UZH Robotics and Perception Group

Peer-reviewed work grounded in high-speed autonomous flight and onboard perception.

No.Paper · what it does and showsIndustry contribution

Dream to Fly: Model-Based Reinforcement Learning for Vision-Based Drone Flight2026 ↗

Uses DreamerV3 model-based RL to learn pixel-to-command drone racing policies, emerging with camera-pointing behaviour that keeps gates in view, and flying real hardware up to 9 m/s.

Positions model-based RL as a practical route to agile autonomy for teams whose data and compute budgets are far smaller than model-free methods require.

Learning Agile Quadrotor Flight in the Real World2026 ↗

Removes the need for precise system identification or offline sim-to-real transfer: adaptive temporal scaling explores the platform’s physical limits while online residual learning and in-flight policy updates raise a conservative 1.9 m/s policy to 7.3 m/s in about 100 seconds of flight.

Replaces fragile identification pipelines with in-flight adaptation, letting drone vendors keep improving aggressive flight performance after deployment.

Learning Quadrotor Control From Visual Features Using Differentiable Simulation2025 ↗

Back-propagates gradients through a differentiable simulator to train quadrotor controllers, learning recovery policies in seconds from vehicle state and in minutes from visual features only.

Demonstrates that gradient-based simulation training can cut RL sample cost by orders of magnitude for drones and other real-time control products.

Learning High-Speed Flight in the Wild2021 ↗

Trains a vision-based policy that flies a quadrotor at high speed through cluttered forest trails using onboard sensing only, with no map and no external motion capture.

Showed onboard end-to-end learning can replace map-and-plan pipelines for drones, the basis of much of today’s autonomous inspection and defence drone autonomy.