Tiny quadrotor learns to fly in 18 seconds

Original Article ↗ Hacker News Discussion ↗

A research project on a tiny quadcopter that learns to stabilize and fly in about 18 seconds using a neural network controller sparks interest in how quickly control policies can be trained in simulation and transferred to real hardware. Commenters compare this approach with traditional PID-based autopilot stacks like ArduPilot, PX4, and Betaflight, debating ease of use, protocol design, and whether machine learning should replace or just tune classical controllers. The thread also touches on parallels (and limits) with animal instinctive learning, as well as the implications of increasingly autonomous, agile drones for hobbyists, industry, and warfare.

DIY quadrotors, firmware, and platforms

  • Several comments discuss DIY builds using ArduPilot, PX4, Betaflight, and TinyWhoop-style quads.
  • ArduPilot is described as mature and feature-rich, but some say it is hard to configure for small quads and has a messy codebase; others argue it’s powerful, hobbyist-friendly, and very capable for long-range/autonomous missions.
  • PX4 is reported as widely used in industry/prosumer products, stronger for scientific/industrial integration, and benefits from a permissive BSD license. Critics say its community is smaller and support less accessible.
  • Betaflight is seen as the go-to for freestyle FPV, with good defaults and presets.
  • Crazyflie, the platform used in the paper, is seen as research-focused; one commenter criticizes its flight performance without motion capture, while the project author counters that the optical flow deck provides robust indoor positioning.

Protocols: MAVLink, DroneCAN, ROS2

  • MAVLink is heavily criticized as overcomplicated, hobby-origin, poorly structured for C++, thread-unsafe in common implementations, and bloated in overhead.
  • Despite this, it is acknowledged to “get the job done” and has active maintainers and vendors.
  • DroneCAN is described as a differential bus with a good ecosystem but criticized for bit-level quirks, poor documentation, and a “use the official library” mindset.
  • ROS2 is mentioned as a more modular, transport-agnostic alternative for higher-level systems, though not directly usable on bare-metal targets.

Animal learning vs innate behavior (“Fulmar problem”)

  • Long subthread debates whether birds (and other animals) “learn” complex behaviors (flight, walking, web-building) or execute mostly hardwired instincts with minimal calibration.
  • Points raised: precocial vs altricial species, human underdevelopment at birth, genetic/epigenetic encoding of behavior, and the need for calibration because biological “hardware” varies.
  • No consensus; some emphasize innate programs, others highlight rapid but real learning driven by feedback from the environment.

Control, PID tuning, and neural networks

  • Discussion on using “AI” to auto-tune PID vs replacing PID entirely with neural controllers.
  • Some note that classic PID assumes linear time-invariant systems; quadrotors are nonlinear, so geometric controllers and optimal control are suggested.
  • Others argue that insisting on PID unnecessarily constrains learning; if compute allows, end-to-end neural control (as in the paper, with NN outputting motor PWMs) is appealing.

“Learning” terminology and realism of the demo

  • Some object that evolutionary or RL-style optimization in simulation is just “refining” an algorithm, not “learning.”
  • The project’s author clarifies that no handcrafted controller was preprogrammed; a neural net directly maps state to motor commands.
  • Another subthread notes the system is trained on many simulated agents, then transferred once to hardware; this is contrasted with how humans or animals learn from few trials.

COTS experimentation and TinyWhoops

  • For experimentation, Crazyflie is cited as the primary open platform from the article, though criticized as fragile and less “fun” than whoop-style quads.
  • TinyWhoop/Mobula-series micro FPV quads are recommended for hobby flying; DJI drones are suggested for cinematic, non-FPV use.

Warfare and ethics

  • Several comments link this work to autonomous weapon swarms and “Slaughterbots”-style scenarios.
  • Some argue such systems should be banned; others argue they should replace humans on battlefields to reduce human casualties.
  • Drone interception (e.g., nets) is mentioned as an active research area.