Inverse Dynamics

Inverse dynamics works backward from motion or state changes to estimate the forces, torques, or actions that caused them, supporting both physical analysis and control. Methods range from rigid-body equations, geometric formulations, and observers to learned inverse dynamics models, including transformers, flow-based models, and latent action predictors; some combine these with forward-model consistency or physics-based constraints. Applications include robot manipulation and locomotion, vehicle and aircraft control, human joint-torque estimation, and learning actions from video or world-model representations. A central challenge is that noisy or incomplete observations and model mismatch can make the inverse ambiguous: accurate predictions do not always yield robust closed-loop behavior, and learned methods must generalize beyond their training data.