AcroRL uses reinforcement learning for aggressive quadrotor inverted flight

A team of robotics researchers from several institutions (Gabriel Rodriguez, Henri Sayag, and colleagues) have introduced AcroRL, a learning-based framework for aggressive quadrotor inversion maneuvers. Bidirectional thrust allows quadrotors to achieve a second equilibrium condition and greater control authority, enabling inverted flight, perching, and sensing. Prior geometric control methods using Hopf fibration-based attitude representations struggled with actuator saturation and motor reversal delay during inversions, requiring heuristic thrust posture scheduling and waypoint tuning.

AcroRL instead uses separate reinforcement learning policies for nominal-to-inverted and inverted-to-nominal transitions, modulating a constant reference trajectory to perform compact, position-constrained inversions. In JAX-based simulation, the method achieved a 32% reduction in position root mean square error and a 57% reduction in settling time compared to the strongest optimization-based baseline. Hardware experiments demonstrated successful inversion across multiple yaw configurations with position RMSE under 0.35m, and compatibility with downstream trajectory generation and control—demonstrated via circular flight in both regimes. The implementation is open-source.