Zhejiang University trains drones to fly through gaps

Researchers at Zhejiang University built a vision-based control system that lets a small quadcopter fly through confined openings using only onboard sensing, according to Hackster.io. The project replaces a traditional robotics stack with a learned sensorimotor policy that maps camera images and other onboard sensor inputs to low-level flight commands, per Hackster.io.

Per Hackster.io, the team trained a reinforcement learning policy in simulation and transferred it to a physical quadrotor. The custom test vehicle measured 38 cm wide and 10 cm tall and carried:

• •a monocular camera
• •a PX4 flight controller
• •an NVIDIA Jetson Orin NX computer

Hackster.io reports that the researchers used trajectories from a model-based planner to guide exploration during training, addressing the sparsity and danger of random trial-and-error when learning aggressive gap traversal.

Industry observers note that end-to-end, vision-driven sensorimotor policies have become a preferred research approach for agile robotics because they can compress perception-to-action latency and reduce payload complexity. Comparable projects typically rely on careful simulation-to-real transfer methods, guided demonstration data, or planner-assisted exploration to avoid unsafe exploration in the real world.

For practitioners and labs, key indicators are whether the team publishes training code or domain-randomization details, test robustness under wind and lighting variation, and scalability to larger platforms or additional maneuvers. Improvements in planner-assisted RL and safer real-world exploration techniques would make similar tactics more practical in applied robotics.