Residual Dynamics Learning

Overview

Residual dynamics learning studies how to correct an approximate physics model using the differences between its predictions and real or simulated behavior. Common approaches learn these corrections with neural networks, Gaussian processes, sparse system identification, or Koopman models, often preserving physical structure and updating online as conditions change. The resulting models are used in settings such as robot control, model predictive control, vehicle navigation, and sim-to-real transfer, with evaluation on trajectory tracking, long-horizon prediction, and contact or disturbance tasks. A central challenge is balancing expressive corrections that generalize across changing conditions with models that remain interpretable, data-efficient, and fast enough for real-time control.

Latest papers

CardsList

32 papers4 papers in the last four weeks, down 33% on the four weeks before. 0.0% of all new papers.

  1. TEFL: Prediction-Residual-Guided Rolling Forecasting for Multi-Horizon Time Series

    Feb 26, 2026Xiannan Huang, Shen Fang, Shuhan Qiu +3Time Series ForecastingResidual Dynamics Learning

  2. Nonplanar Model Predictive Control for Autonomous Vehicles with Recursive Sparse Gaussian Process Dynamics

    Feb 18, 2026Ahmad Amine, Kabir Puri, Viet-Anh Le +1Model Predictive ControlResidual Dynamics Learning