Active Learning for Exciting Motion Generation With Safety Constraint: Toward Reducing Model-Reality Gap in Inertial Parameters

Inertial parameters should be estimated accurately for precise robot control and simulation. Exciting motions, motions that sufficiently excite all robot dynamics, must be generated to obtain these parameters accurately. However, this process requires a certain level of accuracy in the inertial parameters themselves, leading to a circular dependency, especially in cases with a significant model-reality discrepancy. To address this challenge, we propose a constrained method for generating exciting motions within an iterative data acquisition process (active learning). It optimizes the condition number of the regressor matrix and achieves sufficient excitation by combining several motions. We evaluate our method on a manipulator robot not fixed to the ground, where dynamic constraints should be considered to maintain balance during experiments. Despite a substantial gap between the initial model and the actual robot (simulated by adding a 0.575 kg payload at the end effector), our method continuously improved the condition number without motion execution failures–whereas conventional methods resulted in robot tipping. Cross-validation confirmed that the proposed method achieves the lowest root mean square error among all approaches, demonstrating superior performance for inertial parameter identification. These results collectively demonstrate that our method is particularly practical in applications where dynamic constraints are critical for motion planning.