New findings indicate that teaching robots to form their own teams and patiently wait for their companions can lead to faster task execution, showcasing a promising avenue for advancements in manufacturing, agriculture, and warehouse automation.
Recent research conducted by the University of Massachusetts Amherst reveals that when robots are programmed to assemble their own teams and willingly wait for their peers, they can complete tasks more quickly. This study was recognized as a finalist for the Best Paper Award on Multi-Robot Systems at the IEEE International Conference on Robotics and Automation 2024.
According to Hao Zhang, an associate professor at UMass Amherst’s Manning College of Information and Computer Sciences and director of the Human-Centered Robotics Lab, “There has been a longstanding debate about whether we should develop a single, powerful humanoid robot capable of handling all tasks, or utilize a collaborative team of robots.”
In manufacturing, employing a robot team can often be more cost-effective since it optimizes the strengths of each robot. The primary challenge lies in coordinating a varied group of robots—some may be stationary, while others are mobile; some can handle heavy items, and others excel at lighter tasks.
To address this, Zhang and his team introduced a novel scheduling approach for robots termed learning for voluntary waiting and subteaming (LVWS).
“Robots face substantial tasks similar to humans,” Zhang explains. “Take, for instance, a large box that’s too heavy for one robot to lift on its own. In such cases, multiple robots must collaborate.”
The concept of voluntary waiting is also crucial here. “We want the robots to have the ability to wait actively. If they always rush to complete smaller, immediate tasks, they may never tackle the larger job,” Zhang elaborates.
In their experiments with the LVWS strategy, the researchers assigned six robots a total of 18 tasks within a computer simulation, comparing their LVWS method to four alternative techniques. The simulation contained a well-defined, optimal solution for the fastest way to accomplish the tasks. The team assessed how each method performed relative to this ideal solution, a metric referred to as suboptimality.
The evaluation found that the competing methods ranged in suboptimality from 11.8% to 23%. In contrast, the innovative LVWS strategy only showed 0.8% suboptimality. “Our solution is very close to the theoretical optimal,” stated Williard Jose, a co-author of the paper and doctoral student in computer science at the Human-Centered Robotics Lab.
How does having a robot wait improve the entire team’s efficiency? Imagine this scenario: you have three robots—two capable of lifting four pounds each, and one that can lift ten pounds. If one of the lighter robots is occupied and there’s a seven-pound box to move,
“it might be more advantageous for the larger robot to wait for the second lighter robot, allowing both to move the heavy box together. This way, the larger robot can focus its efforts on a different, more suitable task,” José explains.
If it’s feasible to identify an optimal solution, what’s the need for a scheduling system? “The challenge with deriving that precise solution is that it’s time-consuming,” Jose clarifies. “As the number of robots and tasks increases, it becomes exponentially more complex. Finding the exact optimal solution in a reasonable timeframe is not feasible.”
In scenarios involving 100 tasks, where deriving an exact solution is impractical, their approach managed to complete the tasks in 22 timesteps, compared to 23.05 to 25.85 timesteps for the other methods.
Zhang aspires that this research will propel the evolution of automated robot teams, particularly as we contemplate scaling. For example, a lone humanoid robot may be ideal for a compact household, while multi-robot systems can excel in larger industrial settings that require specialized capabilities.
This research received funding from the DARPA Director’s Fellowship and a CAREER Award from the U.S. National Science Foundation.
