Every drone fleet operator, military logistics unit, and emergency response team faces the same hidden bottleneck: the software that flies one drone cannot fly another. Change the motor. Swap the propeller. Use a different frame. The control system breaks. Each new hardware configuration requires its own calibration, its own training data, its own tuned policy. This is not a minor inconvenience; it is one of the core reasons drone deployment at scale remains slow, expensive, and fragile.
A paper published this week in Science Robotics suggests that bottleneck may be removable.
What the Researchers Built
Researchers at UC Berkeley and the Technology Innovation Institute in the UAE developed RAPTOR, a single neural network policy that can fly any quadrotor drone it is handed — without ever having seen that specific drone before, without calibration, and without retraining.
The network has 2,084 parameters. For context, a typical smartphone photo contains roughly 10 million times more data. It is, by the standards of modern AI, almost absurdly small.
Yet in testing, this tiny policy successfully flew 10 completely different quadrotors it had never encountered: drones ranging from 32 grams — lighter than a AAA battery — to 2.4 kilograms, a professional-grade platform. The drones differed in motor type, frame rigidity, propeller blade count, and the underlying flight controller software. In conventional drone engineering, each of these differences would normally require a separate, hardware-specific control system. RAPTOR flew all of them.
Adaptation to a new drone happens in milliseconds, using only the drone’s recent flight behaviour. The policy watches how the vehicle responds to its initial commands and adjusts instantly, without any prior knowledge of the hardware’s specifications.
Why This Matters
The assumption that drone control policies must be platform-specific is not an arbitrary constraint — it is baked into how drone software has been developed for two decades. Controllers are tuned to specific motor curves, weight distributions, and aerodynamic profiles. Changing the hardware means starting the tuning process again.
RAPTOR challenges that assumption directly. The authors describe the problem using an analogy that holds up: a human driver can get into a car they have never driven before and navigate it reasonably well within seconds. They do not need to study the engine specifications or recalibrate their reflexes from scratch. They adapt on the fly from the vehicle’s feedback. Modern drone control software cannot do this. RAPTOR begins to.
The engineering consequences are concrete. A single deployable policy that works across an entire drone fleet — regardless of hardware variation, component wear, or emergency substitution — would dramatically reduce the software overhead of operating at scale. For a military unit deploying mixed drone assets, a disaster response team using whatever hardware is available, or a logistics company managing a heterogeneous fleet, the ability to run one controller across all platforms is not a minor optimisation. It removes an entire layer of operational complexity.
The Mechanism
RAPTOR uses a technique called in-context learning — the same principle that allows large language models to adapt their responses based on the conversation so far, applied here to physical control. The policy’s hidden layer uses recurrence: it encodes a short history of the drone’s recent behaviour and uses that context to infer the platform’s physical characteristics without being told them.
The policy was trained by first generating 1,000 simulated quadrotors with different properties and training a specialist controller for each one. Those 1,000 specialists were then distilled into a single student policy through a process the researchers call meta-imitation learning. The result is a controller that has, in effect, absorbed the knowledge of a thousand specialists into a package small enough to run on constrained hardware in real time.
Testing included outdoor flights, wind disturbance, deliberate physical interference (poking the drone mid-flight), and propeller swaps mid-experiment. The policy adapted in all cases.
Limitations Worth Noting
The 10 drones tested, while genuinely diverse, represent a limited sample of the full design space of quadrotors. Highly unusual configurations — very large platforms, unconventional geometries, heavily damaged hardware — may push outside what the policy can adapt to from flight history alone.
The current policy also focuses on trajectory tracking and stabilisation. More complex autonomous tasks — mapping, payload delivery, precise landing on moving platforms — will require the foundation policy to be extended or composed with higher-level planning systems.
But the core result stands: a general-purpose drone brain exists, it has been physically tested across meaningfully different hardware, and it works. The next question is not whether one policy can fly any drone. It is how far that generalisation extends.
Sources
Eschmann, J., Albani, D., & Loianno, G. (2026). RAPTOR: A foundation policy for quadrotor control. Science Robotics, 11(114), eaec1481. https://doi.org/10.1126/scirobotics.aec1481 https://www.science.org/doi/10.1126/scirobotics.aec1481
Eschmann, J., Albani, D., & Loianno, G. (2025/2026). RAPTOR: A foundation policy for quadrotor control. arXiv preprint arXiv:2509.11481 (v1 submitted September 15, 2025; v2 revised April 6, 2026). https://arxiv.org/abs/2509.11481 https://arxiv.org/pdf/2509.11481
Project page
RAPTOR project page (code, videos, supplementary materials): https://raptor.rl.tools
Data and model repository
Eschmann, J., Albani, D., & Loianno, G. (2025). RAPTOR: A foundation policy for quadrotor control (v1). Zenodo. https://doi.org/10.5281/zenodo.17096679
Quotes in this article are drawn from the research paper published in Science Robotics on May 13, 2026. DOI: 10.1126/scirobotics.aec1481
More Like This
AI Teaches Drones to Keep Their Wireless Signals Locked On Target
Sam Brown holds a BSc in Electronic Systems Engineering from University of Regina, a Masters from the University of Alberta and a PhD from the University of Toronto in Electrical Engineering. His reporting interests include Physics and Artificial Intelligence.