Deep Reinforcement Learning for Control of a Small-Scale Helicopter

This paper presents a deep reinforcement learning (DRL) framework for the control of a small-scale helicopter, aimed at overcoming the limitations of traditional control methods in handling nonlinear dynamics and diverse flight conditions. The proposed approach integrates a full-GPU training pipeline implemented in NVIDIA Isaac Lab, leveraging the RSL-RL library and the Proximal Policy Optimization (PPO) algorithm. This framework supports the parallel simulation of thousands of environments, significantly accelerating policy acquisition through high-throughput data collection. Simulation results demonstrate that the learned policies are robust against external wind gust disturbances and capable of accurately tracking complex reference trajectories, effectively capturing the underlying nonlinear dynamics and generalizing across diverse operating conditions.