Robotics & Embodied AI
An autonomous drone won several races against each human champion and posted the fastest recorded race time
In a 2023 paper published in Nature, a peer-reviewed journal, researchers at the University of Zurich and Intel report that an autonomous system named Swift raced physical drones at the level of human world champions. The system learned through deep reinforcement learning in simulation combined with data collected in the physical world. Deep reinforcement learning is a training method in which a program tries actions in a simulated setting and improves from feedback. In real-world head-to-head races, Swift competed against three human champions, including the world champions of two international leagues. It won several races against each of them and set the fastest recorded race time.
What did the study find?
The study reports a direct comparison between Swift and three human champion drone pilots. Swift won several races against each champion. The researchers also report that Swift posted the fastest recorded race time in the comparison. First-person-view drone racing is a sport in which pilots steer drones through a 3D circuit while seeing video from an onboard camera. The abstract does not report how many total races were run.
How was Swift trained?
The abstract says Swift combined deep reinforcement learning in simulation with data collected in the physical world. It does not specify how much of the training came from each source. The paper describes the challenge this way: the robot needs to fly at its physical limits while estimating its speed and location in the circuit only from onboard sensors.
What else should a reader know?
Did Swift win every race?
The abstract reports that Swift won several races against each human champion. It does not say whether Swift won every race or how many races it lost.
What does deep reinforcement learning mean here?
It is a training method where a program learns by trying actions in a simulated environment and using feedback to improve. In this study, the simulation learning was combined with data from the physical world.
Where was the work published?
The paper appeared in Nature, a peer-reviewed journal, in 2023.
Limits and what's next
The abstract describes one study with three human champions, including the world champions of two international leagues, in real-world head-to-head races. It reports no independent replication. The abstract does not give the total number of races, the number of losses, or the exact share of training that came from simulation rather than physical-world data. The authors write that the work may inspire hybrid learning-based solutions in other physical systems, but the abstract does not state what the researchers plan to study next.
Source: Elia Kaufmann et al., “Champion-level drone racing using deep reinforcement learning,” Nature, 2023.