Constrained Differentiable Policy Learning for Safe Quadrotor Navigation

This research presents a method for safe quadrotor navigation using constrained differentiable policy learning. The approach formulates obstacle avoidance as an explicit constraint and employs primal-dual gradient descent-ascent through a differentiable simulator. Two key components—bounded clearance-based constraints and log-mean-exp aggregation—enable effective constrained training. Theoretical guarantees and experiments demonstrate improved safety-performance trade-offs and generalization to out-of-distribution environments, with zero-shot transfer to both simulation and physical quadrotors.