Artificial intelligence (AI) has once again defeated a human champion
This time, in the drone racing field.
A team led by Dr. Elia Kaufmann from the Robotics and Perception Group at the University of Zurich and Intel's team jointly designed an autonomous driving system called Swift, which can drive drones better than human opponents in one-on-one championship matches.
This groundbreaking research was just published as a cover article in the latest issue of Nature magazine.
In a news and views article published simultaneously in Nature, Guido de Croon, a professor at the Delft University of Technology's Institute of Technology, wrote, "Kaufmann et al.'s research is a good example of how roboticists can overcome real-world gaps. Although Swift is trained using a clever combination of AI learning techniques and traditional engineering algorithms, the system should be further developed in a more realistic and variable environment to fully unleash the potential of this technology."
However, the research team stated that the study marks a milestone in mobile robotics and machine intelligence, or may inspire the deployment of mixed learning-based solutions in other physical systems, such as autonomous ground vehicles, flying vehicles, and personal robots.
Intelligent training combining AI and engineering algorithms At present, AI systems based on deep reinforcement learning have surpassed human champions in games such as Atari games, chess, StarCraft, and Gran Turismo, but all these achievements have taken place in virtual environments, rather than in the real world.
Drone racing presents a challenge for experienced pilots and AI, but it is even more challenging for AI. In virtual environments, resources are almost limitless, but turning to the real world means having to use limited resources. This is especially true for drones, as the sensors and computing equipment that replace human pilots must be carried into the air.
Moreover, the real world is more unpredictable than the virtual one. While a simulated drone can follow a pre-programmed trajectory perfectly, a single command to a drone in the real world can have multiple effects, with unpredictable consequences, especially for drones trained by AI.
Traditional end-to-end learning methods are difficult to transfer the mapping from virtual environments to the real world, as there is a real gap between the two, and this real gap constitutes one of the main challenges in the field of robotics.
In this study, the Swift system achieved intelligent training by integrating AI learning technology with traditional engineering algorithms. First, the system processes images captured by the drone's camera using a neural network to precisely detect the corner of the door. Then, it uses binocular vision software to calculate the drone's speed.
The innovative aspect of the Swift system is another artificial neural network that maps the drone's state to commands for adjusting thrust and rotational rate. It uses reinforcement learning to optimize rewards obtained from the environment through a trial-and-error process in the simulation. In this algorithm, the system employs reinforcement learning, rather than end-to-end learning, thus enabling it to bridge the gap between reality and simulation through abstract concepts.
Since the state encoding has a higher abstract level than the original image, the reinforcement learning simulator no longer needs a complex visual environment. This optimization reduces the difference between the simulation system and the real system and improves the simulation speed, enabling the system to complete learning in about 50 minutes.
According to the paper, Swift consists of two key modules: observation policy and control policy. The observation policy is composed of a visual inertial estimator and a door detector that can convert high-dimensional visual and inertial information into task-specific low-dimensional encoding; the control policy is represented by a two-layer perceptron that can accept low-dimensional encoding and convert it into drone commands.
Exceeding the speed and performance of human pilots The racecourse for this competition was designed by a world-class FPV (first-person view) pilot from outside the organization. The course includes seven square gates arranged in a 30x30x8 meter space, forming a 75-meter track.
Moreover, the course features distinctive and challenging maneuvers, including Split-S. Even if a collision occurs, as long as the drone can continue flying, the pilot can still continue the race. If a collision occurs and both drones fail to complete the course, the drone that is further away from the finish line wins.
Swift competed in multiple races with Alex Vanover (2019 Drone Racing League World Champion), Thomas Bitmatta (2019 MultiGP Champion), and Marvin Schaepper (3X Swiss Champion) among others.
In particular, Swift won 5 out of 9 races against A. Vanover, 4 out of 7 races against T. Bitmatta, and 6 out of 9 races against M. Schaepper.
Additionally, Swift had 10 losses, with 40% caused by colliding with opponents, 40% caused by colliding with gates, and 20% caused by flying slower than human pilots.
Overall, Swift won the majority of races against each human pilot. Furthermore, Swift set the fastest race time record, beating A. Vanover's best human pilot performance by half a second. It can be seen from the data analysis that Swift is generally faster than all human pilots, especially in critical parts such as takeoff and emergency turns. Swift's takeoff reaction time is shorter, on average 120 milliseconds ahead of human pilots. Furthermore, Swift has a greater acceleration, reaching a higher speed at the first gate.
In addition, Swift exhibits tighter maneuvering actions during sharp turns, which may be because it optimizes its trajectory on a longer time scale. On the other hand, human pilots tend to plan actions on a shorter time scale, at most considering the position of the next gate.
In addition, Swift achieved the highest average speed overall on the track, found the shortest race route, and successfully kept the aircraft flying near the limit. In the time trial, Swift was compared with human champions, and the autonomous drone showed consistently higher lap times, with lower average values and variances, while human pilots' performance varied more individually, with higher average values and variances.
The comprehensive analysis indicates that autonomous drone Swift has demonstrated excellent performance in the competition, not only in terms of speed, but also in terms of unique flight strategies that enable it to maintain a high level of performance throughout the competition.
It's not just drone racing This study explores autonomous drone racing based on noisy and incomplete sensor inputs from the physical environment, showing that an autonomous physical system achieved championship-level performance in the race, sometimes even surpassing the world champion in the human world, highlighting the significant meaning of robots achieving world championship-level performance in popular sports, and achieving an important milestone for robotics and artificial intelligence.
However, the system in the study was not trained for recovery after collisions, which limited its ability to continue flying after a collision, while human pilots can continue the competition even with hardware damage.
Moreover, compared to human pilots, the Swift system has weaker adaptability to environmental changes and uses cameras with a lower refresh rate; although the method performed excellently in autonomous drone racing, its generalization ability in other real systems and environments has not been fully explored.
Clearly, Kaufmann and his team's achievements are not limited to the drone racing field, and this technology may find its niche in military applications.
Furthermore, their technology can make drones more stable, faster, and longer-range, helping robots to more effectively utilize limited resources in driving, cleaning, and inspection, etc. But to achieve these goals, the research team still needs to overcome many challenges. As Croon commented in his review article, "In order to beat human pilots in any racing environment, the system must be able to deal with external disturbances such as wind, changing lighting conditions, poorly defined entry and exit points, other racing drones, and many other factors."
