Learning Inertial Parameter Identification of Unknown Object with Humanoid Robot using Real-to-Sim Adaptation

We present a fast learning-based inertial parameters estimation framework capable of understanding the dynamics of an unknown object to enable a humanoid (or manipulator) to more safely and accurately interact with its surrounding environments. Unlike most relevant literature, our framework doesn't require to use of a force/torque sensor, vision system, and a long-horizon trajectory. To achieve fast inertia parameter estimation, a time-series data-driven regression model is utilized rather than solving a constrained optimization problem. Due to the challenge of obtaining a large number of the ground truth of inertia parameters in the real world, we acquire a reliable dataset in a high-fidelity simulation that is developed using a real-to-sim adaptation. The adaptation method we introduced consists of two components: 1) \textit{Robot System Identification} and 2) \textit{Gaussian Processes}. We demonstrate our method with a 4-DOF single manipulator of a wheeled humanoid robot, SATYRR. Results show that our method can identify the inertial parameters of various unknown objects quickly while maintaining sufficient accuracy compared to other methods. Manipulation and locomotion experiments were also carried out to show the benefit of using the estimated inertia parameters from control perspective.