Answers are grounded in this paper and cite their sources.
Velocity estimation with low-cost MEMS inertial measurement units (IMUs) is difficult because accelerometer bias, gravity-projection error, vibration, and integration drift accumulate rapidly. This paper presents a physics-guided residual-learning velocity estimator. A direction-cosine-matrix complementary filter removes gravity, an FFT-guided moving average aligned to the dominant vibration period suppresses vibration, and direct integration yields a preliminary velocity. A two-layer long short-term memory (LSTM) network then predicts only the residual error of this velocity from IMU-derived features, rather than regressing velocity directly. At inference, no zero-velocity update, Kalman filter, or external positioning is used. Weights are learned from straight-line motion only; curved-path validation trials serve solely for early stopping and hyperparameter tuning. Evaluation on a wheeled robot with an RTK-GPS reference covers ten trials at each of three unseen speeds over straight, circular, and stadium-shaped trajectories. The estimator achieves RMSE ≤ 0.0307 m/s and R2 ≥ 0.9872. Compared with an encoder-referenced multi-filter baseline, it significantly reduces RMSE in all conditions, reduces bias by two to ten times, and cuts cumulative-distance error by 52% on average without the encoder. Compared with an LSTM regressing velocity directly from the same features, it reduces RMSE by up to ten times, confirming the benefit of residual learning.
No takes yet. Share an insight, caveat, or question.
Aung et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: