Acoustic Odometry for Wheeled Robots on Loose Sandy Terrain

This work proposes a system capable of estimating the longitudinal velocity of a wheeled robot on loose sandy terrain using only acoustic sensors. These will likely be present anyway in many future ground mobile robots for human-robot interaction purposes. The proposed system consists of an audio feature extraction module, based on gammatone filterbanks, and a prediction module, based on a convolutional neural network. The proposed system has been tested in a single wheel test bed with a wheel driving up to speeds of 0.07 m/s in a wide range of wheel slippage resulting in an average drift of 5 mm/s. A qualitative evaluation of the proposed system against other sources of odometry shows that acoustic and visual methods vulnerabilities do not overlap, which indicates that a system based on acoustic sensors can be a feasible auxiliary source of odometry. The system is able to make a prediction from a single audio frame with a duration of 15 ms in only 2 ms on a user-level commercially available CPU.