Deep reinforcement learning (DRL) policies for planetary landing are usually trained with randomized disturbances to encourage robustness. But random perturbations may not cover the most challenging conditions a lander could face. A new preprint on arXiv argues that this standard stochastic training leaves room for improvement.
The authors propose an adversarial reinforcement learning framework, where a disturbance model is trained to oppose the landing policy. Instead of sampling disturbances from a fixed distribution, the adversary learns to produce inputs that are most likely to cause failure, forcing the landing policy to adapt to worst-case scenarios.
Because this is a single preprint with a truncated abstract, the specific algorithm and experimental results are not detailed here. The core idea, however, is clear: making the disturbance generation part of the training game could yield more robust guidance, navigation, and control for future landing missions.