Learning Agile Quadrotor Flight in the Real World (RSS 2026)

[00:00:03]Just like human motor learning, achieving true agility necessitates the ability to learn and adapt directly within the physical world.

[00:00:13]For quadrotors, learning-based control policies can enable highly accurate trajectory tracking.

[00:00:19]They can even unlock agile maneuvers, such as flying over 7 m/s in a tight 15 m² space.

[00:00:29]But this is not a new drone with more powerful motors, nor is it a different policy reliant on extensive offline training.

[00:00:37]It is a single policy that adapts in the real world, fully evolving in just 100 seconds.

[00:00:46]We propose a self-adaptive framework, enabling a quadrotor to evolve from a conservative policy to its actuation command limits, rapidly tripling its speed.

[00:00:57]We evaluate our method on an agile quadrotor platform using two trajectory tracking tasks, a figure-eight and a straight-line shuttle.

[00:01:06]Starting from a conservative policy, the agent tracks the trajectory while simultaneously collecting real-world data to refine both the model and the policy.

[00:01:16]This learning loop repeats continuously over approximately 100 seconds of flight.

[00:01:22]The result is a rapid improvement in performance. We achieve over a threefold speed-up in both scenarios while maintaining tracking errors within safety bounds throughout the entire adaptation process.

[00:01:37]We further test the robustness of our approach against unknown physical changes.

[00:01:43]Taking the additional payload scenario as an example, initially, the drone struggles to track the trajectory due to the altered mass.

[00:01:53]However, our method quickly captures the new dynamics, significantly reducing the tracking error in just one iteration.

[00:02:03]It then safely increases its speed, achieving an over threefold speed up while keeping tracking error within safety bounds.

[00:02:13]We observe similar resilience with a clipped propeller and even under combined failure modes.

[00:02:30]Quantitatively, as illustrated here, the method maintains the tracking accuracy under the threshold across all scenarios.

[00:02:39]Crucially, although hardware degradation limits the maximum achievable agility, our framework adapts to find a near-optimal policy for the current system, ensuring the drone performs close to its new physical boundary.

[00:02:53]Our method is well-suited for practical applications, such as autonomous inspection. In this experiment, the quadrotor is required to visit three different landmarks in sequence.

[00:03:06]To test resilience, an external fan generates strong wind disturbances during the segment between waypoints two and three.

[00:03:14]Starting from a conservative policy, our method rapidly adapts to this challenging environment.

[00:03:22]It accelerates to complete the task efficiently, ultimately reducing travel time by 50% compared to the initial policy.