Robot Dynamics Identification

Robot dynamics identification studies how to estimate a robot’s motion model so it can predict forces and trajectories for control, simulation, and transfer from simulation to hardware. Common approaches use designed excitation motions with least-squares or Bayesian estimation, online filters and moving-horizon methods, and differentiable simulation; neural networks often capture residual effects that simpler models miss. These methods are tested on manipulators, legged robots, quadrotors, underwater vehicles, and systems with contact or flexible components, where better models can improve tracking and sim-to-real transfer. A central challenge is balancing physically consistent, interpretable models with flexible models that capture complex effects, especially when parameters are hard to distinguish or measurements are limited.