9 Open problems
This chapter collects what the energy metric settles, what it leaves open, and where the current analysis is limited. The research roadmap turns these into threads with concrete next steps, and research notes record progress on them.
9.1 What energy whitening does and does not capture
- Retention is settled. For the question “how much does the body itself hold about the world,” energy is the unique admissible resolution (Theorem 7.1). Every information quantity built on it is coordinate-free, composes across components, converges in constraint and refinement limits, and changes only through work and dissipation (Proposition 7.1).
- Routing is what morphology does. In a thermodynamically consistent body the total is capped by the dissipation at the contact (Proposition 8.2). Designs differ in how much of that total reaches the coordinates a controller can read, at a latency it can use. The efficiency in Table 8.1 is a candidate scalar measure of mechanical intelligence in this sense, and the temporal and frequency decompositions of Part I apply to it unchanged.
- Relevance is not captured. Neither retention nor routing asks whether the information is about what the task needs. The conditional information \(I(W;X\mid A)\) of Patterson (2026) measures relevance to the action, and task-specific variables, such as the force component that matters for balance, would sharpen it.
- Nonlinear and multistable bodies. The framework extends directly: the noise law becomes the Gibbs law \(\exp(-V/T)\) on a non-quadratic potential. A mechanical memory then stores a reliable bit when its energy barrier is large compared with the noise energy, the classical criterion behind Landauer’s bound (Landauer 1961). This connects the analysis to mechanical metamaterials and latching mechanisms, though none of it is tested yet.
- Validation must be predictive. A definition is checked by its consequences. The consistency checks pass: invariance, the limits, and agreement with the bandwidth argument of Wensing et al. (2017). The empirical test is whether retention, routing or relevance best predicts closed-loop performance across designs; Table I of Patterson (2026) gives a first hint.
9.2 Limitations
- Body versus sensor. The energy noise model quantifies what the body retains. Real encoders have per-coordinate noise, and an encoder readout answers a different question. The two should be reported as distinct quantities; Chapter 8 shows how to combine them consistently.
- The environment lives in the metric. \(K\) includes the ground spring, so \(\Em\) differs between stance and flight, and the ground stiffness can flip the sign of the SEA-rigid comparison (the analysis library shows this in
test_the_ground_decides_whether_series_compliance_helps_or_hurts). That is physically right, since the ground does store energy, but any result should say so. - Dissipation is not stored energy. \(D\) is absent from \(\Em\), as it should be. Materials with internal variables, such as viscoelastic Maxwell elements, add states whose stored energy has to be defined first.
- Nonlinearity and contact. \(H(q)\) varies, contact switches \(K\), and gravity preloads the stiffness. Remark 2.2 outlines a trajectory-wise extension. Hybrid contact needs a metric per contact mode and a rule at impacts.
- The noise scale. \(\sigma\) is a resolution, not a temperature. Comparisons do not depend on it only when the designs excite the same number of modes (Proposition 4.3(c)), so ranks should be reported alongside bits.
- Two metrics, two questions. \(I(W;X)\) measures transmission and \(\mathrm{MC}_W\) measures retention (Proposition 5.2), so the two should not be expected to agree.