Deep Compressed Communication and Application in Multi-Robot 2D-Lidar SLAM: An Intelligent Huffman Algorithm

Liang Zhang, Jinghui Deng2024 · DOI

TL;DR

This paper introduces a novel deep learning-based compression framework for multi-robot 2D-Lidar SLAM. By compressing full occupancy grid maps using a lightweight CNN and Huffman/RLE encoding, they achieve a 99% reduction in communication overhead while maintaining high map quality, enabling efficient collaborative mapping in bandwidth-constrained environments.

What Problem Does This Paper Address?

Multi-robot Simultaneous Localization and Mapping (SLAM) systems are crucial for exploring large or hazardous environments, offering advantages over single robots in terms of speed, accuracy, and robustness. However, as the number of robots and the size of the environment increase, the communication bandwidth required to share map data becomes a significant bottleneck. Traditional approaches often involve transmitting partial map information or feature descriptors, which can lead to loss of detail or accumulation of errors. This paper tackles the challenge of enabling efficient, full-map communication in multi-robot SLAM systems without sacrificing map integrity, particularly for resource-constrained platforms and limited communication channels.

Methods

The proposed solution is a compressed communication framework with three main components:

  1. Full-Map Creation: Robots use the Gmapping algorithm to build individual 2D occupancy grid maps of their environment.
  2. Full-Map Compression and Transmission:
    • Lightweight CNN Feature Extraction: A Convolutional Neural Network (CNN) (referred to as CNN_feature) is used to extract essential structural information from the occupancy grid map, creating a compressed m_feature representation. This CNN has three layers designed for feature extraction and downsampling.
  3. Intelligent Encoding: The m_feature is then compressed using a combination of Huffman coding and Run-Length Encoding (RLE). Huffman coding is applied first to assign variable-length codes based on the probability of feature values, followed by RLE to further compress consecutive identical characters in the resulting hexadecimal code. This process generates a code table (T) and a compressed string (e).
  4. Map Recovery CNN: A corresponding CNN (CNN_recovery) is used by other robots to decompress the received string (e) and reconstruct a near-original map. This CNN employs residual blocks to help learn the differences between the input and output, aiding in accurate map recovery.
  5. Transmission: The compressed string is transmitted over Ultra-Wide Band (UWB) communication modules, segmented for efficient transfer.
  6. Full-Map Fusion: Robots can fuse their locally created maps with the recovered maps from other robots. The paper considers two scenarios for fusion: when initial robot positions are known (direct transformation) and when they are unknown (using feature matching and RANSAC).

The entire network's optimization is achieved by minimizing the mean square error between the network's output map and the original full map using backpropagation and the Adam optimizer.

Datasets

  • Training Data: 2,800 occupancy grid maps were used for training the deep learning models. These maps were generated using the Gmapping algorithm in various environments.
  • Evaluation Data: The experiments were conducted using a two-robot system in a hallway environment. The performance was evaluated by comparing maps generated with and without the proposed compression method, and by measuring localization errors.

Key Insights and Ideas

  • Full-Map Compression is Feasible and Beneficial: Contrary to the common practice of transmitting only partial map data, this work demonstrates that compressing the entire occupancy grid map is achievable and highly effective, leading to significant bandwidth reduction without substantial loss of map quality.
  • Synergy of Deep Learning and Traditional Compression: The combination of a lightweight CNN for feature extraction/recovery and classical Huffman/RLE encoding proves to be a powerful approach. The CNN captures the semantic structure, while Huffman/RLE efficiently encodes the resulting data.
  • Lightweight Architecture for Resource Constraints: The use of lightweight CNNs is crucial for deployment on resource-constrained robotic platforms, allowing for real-time processing of map compression and decompression.
  • Minimal Impact on Localization: Despite the aggressive compression, the recovered maps retain sufficient detail for accurate robot localization, with minimal degradation compared to using uncompressed maps.

Key Results

  • Communication Overhead Reduction: The proposed method achieves a communication overhead reduction of up to 99%.
  • Map Quality Preservation:
    • The recovered maps maintain high quality, with average PSNR values of 50.649 dB and SSIM values of 0.975. These values indicate that the recovered maps are very close to the original maps.
  • Localization Accuracy: The mean localization errors (x, y, yaw) were 3.59 cm, 3.43 cm, and 0.72°, respectively, showing that the subtle changes from compression have little effect on localization performance.
  • Transmission Time and Bandwidth: The compressed communication significantly reduces the required bandwidth, allowing it to be satisfied by UWB hardware limitations. The transmission time for compressed maps is also reduced.

Limitations and Future Work

  • Limitations:
    • The paper focuses on 2D occupancy grid maps. While effective, this representation might not be sufficient for all applications that require 3D information.
  • The performance of the CNN recovery is dependent on the quality of the compressed m_feature.
  • The current implementation is validated with a two-robot system; scaling to a larger number of robots might introduce new challenges.
  • Future Work:
    • Extending the framework to 3D SLAM systems.
  • Investigating more advanced deep learning architectures for even better compression and recovery.
  • Exploring different communication protocols and hardware to further optimize transmission.
  • Testing the framework in more complex and dynamic environments.

Summary updated

BibTeX
@misc{zhang2024deep,
  author = {Liang Zhang and Jinghui Deng},
  title = {{Deep Compressed Communication and Application in Multi-Robot 2D-Lidar SLAM: An Intelligent Huffman Algorithm}},
  year = {2024},
  doi = {10.3390/s24103154},
}

Loading related content...

Figure