Browsing by Author "Kalavagunta, Rahul"

Robotic drone platforms require robust, precise control solutions that are able to preserve safety while achieving high task performance in order to operate in high uncertainty, constrained environments. Safety filters provide a scalable method to ensure the safety of the system, but due to the dynamical complexity of drone platforms, synthesizing safety monitors and intervention systems can be challenging. This project presents Safe-Aviary, a modular, gym-like simulation environment that extends the gym-pybullet-drones simulator. Safe-Aviary is optimized for training safety fallback policies using the Iterative Soft Adversarial Actor Critic for Safety (ISAACS) deep reinforcement learning algorithm. We demonstrate that the resulting safety filters are effective in reducing safety violations, and assess the impact that filtering during conventional policy training can have in Safe Reinforcement Learning contexts.