Exploring the Effects of Pedestrian Models on Crowd Navigation Training

Murthy, Surya

Permalink

https://hdl.handle.net/2142/124814 Copy

Description

Title

Exploring the Effects of Pedestrian Models on Crowd Navigation Training

Author(s)

Murthy, Surya

Issue Date

2023-05-01

Keyword(s)

Crowd Navigation, Deep RL, Human Behavior Models

Abstract

Autonomous crowd navigation is the process of creating an agent to navigate a crowd of humans without collisions. Current approaches to the crowd navigation problem use deep reinforcement learning (deep RL) to determine the optimal movement for the agent at each time step. These methods are trained using simulations that use artificial human crowds that move using built-in collision avoidance strategies. Deep RL models like Decentralized Structural Recurrent Neural Networks (DSRNN) have applied this approach to successfully navigate human crowds by incorporating spatial and temporal interactions between agents. However, using specific human behavior models during training may impact a deep RL model’s ability to function in environments with humans that operate using a different model. This work explores the effects of different human behavior models on the performance of deep RL models of crowd navigation. In our testing, we trained separate DSRNN models using simulations with different human behavior models. These behavior models included collision avoidance algorithms and pre-trained deep RL models for crowd navigation. We then perform experiments in a simulated environment to identify the benefits and drawbacks of different human behaviors when training DSRNN models. Through these experiments, we find that Deep-RL human behavior models have the potential to increase a crowd navigation agent’s success rate when compared to collision avoidance algorithms. However, there is an inherent increase in computing resources required to train DSRNN agents using a deep RL human behavior model. Further experiments using environments with larger crowds may be necessary to determine the training benefits of deep RL human behavior models.

Type of Resource

text

Language

eng

Owning Collections