MIT Tech Takes On Warehouse Robot Traffic Jams
Though the increasing number of warehouse robots often optimize efficiency, they can also experience a problem that some consider uniquely human: traffic jams.
A group of engineers at the Massachusetts Institute of Technology (MIT) recently developed technology that could change that.
More from Sourcing Journal
Cathy Wu, an assistant professor at MIT’s Laboratory for Information and Decision Systems (LIDS), worked with graduate student Zhongxia Yan to develop neural network architecture that considers warehouse robots in terms of smaller groups, rather than accounting for the entire fleet simultaneously.
For example, if a warehouse had a fleet of 800 robots, they may be segmented into cohorts of 40. The size of the group may change according to the number of total robots a warehouse has.
Wu said the technology pinpoints certain regions of the warehouse facing issues, then makes changes only to the applicable subsets of robots, one group at a time.
“You can think of what we’re doing [as] holding the other robots fixed so that they’re not going to move,” she said. “For the sake of optimization, it’s much harder to work with 800 robots than to work with 40.”
The researchers’ deep-learning model takes in information about the warehouse, as well as the robots’ planned paths, typical tasks and run-of-the-mill obstacles. Based on that input, it has the intelligent capability to predict which areas of the warehouse need to be decongested to ameliorate efficiency.
That model differs from current technology, which does not have a learning component. That means that although the system can choose a subset of robots to optimize, it’s effectively guessing at which robots to select, since the system cannot identify the locale of the problem, nor the robots that directly cause it, Wu explained.
“For a large warehouse—let’s say 800 robots—choosing a random set of 40, that could be a part of the warehouse that’s already fairly uncongested. Maybe there are some bottlenecks in the warehouse…but it’s hard to identify that automatically,” Wu told Sourcing Journal.
Alternatively, the neural network her team has been testing can determine the root of the problem.
“We’re doing some more automatic learning of those features to actually identify which parts of the warehouse to focus the computation on, so by using this neural network, what we’re doing is identifying not just a single set of, say, 40 robots. We’re going to consider, say, 100 different subsets, and we’re going to use this fast neural network to predict which location is more congested.”
The researchers can run that process repeatedly until they alleviate the majority of the congestion in the warehouse.
“We want to iteratively decongest the warehouse until we’re happy with the solution,” Wu said.
The researchers tested the technique in simulated environments that contained obstacles and “maze-like settings that emulate building interiors.” Though they did not test their operations in actual warehouses, they did set some of the simulations up to mirror them.
Wu said the technology will be most easily applicable to warehouses with homogenous fleets, though she and other researchers may in the future consider how different types of robots could operate together on one neural network.
Wu said real warehouse tests will come soon, and noted that the technology could be available for use in commercial warehouses within the next three years. The neural network could have practical applications for product-picking robots in e-commerce warehouses, where robots sometimes run rampant.
Before that, the team will need to iron out some potential issues, like robot failure and latency in communication between sensors, algorithms and robots.
“We assume perfect information right now, so [for example], full observation of the warehouse. In practice, that information has to come from sensors, and those sensors will have delays or imperfections,” Wu said. “Also, the robots will fail from time to time, and how to handle those failures is going to be another interesting obstacle.”
Amazon supported the researchers’ work; the e-commerce giant has an ongoing partnership with MIT to better understand how automation affects jobs.
Amazon did not immediately return Sourcing Journal’s request for comment.
