RIA-CSM2: A Real-Time Impact-Aware Correlative Scan Matching Algorithm Using Heterogeneous Multicore SoC for Low-Cost Wheeled Robots

As a classic scan-to-map matching method, the correlative scan matching (CSM) algorithm may not be applicable if low-cost wheeled robots (like robot cleaners) are impacted. The first open issue is heavy dependence on trustworthy initial poses. Side slips caused by impacts are unobservable for rotary encoders mounted on wheels, leading to huge localization errors. The second open issue is the efficient processing of global localization using the CSM algorithm, which is essential for impacted robots. The state-of-the-art hardware designs for large-scale multiresolution CSM have low energy efficiency. These two open issues are properly addressed in this work, namely, RIA-CSM2. Based on lightweight deep neural networks, we perform reliable impact detection in unforeseen environments using only low-cost proprioceptive sensors. To bind the rapid error growth of conventional wheel-aided inertial navigation systems (INSs), delayed out-of-sequence measurements from CSM algorithms are integrated into an extended Kalman filter (EKF). The lightweight impact detection networks and INS are generally applicable for most embedded robotic systems with stringent energy, computing, and memory cost limitations. Once impacts are detected, large-scale multiresolution CSM algorithms will be performed on an energy-efficient hardware accelerator. Extensive experiments based on public datasets show that our work achieves high real-time performance and energy efficiency. The frame rate of local-scale high-resolution CSM can reach up to 96.42 frames/s. Field experiments on wheeled robot platforms demonstrate the effectiveness of our impact detection network, which outperforms our preliminary work in precision and false-alarm rate by a significant margin, with precision and recall rates reaching 100% and 97.8%, respectively. The presented impact detection datasets are now publicly available.