Researchers teach drone to fly autonomously by incorporating an obstacle avoidance AI algorithm which was trained using data from 11.5K crashes
Evan Ackerman / IEEE Spectrum :
Context & Ripple Effects
This 2017 report lands one year after consumer drones first shipped onboard autonomy: the Movidius Myriad 2 chip gave the DJI Phantom 4 its obstacle-avoidance features. What the researchers add is a training method — an avoidance algorithm learned from 11,500 recorded crashes rather than hand-coded rules.
The line from there runs straight through the related coverage: within a year Skydio turned the same idea into the R1, a self-navigating 4K camera drone, other labs showed a drone could fly city streets on data borrowed from cars and bikes (training without extra onboard hardware), and by 2023 the UZH/Intel Swift system was beating human FPV racing champions.
First-order effects
- The named players here are the researchers and drone makers: crash-data training offers a cheaper route to obstacle avoidance than sensor-heavy designs, directly challenging the dedicated-chip approach Movidius sold into the Phantom 4.
Second-order effects
- Autonomy becomes the competitive axis for consumer drone startups — Skydio's R1 bet its entire product on self-navigation just months after this work, forcing incumbents like DJI to keep deepening onboard AI rather than competing on flight time alone.
Third-order effects
- If learning-from-failure keeps scaling, the endpoint visible in the corpus is autonomy far beyond quadcopters: Swift out-racing humans in 2023 and DARPA's ACE program flying AI on a modified F-16 point toward aircraft trained the way these researchers trained a drone — on accumulated failure data rather than explicit programming.
The trend: Autonomous flight is progressing from crash-trained lab demos to systems that outperform human pilots, with each generation needing less bespoke hardware and more learned behavior.