A New Implementation of NeoSLAM and a Comparative Evaluation with RatSLAM

Technical Summary: A New Implementation of NeoSLAM and Comparative Evaluation with RatSLAM

Problem Statement
Autonomous mobile robotics faces significant challenges in unknown, dynamic, and GPS-denied environments, such as subterranean, forest, or underwater settings. In these scenarios, robots must perform Simultaneous Localization and Mapping (SLAM) without prior maps while coping with perceptual aliasing, sensor noise, and environmental changes. While biologically inspired approaches like RatSLAM and the more recent neuroscience-based NeoSLAM offer scalable topological representations and improved tolerance to ambiguity, their practical application is often hindered by software limitations. The original NeoSLAM implementation, developed within the ROS 1 ecosystem using Python 2.7 and the legacy NuPIC framework, suffers from architectural bottlenecks. Specifically, its synchronous request/response communication model and reliance on outdated technologies prevent the system from processing incoming sensory data at full rates, leading to significant data loss in real-time applications.

Methodology
This work presents a complete architectural refactoring and software stack upgrade of the NeoSLAM algorithm. The methodology involves three primary technical shifts:

1. Architectural Refactoring: The original monolithic neoslam_server node was decomposed into four independent, concurrent nodes: Visual Feature Extractor, Binary Projector, Neocortex, and Spatial View Cells. By transitioning from a synchronous request/response paradigm to a fully concurrent publish/subscribe pipeline (using ROS 2), the system's throughput is no longer constrained by the total processing latency of all stages combined, but rather by the slowest individual stage (the bottleneck).
2. Software Stack Modernization: The implementation was migrated from ROS 1 and Python 2.7 to ROS 2 Rolling. The legacy NuPIC framework was replaced with the modern htm.core library, implemented in C++17 with Python 3 bindings. Additionally, random matrix files were converted to binary format, and efficient libraries (Eigen and CRoaring) were adopted for matrix operations and sparse binary data handling.
3. System Modules: The new implementation comprises seven ROS 2 nodes:
   • Visual Feature Extractor: Uses a pre-trained AlexNet (conv3) to extract features, with options for frame striding and cropping.
   • Binary Projector: Performs dimensionality reduction via random Gaussian projection and binarization using Sparse Locality Sensitive Binary Hashing (sLSBH) to create a compact binary descriptor.
   • Neocortex: Implements Hierarchical Temporal Memory (HTM) principles, specifically Temporal Memory (TM), to encode temporal context and expand the binary input into a higher-dimensional Sparse Distributed Representation (SDR).
   • Spatial-View Cells: Groups SDR observations into intervals based on similarity thresholds to detect loop closures and assign visual template identifiers.
   • Pose Cells, Experience Map, and Visual Odometry: These modules remain consistent with the RatSLAM architecture.

Key Contributions
The paper outlines three main contributions:

• Architectural Refactoring: A modular redesign that significantly improves computational efficiency, enabling near-full capacity operation in real-time with minimal discarding of incoming sensory data.
• Software Stack Upgrade: A transition to modern, maintainable technologies (ROS 2, C++17, htm.core) that resolves the scalability and long-term support issues of the original Python 2.7/ROS 1 stack.
• Experimental Evaluation: A systematic comparative analysis between the new NeoSLAM, the original NeoSLAM, and RatSLAM across three distinct datasets (Robotarium, iRat Australia, and FIU MMC Lake) under varying environmental conditions.

Experimental Results
The evaluation utilized three datasets: Robotarium (terrestrial, indoor), iRat Australia (terrestrial, outdoor), and FIU MMC Lake (aquatic, Unmanned Surface Vehicle).

• Real-Time Performance: The new implementation drastically reduced data loss compared to the original. On the Robotarium dataset, data loss dropped from 91% (Original) to 1% (New). On the iRat Australia dataset, it dropped from 90% to 0%. This improvement allowed the new system to process frames at the input rate, whereas the original system discarded the majority of frames.
• Mapping Consistency: In the Robotarium and iRat Australia datasets, the original NeoSLAM failed to close loops effectively due to data loss, resulting in incoherent topological maps. The new implementation generated consistent maps comparable to RatSLAM.
• Trajectory Accuracy:
  • iRat Australia: RatSLAM showed a slightly lower average error (0.24 m) compared to NeoSLAM (0.29 m), though both produced topologically consistent maps.
  • FIU MMC Lake: In this challenging aquatic environment with purely inertial odometry, NeoSLAM achieved a slightly lower average error (5.92 m) compared to RatSLAM (6.10 m). The authors noted that tuning the exclude_recent_intervals parameter was necessary to prevent incorrect re-localization.

Significance and Claims
The paper claims that the proposed ROS 2-based implementation of NeoSLAM successfully overcomes the processing throughput limitations of the original version, making it viable for real-time applications. The results demonstrate that the new NeoSLAM achieves map reconstruction performance comparable to RatSLAM across diverse environments. Notably, the authors state that this is the first work to evaluate NeoSLAM on a USV dataset, highlighting its ability to construct topological maps that improve upon degraded pure odometry trajectories even under conditions of significant visual ambiguity and inertial-only navigation. The work emphasizes that while the new NeoSLAM matches RatSLAM in reconstruction quality, its modernized architecture offers superior maintainability and scalability for future integration with current robotics tools.