Learning and Adaptation of Inverse Dynamics Models: A Comparison

Performing tasks with high accuracy while interacting with the real world requires a robot to have an exact representation of its inverse dynamics that can be adapted to new situations. In the past, various methods for learning inverse dynamics models have been proposed that combine the well-known rigid body dynamics with model-based parameter estimation, or learn directly on measured data using regression. However, there are still open questions regarding the efficiency of model-based learning compared to data-driven approaches as well as their capabilities to adapt to changing dynamics. In this paper, we compare the state-of-the-art inertial parameter estimation to a purely data-driven and a model-based approach on simulated and real data, collected with the humanoid robot Apollo. We further compare the adaptation capabilities of two models in a pick and place scenario while a) learning the model incrementally and b) extending the initially learned model with an error model. Based on this, we show the gap between simulation and reality and verify the importance of modeling nonlinear effects using regression. Furthermore, we demonstrate that error models outperform incremental learning regarding adaptation of inverse dynamics models.