CASTRO: robust nonlinear trajectory optimization using multiple models

In this paper we present CASTRO, a new approach for achieving robust planned trajectories for nonlinear systems in the presence of modelling uncertainty. With CASTRO, we simultaneously optimize trajectories for multiple copies of the same system model, each using different estimates for the system parameters. The systems are constrained to use the same policy. With an appropriate sampling of system parameters in the optimization problem, the trajectory will be robust when run on the real system, compared to a trajectory optimized with just one model. We present results on a simulated double-link pendulum swing-up problem.