Optimizing Large-Scale AMR Fleet Operations

Optimizing Large-Scale AMR Fleet Operations

Automotive industry leaders use autonomous mobile robots (AMR) in their production facilities to improve productivity. In this case study, Tesla Material Flow Engineer and former BMW Group PhD Student and AMR researcher, Maximilian Selmair, describes standard industry practice when deploying large-scale transporter fleets and demonstrates how AnyLogic cloud-based simulation helps develop optimal task allocation algorithms.

Why AMR and not AGV?

Autonomous mobile robots are more capable than automated guided vehicles (AGV) because they are more complex. AMR have greater software capabilities and navigate using maps, without the need for guiding wires or strips, so they are not restricted to fixed routes like AGV. As a result, AMR are more flexible in the tasks they can do, and they can be redeployed quickly with only a software update.

Compared to AGV, modern AMR technology is seen as being more cost-effective thanks to reduced infrastructure requirements and quicker deployment that does not cause production interruptions.

Problem: How to efficiently assign tasks to an automated fleet

The optimization of production line transporter operations has two objectives. Primarily, optimization should result in no late tasks. And secondly, the density of traffic in a production facility should be minimized.

Late tasks lead to delays that reduce efficiency and increase costs. For BMW, avoiding production stoppages was the main aim of optimization work related to transporter operations.

In production facilities, the space available for pathways is a limited resource and, as a result, autonomous transporters must share routes with people and other vehicles. The reduction of traffic has several desirable effects, including increased safety, less congestion, and fewer late tasks.

The research aimed to meet the goal of no late tasks with a minimum of transporter driving. A challenge that can broadly be characterized as the assignment problem.

Solution: Simulation to test AMR task assignment methods

To solve the assignment problem, a hybrid simulation model of a vehicle production line facility made it possible to test various methods. AnyLogic simulation software, with its multi-method modeling capabilities and its built-in Material Handling Library, allowed for quick modeling of the workspace, including the addition of automated transporters. The simulation uses both agent-based and discrete event modeling approaches.

Testing autonomous mobile robot task allocation algorithms in an AnyLogic simulation model

Simulation model for testing task allocation algorithms for autonomous mobile robots (click to enlarge). Example cloud model

While AnyLogic includes a variety of methods for assigning tasks to transporters, it also offers the flexibility of including custom code. Using custom code allowed the testing of any assignment algorithm that might best solve the AMR task allocation problem. In testing, both heuristics and exact methods were analyzed. The exact methods are algorithms that always produce one optimal solution, such as with linear optimization. By contrast, the heuristic methods are based on approximation and may not be as accurate as exact methods, but are usually faster.

Methods tested for solving the AMR task assignment problem included the exact methods of Linear Optimization, the Hungarian Method, and the Jonker-Volgenant-Castanon algorithm, and the heuristic methods of Vogel’s Approximation method, VAM-nq, and Greedy-Search

The methods tested when designing a task assignment algorithm for autonomous mobile robots (click to enlarge)

The research results from testing the various methods in an AnyLogic simulation model show that the Jonker-Volgenant-Castanon (JVC) assignment algorithm is superior when assigning tasks to transporters in a vehicle production facility.

Faster AMR scenario analysis with cloud-based simulation

After creating the simulation model to test different assignment methods, it was necessary to conduct many simulation runs. Each assignment method needed testing with parameter variations and in different scenarios.

In this case, 40 different fleet sizes were tested against different assignment methods. And, with each run taking two hours to compute nine hours of simulation, a solution that could speed up the process was welcome.

The AnyLogic Cloud platform is a scalable computing environment that allows parallel run execution. This computing capability allowed for rapid parameter variation and multiple scenario testing. For example, simulation runs for all 40 fleet sizes can be run simultaneously. Also, with the computing taking place on cloud servers, the simulation modeler is free to continue using their computer without hindrance.

Results: An optimal AMR task allocation algorithm

Compared to the baseline scenario that allocates tasks based on the nearest agent, the method developed from the JVC assignment algorithm reduces the number of AMR required by 30%.

In an example scenario of 7,500 tasks, the traditional method required 58 transporters to achieve a late task total of three. For the algorithm that resulted from the testing and research, only 42 transporters were needed to achieve the same number of late tasks.

Such a reduction in the number of AMR transporters required for a desired level of service is significant because of the high upfront cost of AMR, in comparison to AGV and manual methods.

Reducing the number of AMR required to achieve just three late tasks in 7,500 also helps meet the goal of minimizing traffic density.

The combination of simulation, for creating a testing environment, and cloud computing, for running the experiments, allowed the rapid development of a custom assignment algorithm for AMR task allocation in a vehicle production facility. The algorithm improved on standard nearest agent assignment by 30%, achieving the aims of a low late-task rate and reduced traffic density.

Maximilian Selmair presented the research at the AnyLogic Conference 2021:

Learn more about cloud-based simulation on our dedicated AnyLogic Cloud page.

Similar case studies

More case studies

Get a brochure with industry case studies