The same hidden computation is obvious on one robot and invisible on another — and no adversary chose which. The body did. Pick a morphology, pick where the meter sits, and read whether the work clears the floor.
A mobile robot navigating autonomously. What share of its electrical power do the motors draw?
Detection needs the work to clear the residual, not the total. The bar on the right is drawn against a null control — the same body doing the same job with nothing hidden — because a margin with no null is not a result.
| Body | Split | Source |
|---|---|---|
| Industrial arm | actuation-dominated | ISO/TS 25213 scope |
| G1 arm | mech 10.1%, copper 42.4%, viscous 38.2% | arXiv 2606.15915 |
| AMR autonomous | motors 16.6%, GPU 37.3% | arXiv 2511.20467 |
G1 residual σ = 3.58 W hold-out RMSE, 17.2 W
worst case, against a 121.7 W full scale. AMR whole-system prediction is reported
at >96% accuracy. The industrial-arm residual is not known to us — ISO/TS
25213's clause text is paywalled, so whether it even fixes a measurement point is unread.