Title:Task-driven SLAM Benchmarking

Abstract:For assistive robots, one critical use case of SLAM is to support localization as they navigate through an environment completing tasks. Current SLAM benchmarks do not consider task-based deployments where repeatability (precision) is more critical than accuracy. To address this gap, we propose a task-driven benchmarking framework for evaluating SLAM methods. The framework accounts for SLAM's mapping capabilities, employs precision as a key metric, and has low resource requirements to implement. Testing of state-of-the-art SLAM methods in both simulated and real-world scenarios provides insights into the performance properties of modern SLAM solutions. In particular, it shows that passive stereo SLAM operates at a level of precision comparable to LiDAR-based SLAM in typical indoor environments. The benchmarking approach offers a more relevant and accurate assessment of SLAM performance in task-driven applications.