Minimax Iterative Dynamic Game: Application to Nonlinear Robot Control Tasks

Deep reinforcement learning policies provide good control decision strategies in high-dimensional state spaces, particularly in complex autonomous tasks. However, they exhibit weak performance guarantees in the presence of disturbance, model mismatch, or model uncertainties. This brittleness limits their use in high-risk scenarios. We present how to quantify the sensitivity of deep reinforcement learning policies -- this would inform of its robustness capacity. We then formulate a minimax iterative dynamic game for designing robust policies in the presence of an adversarial input. The algorithm is simple and is adaptable for designing meta-learning/deep policies that are robust against disturbances, model mismatch or model uncertainities, up to a disturbance bound. Videos of our results can be seen here: https://goo.gl/JhshTB.