sjtuyinjie/M3DGR

What it solves

It addresses the need for robust Simultaneous Localization and Mapping (SLAM) on ground robots, particularly when operating in "degraded conditions" where sensors may fail or provide unreliable data (e.g., darkness, wheel slippage, or LiDAR degeneracy).

How it works

The project provides a three-pronged approach to improving ground SLAM:

1. M3DGR Benchmark: A comprehensive dataset featuring 32 sequences across various challenging scenarios, including visual challenges (darkness, occlusion), LiDAR degeneracy (corridors, elevators), wheel slippage (grass, rough roads), and GNSS denial.
2. Ground-Fusion++: A modular SLAM framework that integrates heterogeneous sensors to maintain localization and mapping quality even when individual sensors are compromised.
3. Comprehensive Evaluation: A benchmark of over 40 cutting-edge SLAM methods tested against the M3DGR dataset to establish performance baselines.

Who it’s for

It is designed for robotics researchers and engineers developing SLAM and localization algorithms who need a standardized, challenging dataset to test resilience and sensor-fusion capabilities.

Highlights

• Diverse Sensor Suite: Includes dual LiDARs (Livox Avia and MID-360), RGB-D and omnidirectional cameras, wheel odometry, GNSS, and RTK.
• Systematic Degradation: Specifically captures data in scenarios like varying illumination, dynamic obstacles, and rough terrain to test edge cases.
• High-Precision Ground Truth: Uses OptiTrack motion-capture, RTK, and ArUco markers for accurate validation.
• Extensive Baselines: Provides adapted code for 40 different SLAM algorithms to facilitate fair comparison.