Natural Swing-Up of an Underpowered Double Pendulum

Arne Sachtler, Manuel Keppler, Florian Loeffl, Alin Albu-SchäfferarXiv 2026 · arXiv

TL;DR

Researchers have developed a "gain-free" controller that enables a robot with extremely weak motors—delivering only 6% of the torque needed to hold its arm horizontal—to perform a complex swing-up maneuver. By splitting control into "pumping" energy and "steering" along the robot's natural nonlinear normal modes, they successfully navigate the system's chaotic dynamics using a simple, real-time linear program.

What Problem Does This Paper Address?

In traditional robotics, we often use "brute force" control: motors are powerful enough to cancel out gravity and friction to follow any arbitrary path. However, for underpowered robots or those designed for high efficiency, this isn't an option. These robots must "ride" their natural dynamics rather than fight them.

The authors tackle the classic double pendulum swing-up problem under extreme constraints. Their hardware (see Figure 1) uses direct-drive motors that are too weak to even hold the arm out straight. Because a double pendulum is a chaotic system, simple energy-based control (which works for a single pendulum) often results in unpredictable, messy motion that fails to reach the upright position. The challenge is to inject energy ("pumping") while simultaneously keeping the robot on a specific, predictable path ("stabilization") when both tasks must share the same tiny sliver of available motor torque.

Fig. 1. (a) Close-up of the robot, link 1 hanging and link 2 upright: direct- drive motors in both joints, electronics and battery on the links. (b) With full torque in both joints, the model’s static pose is q = (0◦, 16.3◦): link 1 hangs straight down and link 2 tilts by 16.3◦. (c) Model: link lengths li, centers of mass ri, masses mi, inertias Ii, viscous damping di; q1 from the downward vertical, q2 relative to link 1.

Methods

The core of the paper is a shift from tracking trajectories to riding Nonlinear Normal Modes (NNMs). NNMs are families of periodic orbits that represent the "natural" oscillations of the system at different energy levels.

The authors introduce three key technical innovations:

  1. Power-Based Input Split: They decompose the motor torque into two orthogonal components:
  2. Pumping ($\gamma$): Acts along the direction of momentum to change the system's energy.
  3. Steering ($\beta$): A "power-neutral" component that rotates the velocity vector without changing energy, used to keep the robot on the NNM manifold.
  4. Linear Program (LP) Allocation: Instead of manually tuning gains to balance these two tasks, the controller solves a small optimization problem every millisecond. It tries to maximize "pumping" subject to two constraints: the physical limits of the motors ("box constraints") and a Control-Lyapunov Function (CLF) constraint that ensures the robot doesn't drift too far from its intended mode.
  5. Gain-Free Design: Unlike standard PD controllers, this approach doesn't require hand-tuned feedback gains. The required decay rate for stabilization is derived directly from the system's natural "transverse instability" (measured via Floquet exponents), effectively letting the physics of the robot dictate the control effort.

Datasets

This research is experimental and model-based rather than data-driven in the machine learning sense. The authors used:

  • Identified Robot Model: A high-fidelity model of their custom planar two-link robot, including link inertias, centers of mass, and viscous damping.
  • Hardware Experiments: Real-world runs on the DLR "SwingBot" platform to validate the controller against two baselines: "sum-and-clipping" (a common heuristic) and "prioritized stabilization."

Key Insights and Ideas

  • Non-Strict Modes are Tricky: On a "strict" mode, you can pump energy without leaving the path. However, the authors prove that their robot's modes are "non-strict." This means injecting energy necessarily pushes the robot off its path, making the coordination between pumping and steering critical.
  • The Kinetic-Energy Metric Matters: The authors found that analyzing the robot's motion in a warped "kinetic-energy metric" space (see Figure 2b) provides a much clearer picture of the system's geometry than standard Euclidean coordinates.
  • Timing Over Force: The swing-up is achieved not by pushing hard, but by pushing at exactly the right moments to build resonance, much like a child on a swing.
Fig. 2. (a) Brake orbits of NNM 1 and NNM 2 at 2 to 34 J (light to dark) on the configuration torus Q = S1 × S1, drawn as a donut (NNM 1 left, NNM 2 right, also in (b)); q1 runs around the large circle, q2 around the tube, and q2 = 0 is the outer equator (grey), with the hanging point at the front; arrows: increasing q1, q2; fine dashes: parts behind the surface. (b) The same torus in the kinetic-energy metric, embedded isometrically as a surface of revolution, one copy q2 ∈[−π, π]; grey: lines q1 = const. Dashed red in (a) and (b): the cut q2 = ±π, red dot: (q1, q2) = (π, ±π); in (b) the two rims are the same circle, identified after a rotation by 54.7◦. (c) Period T and Floquet exponent λF of the free orbits over their energy; dashed: Eupup.

Key Results

The proposed controller successfully performed swing-ups along two different natural modes:

  • NNM 1 (Links swing together): Reached the target energy in 13.6 seconds.
  • NNM 2 (Links swing against each other): Reached the target energy in just 7.7 seconds.

Quantitatively, the LP-based allocation stayed significantly closer to the natural dynamics than the baselines. For example, on NNM 2, the proposed method maintained a median distance to the manifold of 0.07 $\sqrt{J}$, while the "prioritized" baseline drifted to 0.39 $\sqrt{J}$. The "sum-and-clip" method failed to reach the target energy entirely on the second mode.

Tip paths and energy profiles showing successful swing-ups.

Limitations and Future Work

  • The "Catch": The controller handles the swing-up to "almost upright," but a separate precomputed trajectory and PD controller are still needed to "catch" the robot in a stable handstand.
  • Greedy Optimization: The current LP maximizes pumping at every instant. While energy-efficient, it is not necessarily "time-optimal" for the entire maneuver.
  • Complexity Scaling: While the theory extends to $n$ joints, the paper demonstrates it on a 2-DOF system. Future work will look at higher-degree-of-freedom robots and "phase-free" stabilizers that don't rely on a pre-calculated energy-phase chart.

Summary updated

Abstract

Robots with weak motors must ride their natural dynamics rather than fight them. We swing up a direct-drive planar two-link robot, whose motors deliver only about 6\,\% of its shoulder gravity torque, along its nonlinear normal modes, two-dimensional invariant manifolds consisting of periodic orbits. On a non-strict mode, injecting energy necessarily moves the state off the manifold, so pumping and stabilization compete for the same bounded torque. We split the input torque into a power-neutral steering component and a scalar pumping input, the only component that changes the energy. The allocation then becomes a linear program in two variables: maximize pumping subject to the torque limits and one control-Lyapunov constraint. The constraint uses the model-based rate of the error, so no stabilizing feedback gain is needed. Experiments in hardware show swing-ups along both modes and compare the allocation with summing-and-clipping and with prioritized stabilization.
BibTeX
@misc{sachtler2026natural,
  author = {Arne Sachtler and Manuel Keppler and Florian Loeffl and Alin Albu-Schäffer},
  title = {{Natural Swing-Up of an Underpowered Double Pendulum}},
  year = {2026},
  eprint = {2611.17575},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17575},
}
Figure