DT-RaDaR: Digital Twin Assisted Robot Navigation using Differential Ray-Tracing

TL;DR

DT-RaDaR is proposed, a robust privacy-preserving, deep reinforcement learning-based framework for robot navigation that leverages RF ray-tracing in both static and dynamic indoor scenarios as well as in smart cities and the feasibility of the proposed framework in indoor environments and smart cities is demonstrated.

Abstract

Autonomous system navigation is a well-researched and evolving field. Recent advancements in improving robot navigation have sparked increased interest among researchers and practitioners, especially in the use of sensing data. However, this heightened focus has also raised significant privacy concerns, particularly for robots that rely on cameras and LiDAR for navigation. Our innovative concept of Radio Frequency (RF) map generation through ray-tracing (RT) within digital twin environments effectively addresses these concerns. In this paper, we propose DT-RaDaR, a robust privacy-preserving, deep reinforcement learning-based framework for robot navigation that leverages RF ray-tracing in both static and dynamic indoor scenarios as well as in smart cities. We introduce a streamlined framework for generating RF digital twins using open-source tools like Blender and NVIDIA's Sionna RT. This approach allows for high-fidelity replication of real-world environments and RF propagation models, optimized for service robot navigation. Several experimental validations and results demonstrate the feasibility of the proposed framework in indoor environments and smart cities, positioning our work as a significant advancement toward the practical implementation of robot navigation using ray-tracing-generated data.

Figures (25)

• Figure 1: Comparison of sensor-based navigation and $\text{DT-RaDaR}$. We show that the $\text{DT-RaDaR}$ is privacy preserving and data efficient.
• Figure 2: Overview of the proposed system architecture. $\text{DT-RaDaR}$ continuously monitors the environment. In Module 1, we generate the digital twins. The re-training is performed over an interval ( Module 2), we update the Module 3 which eventually updates path planning policy by optimizing the navigation ( Module 3). The solid line shows the training path, and dashed line represent inference path.
• Figure 3: Overview of the proposed DQN agent. We formulate the state and reward space to optimize navigation.
• Figure 4: Overview of various used tools for implementing $\text{DT-RaDaR}$.
• Figure 5: The scene maps and coverage maps for both indoor and outdoor environments (details in Sec. \ref{['Sec:proposed_mtwo']}). RW, CM and DT represents real-world, coverage map and digital twin, respectively.

Theorems & Definitions (3)

• Remark 1
• Remark 2
• Remark 3