Swift drone uses AI technology to defeat top human players in FPV event

Artificial intelligence has once again defeated the human world champion, this time in drone racing.

This major result was published in the magazines "Nature" and "Science Robotics" by a team of engineers from the University of Zurich in Switzerland. After the first race against an AI drone in 2011, no autonomous drone could beat a human pilot until Swift came along. Swift defeated world champion human players, including 2019 Drone Racing League World Champion Alex Vanover, two-time MultiGP International Open Champion Thomas Bitmatta, and three-time Swiss National Champion Marvin Schaepper

In the virtual realm, artificial intelligence has defeated humans in games such as chess, checkers, Go and StarCraft. Now, for the first time, it has also successfully defeated humans in a physical challenge

This paragraph has been rewritten as follows: The game was conducted by a "First Person View" (FPV) drone equipped with a high-resolution camera. An article was published in Science Robotics magazine detailing the historic victory achieved by the Swift drone development team. In the fierce competition with three drone competitors, this drone won 15 out of 25 challenges. The person in charge of the project said, "Our results mark the first time that an AI-driven robot has defeated humans in a real-life experience event designed for humans and led by humans."

The secret to the success of the "Swift" drone lies in its powerful artificial neural network, which can optimize the drone's route and speed. It collects environmental details in real time through onboard cameras to provide precise guidance for drones, while human pilots rely on video signals transmitted to headphones to experience a "first-person perspective."

In this drone race, the human operator controls the drone through the 3D track through the onboard camera. The innovation of the Swift system is the ability to map the drone's status to commands to adjust thrust and spin rate. This achievement is a milestone in the field of mobile robotics and machine intelligence.

Swift technology introduction

Swift is a quadcopter that is autonomously controlled using only onboard sensors and computing. The aircraft consists of two key modules:

1. Perception system converts high-dimensional visual and inertial information into low-dimensional representation;
2. Control strategy, ingests the low-dimensional representation generated by the perception system and generates control commands.

Among them, the control strategy is represented by a feedforward neural network and trained using model-free on-policy deep reinforcement learning (RL)

Given the differences in sensing and dynamics between simulation and the real world, optimizing strategies only in simulation will result in poorer actual performance of the drone. Therefore, the research team decided to use data collected from physical systems to estimate a non-parametric empirical noise model

Research results show that these empirical noise models play a positive role in successfully transferring control strategies from simulation to reality

Specifically, Swift converts sensor readings on the aircraft into control commands. This conversion process includes two parts:

(1) Observation strategy to refine high-dimensional visual and inertial information into task-specific low-dimensional encoding;

(2) Control strategy, convert encoding into drone commands.

Of the 10 losses recorded by Swift, 40% were due to collisions with opponents, 40% were due to collisions with competition gates, and 20% were due to being slower than human-controlled drones. Overall, Swift won the most races against human-controlled drones, and it also set the fastest race record, beating the best time of a human-controlled drone (A. Vanover) by half a dozen Second.

While the Swift was faster than all human-controlled drones overall, it wasn't faster on every segment of the track.

Careful analysis by the research team found that: when taking off, Swift has a shorter reaction time and takes off 120 milliseconds earlier than human pilots on average; Swift also accelerates faster and enters the first competition gate at a higher speed. In tight turns, the Swift's movements are tighter.

The research team also proposed a hypothesis that Swift optimizes trajectories on a longer time scale than human operators. It is known that model-free reinforcement learning can be achieved by optimizing long-term rewards. In contrast, human operators have a shorter time scale for planning movements and can only predict one competition gate in the future at most

The above is the detailed content of Swift drone uses AI technology to defeat top human players in FPV event. For more information, please follow other related articles on the PHP Chinese website!