An Iterative Online Approach to Safe Learning in Unknown Constrained Environments

This paper presents an iterative learning technique to safely guide a nonlinear system with unknown dynamics through an environment with unspecified constraints. The presented approach leverages the system's local dynamics to incrementally explore the environment and learn the appropriate control, which allows us to avoid the data-intensive task of learning an accurate global system model. Due to the local nature of this approach, the system's safe operating region does not need to be pre-specified as long as local areas of the constraints can be identified when the system approaches those areas. The functionality and efficiency of the proposed approach are demonstrated through simulation of a unicycle and a high-dimensional nonlinear quadcopter, indicating the system's ability to learn dynamics from data and safely navigate unknown environments.