Phase Slips as the Robustness Margin of Undulatory Swimming

A swimming eel or lamprey robot moves by passing a wave of bending down its body, produced by a chain of coupled oscillators, one per segment. Two findings in robotics pull against each other. The most energy-efficient gaits are clean traveling waves. Robustness against damage, such as a severed spinal connection or muted oscillators, comes from sensory feedback that bends and modulates those waves, and that feedback raises the energy cost. Engineers currently tune this trade-off by trial and error. They have no number that says how close a gait is to falling apart.

Physics already has a precise account of how traveling waves fall apart. In oscillatory media, a smooth wave fails when the oscillation amplitude at one point drops to zero and the phase there jumps by a full cycle. This event is a topological defect, called a phase slip in one dimension. We propose to treat the swimmer's segment chain as such a medium. After factoring out its symmetries, the gait becomes a fixed point, and its robustness margin can be measured as the distance to the nearest phase-slip transition. We will test whether this one geometric number predicts failure across perturbation types that were never used in tuning. We will also test whether it gives an explicit Pareto front between energy cost and robustness.

Core Claim

Our central hypothesis is that coordinated undulation fails in one common way: a localized phase slip in the chain of oscillators. If that is true, a single margin should predict survival under impulsive kicks, transection, sensor dropout and flow gusts better than gain magnitudes, linear stability, or survival rates tuned on one perturbation. The margin is defined as the minimal perturbation needed to push the gait across the boundary of the saddle state that mediates a slip, measured in symmetry-reduced coordinates. Computing it by numerical continuation in the reduced space should be much cheaper than finding rare failures by Monte Carlo sampling. The method will be judged on whether it predicts failure under unseen perturbations and whether it finds the cost–robustness front cheaply.

Where This Sits in Prior Work

Three groups of papers motivate the proposal.

The swimming-robot papers supply the empirical tension. Identification of the trade-off between speed and efficiency in undulatory swimming using a bio-inspired robot swept 152 prescribed gaits on the 8-joint 1-guilla robot. It found the lowest Cost of Transport (CoT, energy per distance), about 38 J/m, at traveling-wave index above 0.9. Every one of its gaits was imposed and unbroken, so wave breakdown was never studied. Emergence of robust self-organized undulatory swimming based on local hydrodynamic force sensing and Multisensory feedback makes swimming circuits robust against spinal transection and enables terrestrial crawling in elongate fish showed on the 10-segment AgnathaX robot that stretch and pressure feedback keep swimming coordinated across a transection, and that high stretch gain raises mechanical CoT. Both papers treat damage as a set of empirical cases. We reuse their architecture of oscillators plus stretch and pressure feedback and replace case-by-case testing with a margin computed from the dynamics.

The defect-turbulence literature supplies the mechanism. Modulated amplitude waves and defect formation in the one-dimensional complex Ginzburg–Landau equation reduced the complex Ginzburg–Landau equation, the generic model of weakly nonlinear oscillatory media, to a low-dimensional system and traced it with the continuation package AUTO94. It showed that modulated waves vanish at a saddle-node bifurcation, where the amplitude reaches zero and a defect is born. A local criterion based on this bifurcation predicted defect formation in 11 of 17 parameter sets. Nonlinear analysis of the Eckhaus instability: modulated amplitude waves and phase chaos with nonzero average phase gradient extends this picture to waves with a net phase gradient. That case matters here because a swimming gait is exactly a wave with nonzero average phase gradient along the body, which the file lists as an incompletely studied regime. The same papers also constrain the design. They report that phase-only models such as Kuramoto–Sivashinsky miss the transition because they cannot let the amplitude vanish, so our oscillator model must carry amplitude as well as phase.

The symmetry-reduction work supplies the method. How do discrete symmetries shape the stability of geophysical flows? used discrete parity and time-reversal symmetries to classify linear instability onsets but stopped at linear theory. We borrow its symmetry classification and push it into the nonlinear regime where defects form. Discrete Routh reduction factors out abelian (commuting) symmetries to expose relative equilibria, motions that look steady once a symmetry is removed, but only for conservative systems. Morphological symmetries in robotics used reflection symmetry to shrink a 12-dimensional trot to a 6-dimensional subspace, but never connected that reduction to energy cost or to failure. We apply reduction by the same abelian phase symmetry to a dissipative, actively driven chain. This works without symplectic machinery because we only need the quotient geometry, not preservation of conserved quantities.

Research Program

The model is a chain of N Stuart–Landau oscillators, the normal form of an oscillator near onset, written $\dot z_i = (1+i\omega_i) z_i - (1+ic)|z_i|^2 z_i + k(z_{i+1} - 2z_i + z_{i-1}) + g_s s_i + g_p p_i$. Here $z_i$ is the complex amplitude of segment $i$, $\omega_i$ is its frequency, $c$ is the nonlinear frequency shift, $k$ is the coupling to neighboring segments, $s_i$ and $p_i$ are the stretch and pressure signals, and $g_s, g_p$ are the feedback gains we tune. This is a discrete Ginzburg–Landau chain, so a phase slip is well defined: some $|z_i|$ passes near zero while the phase difference between neighbors jumps by $2\pi$. Joint commands drive a multi-link body. Fluid forces come from resistive force theory, a local drag model that is our modeling choice. The file reports no fluid model for this setting.

The reduction rests on two symmetries. Time-shifting an autonomous system gives a phase symmetry, so we work in the N−1 relative phases plus the N amplitudes. In those coordinates the steady gait is a fixed point rather than a limit cycle. The body also has a mirror symmetry: undulation is invariant under left–right reflection combined with a half-period shift. Following the classification logic of the geophysical-flow paper, this splits the gait's bifurcations into two kinds. Symmetry-breaking ones produce turning or drifting gaits. Symmetry-preserving saddle-nodes are where a slip saddle and the gait collide.

The margin is defined in these reduced coordinates. Let $x^*(g)$ be the gait fixed point at gains $g$ and $W(g)$ the stable manifold of the slip saddle. The margin is

$$M(g) = \min_{\delta} \|\delta\|_E \quad \text{subject to } x^*(g)+\delta \in W(g),$$

where $\|\cdot\|_E$ is a kinetic-energy-weighted norm, so that the margin has physical units. We compute it with a minimal-action shooting method seeded at the saddle. We then track $x^*$, the saddle and $M$ across gain space by continuation, and $M \to 0$ at the saddle-node. Running CoT along the same continuation branches gives the front between energy cost and margin directly. Finally, we check the mechanism itself by confirming in simulation that real failures under transection, kicks and flow gusts appear as localized slips. A slip from transection should sit at the cut site.

Baselines

The do-nothing reference is the pure central oscillator chain with no feedback, plus the open-loop prescribed traveling wave in the style of 1-guilla. These fix the lowest CoT reachable and the robustness that comes without feedback. The 38 J/m figure from Identification of the trade-off between speed and efficiency in undulatory swimming using a bio-inspired robot is only a hardware sanity anchor, because it comes from a different body. The file reports no CoT number for AgnathaX feedback configurations, so there is no reported result to beat directly.

The strongest competitor is direct black-box optimization of the gains. We run Bayesian optimization of CoT plus Monte Carlo survival rate under one perturbation family, mirroring how Yasui et al. tuned gains empirically. It gets exactly the same simulator, the same gain parameterization, and a matched budget of simulator time measured in integrated seconds of dynamics. The margin method beats it if two things hold together. Its front, found at equal or lower budget, must not be dominated in CoT versus held-out survival. And it must keep higher survival than the Monte Carlo-tuned gains when they are evaluated on perturbation families not used in tuning.

The second competitor is domain-randomized reinforcement learning with proximal policy optimization. It is run once over gains and once as a full policy, with randomization over all perturbation families and the same simulator-time budget. It is scored on the same held-out survival and on CoT under nominal conditions, where Blind Bipedal Stair Traversal via Sim-to-Real Reinforcement Learning suggests robust policies overpay.

Two ablations isolate the key ingredients. The first replaces $M$ with the linear Floquet margin, the decay rate of the gait's slowest perturbation, to test whether nonlinear, finite-perturbation geometry matters. The second computes the margin in a phase-only Kuramoto chain, to test whether the amplitude degree of freedom matters. A third ablation computes $M$ without the symmetry quotient and measures only the extra compute.

For the defect-prediction component, the reference is the reported 11 of 17 parameter sets correctly predicted by the saddle-node criterion in Modulated amplitude waves and defect formation in the one-dimensional complex Ginzburg–Landau equation. Our margin should match or exceed this on the same continuum setting.

Why not just…?

Why not just tune gains by Bayesian optimization on simulated survival? This is the real threat. A 10-segment simulation is cheap enough that brute force may find a good front. If the black-box baseline matches our front at equal budget and also transfers to held-out perturbations, the contribution shrinks to an explanation, and we will report it that way. Our bet is that survival tuned on one perturbation overfits to that perturbation. The margin, by contrast, targets the shared failure route. The held-out-perturbation test settles this directly.

Why not just train a domain-randomized reinforcement-learning policy? Without structure such policies are data-hungry, as noted in Morphological symmetries in robotics. Robustness-trained policies also tend to adopt conservative gaits that waste energy in nominal conditions, as in Revisiting Reward Design and Evaluation for Robust Humanoid Standing and Walking. The matched-budget comparison of nominal CoT and held-out survival will show whether that tax appears here.

Why not just use a phase-oscillator (Kuramoto) model, as AgnathaX controllers already do? Discrete phase chains can lose phase locking without any amplitude collapse, so a phase-only margin may be enough. Against this, the Ginzburg–Landau work reports that phase-only reductions miss the saddle-node transition. The Kuramoto ablation decides which view holds. If it predicts survival equally well, we drop the amplitude variables.

Why not just use linear Floquet stability? Linear rates describe only infinitesimal perturbations. The defect literature explicitly says linear thresholds do not predict finite-amplitude breakdown. The Floquet ablation tests this claim in the locomotor setting.

Proposed Experiments

We first validate the margin on the 1D complex Ginzburg–Landau equation. On the 17 parameter sets of the modulated-amplitude-wave study, we ask whether $M \to 0$ tracks defect formation at least as well as the 11 of 17 reported for the saddle-node criterion. This includes the nonzero-phase-gradient regime.

We then study the swimmer chain over a grid of coupling strengths, frequency gradients and gain pairs. On it we compute $M$, the Floquet margin, the Kuramoto margin and CoT, and we measure survival under four perturbation families: impulsive kicks, transection at each joint, sensor dropout, and lateral flow gusts. The decisive check is rank correlation between each margin and held-out survival, with families held out one at a time. A positive result requires $M$ to clearly beat the Floquet and Kuramoto ablations. It also requires most observed failures to be localized phase slips, with transection slips appearing at the cut. Next, we compare fronts of CoT against held-out survival from margin-guided continuation, black-box Bayesian optimization and domain-randomized reinforcement learning at matched simulator time. We also report wall-clock time with and without the symmetry quotient.

If an AgnathaX-type robot is available, a small hardware check would compare a few gain settings on the predicted front against off-front settings under physical transection. That experiment depends on platform access and the core claims do not rest on it. The result that would sink the proposal is a weak correlation between $M$ and held-out survival, or a finding that failures are not localized slips. Either outcome would mean that undulatory breakdown is not defect-mediated.