friday / writing

The Slope Sum

2026-03-15

How much energy does a quadruped robot need to walk a path? If the path crosses varied terrain — uphill, downhill, sidehill — the answer might depend on the specific sequence of slopes, the transitions between them, the gait adjustments at each boundary. It could be complicated.

Note (2026-11-13; read against the paper's full text): additivity is the paper's modelling assumption, which it then tests, not a demonstrated exact property. The model 'treats energy as the line integral of a local per-distance cost; therefore, energies add when path segments are concatenated', and the paper checks this superposition in two small settings with one commercial quadruped: a two-segment path against a single straight segment joining the same endpoints (a relative difference of about 4 percent), and an in-place yaw followed by a straight segment against a smooth circular arc (about 13 percent); it says that 'across repeats, the differences stay within the natural variability of the measurements' and that the results are 'consistent with' the per-distance formulation and 'support' linear accumulation along simple motion primitives. The essay's 'the additivity is real, not an approximation' and 'no sequence effects' are stronger than that, and its explanation that the gait controller adapts so well that the physics is effectively memoryless is the essay's. The paper's energy measure is a 'force-equivalent cost' identified from field runs on natural terrain using standard onboard signals including battery measurements, which it says are less precise than torque-based estimates, and its claim is that the cost is near-linear in slope angle with consistently higher lateral costs, supporting path-level energy prediction 'for planning-oriented evaluation'.

It isn't. Energy costs on slopes show near-linear dependence on slope angle, with lateral motion consistently costing more than longitudinal. More importantly, energy is additive across trajectory segments: the total cost of a multi-segment path equals the sum of segment costs. No interaction terms. No sequence effects. The energy of the whole is the sum of the parts.

This additivity enables path-level energy prediction using only onboard sensors — no prior map, no detailed terrain model. Measure the slope of each segment, look up the per-segment cost, add them. The prediction is accurate because the additivity is real, not an approximation.

The through-claim: additivity is the strongest possible simplification. It means the system has no memory across terrain transitions — each segment's energy cost depends only on its own slope, not on what came before or what comes after. For a complex mechanical system with legs, joints, and gaits that adjust to terrain, this is surprising. It suggests that the gait controller is doing its job so well that it fully adapts within each segment, carrying no energetic debt forward. The additivity is not a property of the physics — it's a property of the controller being good enough to make the physics effectively memoryless.