Chinese algorithm beats veteran human controller in FPV drone race
Chinese research team enables drones to perform high-risk and sophisticated aerobatic manoeuvres with a 100 per cent success rate
A Chinese team has developed a unique algorithm that gives first-person-view (FPV) drones the ability to perform autonomous aerobatic manoeuvres, unlocking their full potential to outperform humans in intense flight missions.
Details of the feat were published on April 16 in the peer-reviewed journal Science Robotics.
Aerobatic flight, as the name suggests, involves high-risk manoeuvres that require unstable postures – movements rarely executed in conventional flight operations for most aircraft.
However, in nature, aerial acrobatics are a vital skill for many species.
For instance, sparrowhawks and falcons can rapidly adjust speed and direction through vertical or inverted flight to hunt or avoid obstacles. Bats excel at mid-air flips and hanging upside down, while ravens perform complex aerobatics to attract their peers.
“This biological wisdom – transforming ‘high-risk manoeuvres’ into ‘high-survival rewards’ – holds the key to redefining traditional drone flight paradigms,” said Gao Fei, an associate professor at Zhejiang University.
His team introduced a comprehensive system that enables drones to generate and execute aerobatic manoeuvres even in environments with obstacles, achieving performance deemed better than professional pilots.
“Conventional methods focus on enhancing hardware performance, employing more powerful motors or higher-precision sensors, yet few have systematically investigated the rationality of manoeuvre-planning itself,” Gao said.
“We have demonstrated that quadrotor drones can unlock more sophisticated flight manoeuvres and graceful postures solely through intelligent algorithms alone – no hardware upgrades required,” he added.
The system works based on two key frameworks: motion-intent translation, which converts desired flight manoeuvres into actionable goals, and risk-reward evaluation, which balances obstacle avoidance, energy efficiency, and aerobatic performance.
“When physical limits cannot be breached, motion intelligence becomes the new frontier of performance,” the team noted.
After users provide their desired aerobatic trajectory, the system can generate complex flight paths and ensure stable execution in known environments, all without human intervention. The drone can autonomously execute these manoeuvres in intricate settings.
The researchers tested the system on two classic FPV drone models with different frame sizes and flight abilities. In an obstacle course, the drone demonstrated impressive agility, resembling a hummingbird. It executed inverted flights through 80cm (31.5 inches) rings, while performing continuous serpentine manoeuvres within tunnels.
To test the algorithm’s effectiveness, the team arranged a competition between the system and a professional human pilot with five years of experience. The challenge involved continuously flying through six square gates measuring 1.2 metres (four feet), executing six large loops while inverted.
The system achieved a 100 per cent success rate, while the human pilot succeeded just three times out of 24 attempts. Additionally, the human pilot needed more space between manoeuvres to stabilise the FPV, while the algorithm’s trajectory maintained tighter, more efficient movements.
“This confirms that the system’s ability to generate and execute complex manoeuvres has reached a level beyond human capability,” Gao said.
Drones have played a pivotal role in the Russia-Ukraine conflict, assisting in reconnaissance, surveillance, and strike missions.
If current drones gain aerobatic capabilities, their flexibility could enable smaller drones to navigate hidden terrain for reconnaissance or swiftly approach and accurately strike targets.
Although the system developed by the team allows users to customise flight plans and generate obstacle-avoidance trajectories, it currently depends on pre-existing environment maps for global planning. Also, it is not yet suited for swarm control.
“We believe that aerobatic flight can enhance a vehicle’s adaptability and flexibility in complex environments, improving its performance in practical applications,” Gao said.
“This research opens up new possibilities for quadrotor drones to perform aggressive missions in extreme environments,” he added.
“Examples include deploying probes near volcanic eruption vents, rapidly exploring narrow gaps in collapsed buildings during disaster rescue, and enabling spacecraft to navigate space debris.
“It could also significantly benefit aerial cinematography by allowing drones to autonomously capture smooth, stable footage without manual piloting or post-production stitching.”