Solving Multi-Agent Pathfinding for Mortgage Systems | AI Magazine
Table of Contents
- Introduction
- Automated Warehouse Systems
- Multi-Agents Coordination
- Multi-Agents Pathfinding (MAPATH)
- Conflict-Based Search (CBS)
- New In-Approximability Results for MAPATH
- Optimality Guarantees for New MAP Variant
- Multi-Agent Navigation Tasks
- Software Development for New RIBS 2020 Flatland Challenge
- Plan Generation and Execution Framework for Real Robots
- Industrial Simulator for Sortation Centers
- Ongoing Research on Learning-based MAP Algorithms
Multi-Agent Planning for Large-Scale Mortgage Systems
Large-Scale mortgage systems have become a significant area of research focus in recent years, with an increasing number of logistic companies such as Amazon and Alibaba constructing automated fulfillment and sortation centers for online orders. These centers require hundreds and even thousands of robots to navigate safely through the network readers and basic traffic of warehouses.
To tackle the fundamental issues in multi-agent planning for large-scale mortgage systems, the concept of intelligent planning has been introduced. At the Core of this approach is solving the multi-agents pathfinding problem (MAPATH), which involves finding collision-free paths for multiple agents on a given graph. The objective of MAPATH is to minimize either the makespan or the flow time.
Automated Warehouse Systems
Automated warehouse systems are already in operation today, with companies such as Amazon and Alibaba using them extensively. These systems require vast numbers of robots to navigate safely through a warehouse to fulfill online orders. The key challenge is to ensure that each robot has a unique task, and none of them collide in the process.
Multi-Agents Coordination
The MAPATH problem can be solved using conflict-based search (CBS), which is a two-level algorithm that first finds individual paths for each agent before resolving any collisions. Existing optimal algorithms for MAPATH are based on CBS, but they are not scalable beyond a certain number of agents.
Multi-Agents Pathfinding (MAPATH)
MAPATH involves finding collision-free paths for multiple agents on a given graph. The makespan or the flow time is minimized to achieve this goal. However, both mixed band and flow time minimization for MAPATH are AP-hard.
Conflict-Based Search (CBS)
The CBS algorithm is used to solve the MAPATH problem. It performs a best-first search on a binary key on a high level to resolve collisions in the computed path from the low-level system. We have made significant scientific contributions to improving the key search of CBS, making it more scalable to larger numbers of agents.
New In-Approximability Results for MAPATH
We have contributed new in-approximability results for MAPATH, making the algorithm more scalable to many more agents. Using the algorithms developed, we can already plan paths for 100 agents optimally within a few minutes.
Optimality Guarantees for New MAP Variant
We have applied our MAPATH techniques to the problem of navigating agents with different shapes, such as drones that share a workspace. We have developed a CBS-based algorithm that keeps the completeness and optimality guarantees for the new map variant.
Multi-Agent Navigation Tasks
Our MAPATH techniques have also been applied to multi-agent navigation tasks for robots and video game characters, where agents try to keep desired information while moving towards their goal locations.
Software Development for New RIBS 2020 Flatland Challenge
We have combined several of our MAP algorithms and techniques to develop software that won the first place in the new RIBS 2020 Flatland challenge. The challenge required solving a MAP problem for multiple chains with uncertain delays whenever the chains attempt to move. Our algorithm based on prioritized planning for the chains and local search improves the solutions.
Plan Generation and Execution Framework for Real Robots
We have developed a plan generation and execution framework to handle kindle dynamic constraints of real robots. We have started a new program that combines target assignment and pathfinding for multiple agents and developed an optimal algorithm for it. We have also developed an algorithm for online tasks in a setting where one can use a total planning time of fewer than seconds for the operation of more than 30 minutes, which involves 250 agents and 2000 tasks. Here, the agents need to pick up and deliver shelves in the simulated warehouse.
Industrial Simulator for Sortation Centers
We tested our lifelong algorithm in an industrial simulator using real-world data for sortation centers where the objective is to optimize the idle time of stations. Our algorithm can compute good solutions for 350 agents in a two-Second runtime limit.
Ongoing Research on Learning-based MAP Algorithms
Our ongoing research includes developing learning-based MAP algorithms and a better theoretical understanding of lifelong planning. We also plan to study more complex joint task assignment and task-finding problems.
Highlights
- Multi-agent planning is crucial for large-scale mortgage systems.
- Intelligent planning involves solving the multi-agent pathfinding problem (MAPATH).
- Conflict-based search (CBS) is a two-level algorithm that resolves collisions between the computed paths.
- We have made significant scientific contributions to improving the scalability of CBS and developing new in-approximability results for MAPATH.
- We have applied MAPATH techniques to multi-agent navigation tasks and developed a plan generation and execution framework for real robots.
- Our lifelong algorithm can optimize the idle time of stations using real-world data for sortation centers.
- Our ongoing research includes developing learning-based MAP algorithms and studying joint task assignment and task-finding problems.
FAQ
Q: What is the MAPATH problem?
A: The MAPATH problem involves finding collision-free paths for multiple agents on a given graph, with the goal of minimizing either the makespan or the flow time.
Q: What is CBS?
A: CBS is a two-level algorithm that performs a best-first search on a binary key to resolve any collisions in the computed paths from the low-level system.
Q: What is the RIBS 2020 Flatland challenge?
A: The RIBS 2020 Flatland challenge required solving a MAP problem for multiple chains with uncertain delays whenever the chains attempt to move.
Q: What is the objective of the lifelong algorithm?
A: The objective of the lifelong algorithm is to optimize the idle time of stations using real-world data for sortation centers.
Q: What are the ongoing research areas in MAP planning?
A: Our ongoing research includes developing learning-based MAP algorithms and studying joint task assignment and task-finding problems.