Author: Cioffi
Year: 2023
- hybrid dynamics model: point mass vehicle model + learned component to model aerodynamics effect
- estimates external forces + robot state
- history of thrust and IMU to predict dynamics
- soa VIMO only consider drône as a point mass without aerodynamics effects
- IMU + thrust commands -> temporal NN
- the TNN needs only position and velocity signals for training (no gt needed)
- dynamic equation: $$ \dot{\boldsymbol{v}}{\mathcal{B}k}^{\mathcal{W}}=\boldsymbol{R}{\mathcal{B}k}^{\mathcal{W}}\left(\boldsymbol{f}{t_k}^{\mathcal{B}}+\boldsymbol{f}{\text {res } s_k}^{\mathcal{B}}+\boldsymbol{f}_{e_k}^{\mathcal{B}}\right)+\boldsymbol{g}^w $$
- $\boldsymbol{f}{t_k}^{\mathcal{B}}$ is the mass normalized thrust, $\boldsymbol{f}{e_k}^{\mathcal{B}}$ is the external force acting on the platform,
$\boldsymbol{f}_{\text {res } s_k}^{\mathcal{B}}$ is a residual term that takes into account aerodynamic effects - use only the gyro for rotation
- sliding window MAP estimation with vision, pre integ, dynamics and marginalization
- the dynamic residual is the difference between the actual delta position and velocity and the one obtained by integrating dynamic equations with estimated forces using discrete time Euler numerical integration
- TCN = as powerfull as RNN but with less computations
- the loss of the CNN is the same as in the factor graph but opitimising on
$\boldsymbol{f}_{e_k}^{\mathcal{B}}$ - Similar perfo of the TCN wrt methods using full state
- TCN exhibits good generalization to situations unseen while training