Note on learning sensitivity metrics for a quadrotor
Résumé
Robust trajectory planning can involve the use of ’uncertainty tubes’
which bound the system’s states and inputs. These tubes can be derived
from a variety of metrics, involving complex simulation computations.
This is the case for the sensitivity-based uncertainty tubes computation
we consider in this work. In fact, this computation was shown to be com-
putationally expensive, and even more in the context of sampling-based
planners, where it has to be performed at least once for each new sample.
In order to solve this problem, a learning-based approach is presented in
this work to predict these sensitivity-based uncertainty tubes. The report
is organized as follows: first, it recalls the sensitivity basis and the tube
computation, then presents the chosen system and controller. Next, the
network architecture is presented, and finally the results including dataset
generation, training to different types of network, from LSTMs and RNNs
to GRUs, and evaluation are given in a final section