Learning high-speed flight in the wild

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Quadrotors exhibit high agility for traversing complex environments.
• Current autonomous quadrotor systems are limited to low speeds due to sequential processing in sensing, mapping, and planning.
• High-speed navigation in cluttered environments is challenging for traditional methods because of processing latency and error propagation.

Purpose of the Study:

• To develop an end-to-end autonomous system for high-speed quadrotor flight in complex environments.
• To enable quadrotors to operate using only onboard sensing and computation.
• To overcome limitations of traditional methods by reducing processing latency and improving robustness.

Main Methods:

• Proposed an end-to-end approach directly mapping sensory observations to collision-free trajectories using a receding-horizon method.
• Utilized a convolutional neural network for sensorimotor mapping, trained in simulation via privileged learning.
• Incorporated realistic sensor noise simulation for zero-shot transfer to real-world scenarios.

Main Results:

• Achieved high-speed autonomous flight in complex, previously unseen environments (forests, collapsed buildings).
• Demonstrated significantly reduced processing latency compared to traditional pipelines.
• Outperformed conventional obstacle avoidance methods in challenging real-world tests.

Conclusions:

• End-to-end policies trained in simulation can enable high-speed autonomous quadrotor flight.
• Direct sensorimotor mapping is a viable strategy for robust and low-latency navigation.
• The approach shows promise for real-world applications requiring rapid autonomous traversal of cluttered spaces.