Panasonic Hd Develops Ai Technology For Efficient Path Planning For Autonomous Robots In Environments Where Various Robots Coexist With Humans

Figure 1. Overview of task setting for in-building robot delivery (cited from the published paper [1]). Delivery robots (blue line) deliver products from the store to customers in conference rooms in the building, while a remote operator monitors multiple robots. The robots proceed while avoiding pedestrians and cleaning robots (orange line) that are directed by other systems.

In such an environment, autonomous robots must slow down or pause to avoid moving obstacles. Consequently, there is a stochastic delay in the time for the robot to traverse the corridor. And in many cases, the probability distribution of this delay is unknown beforehand, or only limited prior information is available. Therefore, it is necessary to develop a method to create the next plan based on the delay observed while the robot operates.

MAPF (Multi-agent Path Finding) has been studied and applied to the problem of creating collision-free paths for multiple robots on a grid-like graph. Still, conventional methods assume that the probability distribution of delays can be accurately modeled. Applying these methods to real-world situations where the number of people and robots in the corridor fluctuates is difficult.

Therefore, we defined a type of MAPF on non-grid like graph, where the travel times of robots are stochastically distributed, and the true distribution is unknown. We proposed a method to update the parameters of the probability distribution using Bayesian inference.
We modeled the delay of a robot running a particular path (edge in the graph) with a gamma distribution with unknown parameters. The parameters of the delay probability distribution were inferred using observed values of delays obtained during plan execution. By creating an optimal route plan based on the estimated parameters, we achieved efficient route planning even in an environment that is highly difficult for autonomous mobile robots to traverse. We demonstrated the effectiveness of this method through comparative experiments with existing methods. (Figure 2)