inefficient for this complex problem, so the researchers designed a unique environment that mimics what happens in a real warehouse.
The hybrid learning-based approach achieved, on average, 25% higher throughput than traditional algorithms and random search methods in terms of the number of packages delivered per robot. Their approach can also generate feasible robot path plans that overcome congestion caused by traditional methods.
“As the density of robots increases, especially in warehouses, the complexity increases exponentially and these traditional methods quickly start to break down. Our method is much more efficient in this environment,” Zheng says.
Although their systems are still far from real-world implementation, these demonstrations highlight the feasibility and benefits of using machine learning-based approaches in warehouse automation.
The researchers hope to include task assignment in the problem formulation in the future, since determining which robot completes each task affects crowding. We also plan to expand the system to a larger warehouse with thousands of robots.
Massachusetts Institute of Technology News
