In context
In 2025, quadrotor control predominantly relied on unidirectional motor thrust, limiting maneuverability and the potential of bidirectional motors. Deep reinforcement learning (DRL) had shown promise in low-level control, but its application to bidirectional thrust remained unexplored, hindering agile flight and rapid recovery scenarios.
What was reported
A team from the National University of Defense Technology introduced a DRL-based bidirectional thrust control method for quadrotors, combining for the first time deep reinforcement learning with bidirectional thrust. They developed a dynamic model incorporating symmetric propellers and bidirectional brushless motors, and designed a neural network controller using the PPO algorithm to output desired thrust for four motors.
The controller, with two hidden layers of 64 neurons each, was trained in a modified Flightmare simulator (FlightLxx) with randomized initial states and parallel training across 100 quadrotors. Simulations covered extreme conditions including large attitudes, high speeds, and high angular speeds, achieving rapid hovering.
Compared to a positive-thrust controller, the bidirectional thrust controller produced smoother actions, smaller state fluctuations, shorter control times, and stronger robustness. The output thrust-to-weight ratio ranged from -2g to 2g, enabling inverted flight and recovery from large disturbances.
“Compared to existing controllers using positive thrust, the proposed bidirectional thrust controller performs smoother actions, with smaller state fluctuations, shorter control times, and stronger robustness.”
Why it mattered
This work expands the action space of quadrotors, enabling new maneuvers such as inverted flight, landing on steep slopes, and recovery after flipping. It demonstrates that DRL can effectively handle bidirectional thrust, paving the way for more agile and versatile aerial robots in industrial inspection, logistics, and emergency response.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-06-09 · “基于深度强化学习的四旋翼无人机双向推力控制”