Extended Data Table 1 Network parameters and detailed metrics show how our Swift system compares with other approaches

From: Champion-level drone racing using deep reinforcement learning

  1. a, Training hyperparameters. b, Comparison with baselines that are provided with the same observation noise model used by our approach. c, Evaluation in simulation, with idealized dynamics (top) versus realistic dynamics (bottom) and ground-truth observations (left) versus noisy observations (right). We report the fastest achieved collision-free lap time in seconds, the average and smallest gate margin of successfully passed gates and the percentage of track completed. We compare our approach with a learning-based approach that performs zero-shot transfer, with and without domain randomization during training, as well as a traditional planning and control approach28. d, Comparison of the average speed, power, thrust, time and distance travelled for each pilot during the fastest flown race. Best numbers are indicated in bold.
Back to article page