Where is the energy reporting?
A humanoid carries its own supply, so every joule it spends is converted into runtime, fleet size, swap infrastructure and a thermal ceiling before it ever appears on a utility bill. The quantity that decides all of those is energy per unit of work, and no machine on the market states it. This topic asks why, and finds that the answer has changed: the measurement is arriving and the report is not.
Tasjah Hall, MS, Academic Dean · The Hall Lab, with Dean David Jean Charlot
A background survey compiled for this track concluded that embodied energy is unmeasured. Checked against primary sources, three of its central claims are wrong, and they are wrong in the same direction. Humanoid cost of transport is published at 0.41 to 0.42, roughly four times better than the figure the survey carried. A physics-based electrical power model for a commercial humanoid arm was identified in June 2026 from 897 measured trajectories. An ISO Technical Specification for measuring industrial robot energy consumption, led by ABB with eleven countries, was due for completion in August 2026. Energy is being measured, modelled per joint, and standardised.
What none of that produces is a report. A journal result describes one machine under conditions its authors chose; a specification describes how to measure and not what any given machine measured. So the binding constraint is disclosure rather than instrumentation, and that distinction decides everything after it: instrumentation problems are closed by better sensors, and disclosure problems are closed by an obligation, a format, and someone qualified to sign. That is why this work sits in a lab built on aviation practice, where a release is valid because a named person holding a current qualification attests to it and carries the consequence of being wrong.
The urgency is a number rather than a mood. Cost of transport is dimensionless, so locomotion power follows from a machine's mass and gait, and compute power for a current embodied platform is a published envelope. Hold the computer fixed and improve only the legs, and going from the older literature's best humanoid to a 2026 one raises the compute share of shift power roughly threefold. An efficient body promotes the computer to a first-order term, which means a ledger that prices actuation and omits compute stops being adequate exactly as the machines become good.
TR-2026-25, the substrate case ↗OER/1, the receipt schema ↗TR-2026-42, what sensing costs ↗
Watch the crossover.
Cost of transport is dimensionless, so the locomotion term follows from mass and gait. Move the gait between the reference points, all of them measured and published, and watch the computer go from a rounding error to the largest single line.
A background survey for this track concluded that embodied energy is unmeasured. Checked against primary sources in August 2026, three of its central claims are wrong: humanoid cost of transport is published at 0.41 to 0.42, a physics-based power model for a commercial humanoid arm was identified from 897 measured trajectories in June 2026, and an ISO Technical Specification for industrial robot energy consumption was due for completion this month. The measurement is arriving. What has not arrived is the report: a statement of energy per unit of work, at a declared grade of evidence, that a purchaser or an inspector can rely on. That makes the binding constraint disclosure rather than instrumentation, and disclosure problems are closed by an obligation, a format and a qualified signature rather than by better sensors. The engineering change is dated and available, because extending an existing measurement method from a mains-powered arm to a self-powered body is a scope decision rather than a research problem. What becomes possible is a machine that can state its own energy per task, which is the one thing a purchaser comparing two robots, an operator sizing a fleet, an insurer pricing a duty cycle and an instructor setting an examination all currently cannot get.
One of eight, and only one of them is physics. How we read a frontier →