When the computer becomes the load
A humanoid's energy is usually argued as a battery problem, because for a decade the legs dwarfed everything else. That is ending. As locomotion efficiency improves the compute module stops being a rounding error and becomes a first-order term, and the crossover is computable from numbers already published.
The machine and its gait
Cost of transport is dimensionless: CoT = P / (m·g·v). It is the honest way to compare a 30 kg robot with a 70 kg human, and the one number no humanoid vendor publishes.
Reference points on the CoT slider, all measured and published: 0.19 Cornell Ranger, 0.20 human walking, 0.25 RAIBO2 quadruped over a marathon, 0.41–0.42 humanoids reported in 2026 (DUKE; lower-limb structural optimisation), 1.61 DURUS, the best full-size humanoid figure the earlier literature carried.
The same machine, three eras of leg
Only the gait changes. The computer is held fixed, which is the point: nothing about the brain improves, and its share of the bill triples.
| Cost of transport | Locomotion | Compute | Compute share | Pack hours |
|---|
What this is. A first-order power ledger, not a robot simulator. Locomotion power is CoT·m·g·v by definition; standing, sensing and compute are held as flat draws over the shift; no regeneration is credited, and no thermal derating is modelled. Pack figures are vendor-published capacities. What it is for. Every term here is measured and published somewhere, and no document anywhere puts them in one ledger for one machine over one shift. That absence is the subject, and the crossover it hides is the finding: an efficient body makes the computer the thing to argue about. Part of the Hall Lab × Charlot Lab track on energy reporting in Physical AI.