Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Task-Oriented Wireless Transmission of 3D Point Clouds: Geometric Versus Semantic Robustness

Vukan Ninkovic, Tamara Sobot, Vladimir Vincan, Gorana Gojic, Dragisa Miskovic, Dejan Vukobratovic

Abstract

Wireless transmission of high-dimensional 3D point clouds (PCs) is increasingly required in industrial collaborative robotics systems. Conventional compression methods prioritize geometric fidelity, although many practical applications ultimately depend on reliable task-level inference rather than exact coordinate reconstruction. In this paper, we propose an end-to-end semantic communication framework for wireless 3D PC transmission and conduct a systematic study of the relationship between geometric reconstruction fidelity and semantic robustness under channel impairments. The proposed architecture jointly supports geometric recovery and object classification from a shared transmitted representation, enabling direct comparison between coordinate-level and task-level sensitivity to noise. Experimental evaluation on a real industrial dataset reveals a pronounced asymmetry: semantic inference remains stable across a broad signal-to-noise ratio (SNR) range even when geometric reconstruction quality degrades significantly. These results demonstrate that reliable task execution does not require high-fidelity geometric recovery and provide design insights for task-oriented wireless perception systems in bandwidth- and power-constrained industrial environments.

Task-Oriented Wireless Transmission of 3D Point Clouds: Geometric Versus Semantic Robustness

Abstract

Wireless transmission of high-dimensional 3D point clouds (PCs) is increasingly required in industrial collaborative robotics systems. Conventional compression methods prioritize geometric fidelity, although many practical applications ultimately depend on reliable task-level inference rather than exact coordinate reconstruction. In this paper, we propose an end-to-end semantic communication framework for wireless 3D PC transmission and conduct a systematic study of the relationship between geometric reconstruction fidelity and semantic robustness under channel impairments. The proposed architecture jointly supports geometric recovery and object classification from a shared transmitted representation, enabling direct comparison between coordinate-level and task-level sensitivity to noise. Experimental evaluation on a real industrial dataset reveals a pronounced asymmetry: semantic inference remains stable across a broad signal-to-noise ratio (SNR) range even when geometric reconstruction quality degrades significantly. These results demonstrate that reliable task execution does not require high-fidelity geometric recovery and provide design insights for task-oriented wireless perception systems in bandwidth- and power-constrained industrial environments.
Paper Structure (23 sections, 5 equations, 5 figures)

This paper contains 23 sections, 5 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Point Cloud Representation and Processing
  4. Conventional Point Cloud Compression
  5. Learning-Based Point Cloud Processing
  6. Semantic and Task-Oriented Communications
  7. System Model and Proposed Architecture
  8. End-to-End Semantic Transmission Model
  9. Transmitter Architecture & Channel Model
  10. PointNet++ Encoder
  11. Wireless Channel Model
  12. Parallel Semantic and Geometric Decoders
  13. Geometric Decoder
  14. Semantic Decoder
  15. Performance Evaluation
  16. ...and 8 more sections

Figures (5)

  • Figure 1: End-to-end semantic PC transmission framework with parallel geometric reconstruction and semantic classification. The background removal step (red rectangle) is optional and evaluated separately in Section IV.
  • Figure 2: Industrial collaborative robotics data acquisition setup.
  • Figure 3: Representative dataset samples: a) Point cloud of an L-shaped solid object; b) Background removal prior to normalization and sampling.
  • Figure 4: Geometric reconstruction performance vs. SNR for different latent dimensions, sampling strategies, and background removal configurations.
  • Figure 5: Semantic classification accuracy vs. SNR for binary clamp detection and six-class industrial object recognition ($\lambda_{\mathrm{cls}}=1$ in Eq. \ref{['total_loss']} in all experiments).