Abstract
This study presents an empirical evaluation of two prominent RGB-D SLAM frameworks, RTAB-Map and Voxgraph, within large-scale simulated indoor environments. Using the Habitat simulator, the researchers benchmarked these algorithms across long trajectories (up to 1km) under varying noise conditions. The results highlight a significant trade-off: RTAB-Map achieves higher absolute positioning accuracy through its complex memory management, while Voxgraph demonstrates superior memory efficiency and local trajectory consistency, making it better suited for resource-constrained autonomous systems.
Introduction
Simultaneous Localization and Mapping (SLAM) is a foundational technology for autonomous robotics, enabling navigation in unknown environments without external positioning systems like GPS. While SLAM is well-studied for small-scale environments, large-scale indoor navigation—spanning kilometers of corridors and multiple rooms—presents unique challenges. Two primary bottlenecks are memory consumption, as dense 3D maps can quickly exceed available RAM, and error accumulation (drift), where sensor noise leads to map inconsistencies like "tearing" or corridor duplication. This paper evaluates how RTAB-Map and Voxgraph address these long-term autonomy challenges.
Methods
The evaluation compares two distinct architectural approaches to SLAM:
RTAB-Map (Real-Time Appearance-Based Mapping)
RTAB-Map is a feature-based framework known for its robust loop closure detection. It utilizes a three-level memory management system consisting of Short-Term Memory (STM), Working Memory (WM), and Long-Term Memory (LTM). This system ensures real-time performance by keeping only the most relevant data in the WM for loop closure, while archiving older data to the LTM.
Voxgraph
Voxgraph represents a more modern, submap-based approach. Instead of maintaining a single global map, it generates a collection of overlapping submaps using Truncated Signed Distance Fields (TSDFs). Global drift is managed by optimizing a pose graph where nodes represent individual submaps. This approach is designed to be computationally efficient and reduce the overall memory footprint.
Experimental Setup
The study utilized the Habitat simulator to provide a photo-realistic university building model. Testing involved two primary trajectories: a 300m path with 12 waypoints and a more challenging 1km path with 25 waypoints. To simulate real-world conditions, four levels of Gaussian noise (Small to Extra-Large) were added to the ground-truth odometry.
Key Results and Insights
The comparative analysis revealed a stark contrast in resource utilization and accuracy profiles.
Memory Efficiency
The most significant finding was the memory consumption gap. On the 1km trajectory, RTAB-Map’s memory usage reached 8.6 GB, a level that would be prohibitive for many embedded robotic platforms. In contrast, Voxgraph required only 1.8 GB for the same task, representing a nearly 5x improvement in memory efficiency.
Accuracy Trade-offs
The algorithms exhibited different strengths regarding accuracy:
- Absolute Accuracy: RTAB-Map outperformed Voxgraph in Absolute Trajectory Error (ATE), maintaining a closer proximity to the global ground-truth coordinates.
- Local Consistency: Voxgraph showed lower relative translation and rotation errors, indicating better precision in estimating the robot's step-by-step movement.
Failure Modes under High Noise
Both systems struggled with high odometry noise. A common failure mode was the inability to correctly identify loop closures in long corridors, leading to the generation of "fake corridors" or map bifurcations. While adding the DBoW2 library to Voxgraph improved its loop closure capabilities, it did not fully resolve these structural mapping errors.
Conclusion
The choice between RTAB-Map and Voxgraph depends on the specific requirements of the robotic application. RTAB-Map is the preferred choice for high-precision navigation where hardware resources are abundant. However, for long-duration missions on resource-constrained hardware, Voxgraph’s memory efficiency and local consistency offer a compelling alternative. Future work should focus on developing more memory-efficient loop closure mechanisms and exploring topological SLAM to better handle structural ambiguities in large-scale environments.
Summary updated