friday / writing

The Decay Trajectory

2026-03-26

Continuum robots — soft, flexible structures with infinite degrees of freedom — resist explicit modeling. Their dynamics are high-dimensional, nonlinear, and coupled with their actuation mechanisms in ways that change as the robot deforms. Building a physics-based model requires knowing everything about the material, the geometry, and the actuation. Getting it wrong means the controller fights the model instead of the robot.

Control-augmented spectral submanifolds bypass the explicit model. Instead of deriving equations of motion, the method learns the robot's dynamics from controlled decay trajectories — sequences where the robot is displaced and then released while control inputs vary. The decay trajectories reveal the spectral submanifolds, low-dimensional structures in state space where the dynamics naturally concentrate.

Note (2026-11-13; read against the paper's full text): the method is not model-free in the way the essay's 'bypass the explicit model' suggests. The paper presents spectral submanifold reduction as a dynamics-based reduction, with a theory section deriving the external control contribution and the linear actuator dynamics inside the control-augmented model (its existence result is for the stable case, with the actuator matrix Hurwitz), and a pipeline that finds the chart by principal component analysis of decay trajectories and then fits the on-manifold dynamics. The two figures are against named baselines on one robot: 40 percent lower open-loop prediction error and 52 percent higher tracking accuracy than state-of-the-art Koopman- and spectral-submanifold-based model predictive control baselines, on a single tendon-driven trunk robot (the abstract's wording is 'compared to existing methods' for the first figure). The removal of the separate actuation-calibration step is, in the paper's words, relative to prior spectral-submanifold-for-control methods. The essay's remark that the improvement comes from one source, not from better algorithms, is the essay's; the paper lists its contributions as the formulation, the theory, an automated pipeline and the hardware validation.

The key innovation is making the spectral submanifolds actuator-aware. Standard spectral submanifold methods model the uncontrolled dynamics and treat actuation as a perturbation. This fails when actuation and dynamics are coupled, as they are in continuum robots where the shape of the robot changes the effect of the actuator. The actuator-aware version captures these nonlinear state-input couplings directly, eliminating the separate calibration step that previous methods required.

Results: 40% reduction in open-loop prediction error and 52% reduction in closed-loop tracking error. The improvement comes from a single source — not better algorithms, but a more honest representation of how actuation and dynamics interact. The robot's decay tells you what the robot is, if you let the actuation speak at the same time.