Aerospace engineer Peter Ryseck published his self-balancing pole experiment in September 2026. Hackaday examined it shortly afterward, and renewed technical coverage on 8 November brought the control-system details back into view. Ryseck's reported outdoor run lasted more than three hours on a single 2,200 mAh 3S battery. The unusual endurance is useful evidence about the experiment, but the device is not hovering, carrying its own weight in the air, or demonstrating three-hour quadcopter flight.
The pole offers a compact lesson for RC aircraft builders: actuator delay, structural flex and estimator drift can defeat a controller that looks perfectly reasonable in a rigid-body simulation.
The ground carries the weight; the propellers correct the fall
The lower tip of Ryseck's pole rests on a contact area roughly the size of a coin. Most of its weight passes down the carbon-fibre shaft into the ground. A four-motor head at the upper end applies small lateral corrections when the pole tilts. The motors run near idle for much of the demonstration, which is why the flight battery lasts so long.
The distinction between supporting weight and controlling attitude is essential. A flying quadrotor must continuously generate thrust near its weight and still reserve enough additional authority to maneuver. Ryseck's pole uses the ground as a load path. The motors mainly counter the overturning torque caused by the center of mass moving off the support point, plus disturbances from wind and touches.
It is a form of actively stabilized inverted pendulum. As the pole begins to lean, gravity increases the tendency to fall. The control system tries to produce a corrective moment before that lean grows beyond what the available motors can arrest.
A longer pole can be easier to control in one respect because its falling motion evolves more slowly, although a longer, flexible shaft may introduce worse bending. Those two consequences pull design choices in opposite directions. The experiment's value lies in making both effects visible with ordinary drone hardware.
The first motors had enough thrust but responded too slowly
Ryseck initially tried larger propellers. Their attraction was understandable: large rotors often produce a useful force efficiently at lower rotational speed. Yet the prototype could not stabilize the pole. The difficulty, according to Ryseck's account, was the time the propellers needed to change thrust after a new command.
A spinning propeller stores angular momentum. Changing its speed takes motor torque, and the available acceleration depends partly on rotating inertia. A controller that demands a rapid change may receive the right amount of thrust after the pole has already tipped too far. Raising proportional gain cannot remove that physical delay. It can make the response overshoot and oscillate.
The builder tried smaller propellers, including 3-inch ones, then settled on 5-inch blades. The selected hardware offered sufficient lateral correction while allowing a faster response. That is an empirical result from his particular pole and motor combination, not a general recommendation that 5-inch propellers are optimal for other stabilizers.
For RC builders, the important test quantity is not only maximum static thrust. It is the command-to-force response: how quickly the actuator produces a repeatable change, whether it responds symmetrically when speeding up and slowing down, and how much response remains near the chosen idle speed.
That last point explains why setting idle to zero did not necessarily help this project. In the reported test, a small continuous motor speed yielded a more stable response than repeatedly starting motors from rest. An aircraft's motor-stop and failsafe rules are different, and this result should not be transferred into flight configuration without purpose-built testing.
Rigid-body simulation missed the pole's own bending mode
Ryseck began tuning with a simulation based on the measured mass and inertia of the pole. His controller used proportional and derivative terms: one reacted to angular displacement and the other to angular rate. Those terms can stabilize a simplified inverted pendulum when actuator response is sufficiently fast.
The real pole was a carbon-fibre pool-cleaning-net shaft, not a perfectly rigid link. Its tip moved relative to the motor assembly, creating an additional flexible mode. The gyro therefore measured a mixture of meaningful overall tilt motion and rapid vibration from shaft bending.
The renewed technical report describes a prominent bending feature around 18 Hz, while Ryseck places the desired balancing motion mostly below 5 Hz. In that setup, the derivative term could amplify the wrong signal: it treated high-frequency bending as a command to correct the pole, and the resulting motor motion fed energy back into the flex.
The source of the oscillation was not simply a poorly chosen gain. A flexible structure adds dynamics the rigid-body model did not include. The effective loop then contains a second moving object coupled to the first, and the controller can excite it.
This problem appears elsewhere in RC engineering. A long camera mast, soft flight-controller mount, flexible wing or high-gain gimbal may have a structural response that differs sharply from rigid-body motion. Each needs measurements of the actual installed assembly rather than assumptions based on bare component specifications.
Filtering helped only because the useful and unwanted motion were separated
A straightforward response to unwanted gyro vibration is to low-pass the measurement. That suppresses high-frequency movement, but it also delays the signal that tells the controller when to respond. With a falling inverted pendulum, enough delay can destabilize the useful slow loop.
Ryseck's solution combined limited low-pass filtering with a notch that identified the pole's resonance when the system armed. The notch targeted the narrow flexible mode while leaving more of the slower balancing response available. The video and reports describe the motor sound changing as the resonance filter engaged.
The distinction between low-pass and notch filtering is practical. A low-pass reduces a broad range of frequencies above its cutoff. A notch attacks a narrower band near a measured resonance. Both add phase effects, and neither can recover information that a sensor never captured correctly or that sampling has already aliased into the control band.
ArduPilot's harmonic notch documentation describes related principles for motor and propeller noise: identify vibration frequencies from logs, choose a tracking source, and verify attenuation after filtering. Its system follows motor-related frequencies using throttle, RPM telemetry, an RPM sensor or FFT estimates. Ryseck's target, as reported, was the pole's structural bending frequency, so it is not the same implementation or a claim that ordinary ArduPilot harmonic-notch settings reproduce his result.
There is also a CPU budget. ArduPilot warns that multiple notches and harmonics cost processing time. A bench project with one obvious mode and a purpose-built controller gives fewer reasons to add a stack of broad filters than a complex aircraft with several motor tones. The right question is which measured frequency interferes with the control objective.
The wind problem reveals a useful role for estimator bias
A balanced pole in still air has a natural upright reference. In steady wind, forcing an exact vertical reference may require the motors to keep producing a continuous lateral moment. Ryseck allowed his target attitude to shift gradually so the pole could lean a little into the wind. That lets gravity supply more of the counteracting moment and reduces persistent motor effort.
According to his explanation, the adaptation also compensates for small inertial-sensor alignment errors. A gyro and accelerometer combination can estimate tilt, but a tiny bias in the angle reference can translate into constant control demand when the goal is defined as perfectly vertical.
The adaptation is useful because the disturbance is relatively slow. It would be unsafe to let a target angle wander quickly in response to every gust or every inertial transient: the controller could start redefining falling as the new upright. Separating fast stabilization from slow trim adaptation is what makes the concept plausible.
This illustrates a broader flight-control distinction between the command a vehicle is trying to track and the slow corrections needed for a persistent environmental load. For a conventional quadrotor, an estimator reset, wind estimate or trim mechanism must still satisfy navigation and safety constraints. It cannot simply chase whatever reduces motor current.
Three hours is a control-power result, not a flight-endurance result
The reported pole used a 2,200 mAh three-cell pack, a Pixhawk-class controller, four drone motors and a printed head. The November report notes a pack voltage of 11.1 V at the three-hour mark and warns against a circulating interpretation that the device ran for 3.75 hours: in the source video, 3.75 is the voltage per cell, not the time.
A voltage reading alone cannot establish remaining capacity or average power consumption. Voltage changes with load and chemistry, and a single figure at the end of an outdoor run says little about how much energy went to motors, the flight computer or intermittent corrections. Neither the pole's total weight nor measured average current is provided in the cited video reporting.
A strong reproducibility package would log current, voltage, commanded and measured motor speed, angular rate, wind conditions, pole dimensions and deflection. That would let another builder distinguish energy savings from favorable weather, drivetrain selection or control tuning. The public demonstration gives an observed duration, not a controlled endurance benchmark.
What this experiment says about an airborne drone
It would be a mistake to read the project as proof that a conventional drone could stay aloft for hours on a small 3S pack. Its physics are different. Yet the experiment isolates failure mechanisms that are easy to miss when many things happen at once during flight:
- A propeller may have adequate maximum thrust while failing the required response-time test.
- More aggressive control gain can worsen oscillation when the plant includes motor delay or structural flex.
- A flexible mount can create a measured rotation signal that does not mean the vehicle body is actually turning at that rate.
- A poorly chosen filter can remove the noise and still destroy the phase margin needed to stabilize the real motion.
- Persistent trim demands may reveal wind load or sensor bias, and those explanations need to be separated.
These are engineering interpretations of the demonstrated setup, not results independently measured across different drones.
The prototype also carries obvious physical hazards. Four exposed propellers at head height on a long pole do not become harmless because the device usually runs near idle. Responsible replication belongs in a guarded, controlled test area with an emergency motor cutoff. The reported park demonstration is not a safety case for public installation.
The strongest lesson is about measuring the real plant
The pole began as a simple inverted pendulum. Its first controller model omitted an actuator response limit and then a flexible structural mode. Each omission mattered enough to make the real system unstable despite apparently sufficient motor thrust.
Ryseck eventually made it stand by matching propellers to the needed response time, measuring the bending mode, filtering narrowly and allowing slow adaptation to persistent wind and alignment error. The result is visually memorable. The lasting technical point is less theatrical: a controller is only as good as its assumptions about the dynamics between command and measured response.