Robots that operate in real-world conditions often
perform complex tasks in the presence of stochastic disturbances.
The source of the disturbances can be widely varied, including
but not limited to, hardware imperfections, atmospheric
changes, and measurement inaccuracies. These disturbances
pose a great control challenge because stochastic drift induces
changes in the robot’s speed and direction. This paper presents
an online trajectory generation method for robots to complete
preference-balancing tasks under stochastic disturbances. Task
learning is done off-line assuming no disturbances, and then
trajectories are planned online in the presence of disturbances
using the current observed information. We model the robot
as a stochastic control-affine system with unknown dynamics
impacted by a Gaussian process. This paper introduces a
supervised machine learning method in lieu of a traditional
greedy policy. We verify the method in simulation for an
aerial vehicle cargo delivery and a flying inverted pendulum
task. Results show the presented method works on a range
of problems and outperforms the deterministic method in the
presence of non-zero mean disturbances.