Towards Online System Identification: Benchmark of Model Identification Techniques for Variable Dynamics UAV Applications

Providing self-modelling capabilities to robotic systems that could change their dynamics variables during their natural operation, like aerial manipulators, can significantly increase model-based algorithm’s resilience and performance. Some samples of those techniques are model predictive control or model-based trajectory planning techniques. This paper aims to benchmark how classical model identification techniques perform when adapted to be used at run-time. This self-awareness capability will improve transparency giving the robot the capability to understand what it can do and how to perform optimally. To do so, this paper compares four different model identification techniques and how they perform online in terms of accuracy and computational cost. The online adaptation of the model identification methods has been developed generically to apply to any dynamic system. The methods are numerically evaluated simulating an Unmanned Aerial Vehicle (UAV) with time-varying mass. The system has been evaluated in 5 different manoeuvres with 5 different mass-behaviour. This creates 25 experiments for each of the four model identification algorithms. In total, this paper presents the result of 100 studied cases.