Humanoids and Legged Robots
Why Humanoid Robots Run Out of Power in Two Hours
The energy arithmetic behind humanoid runtime: standing costs power, walking costs more, and a pack that would fix it weighs more than the robot can carry.
A humanoid spends energy just standing up. Unlike a wheeled robot, which draws close to zero at rest, a legged machine holds its posture against gravity through joint torque, and that costs 100 to 400 W before it moves anywhere. Add locomotion and manipulation and typical draw reaches 500 W to 1.5 kW, which is why a 1.5 to 3 kWh pack yields the 2 to 5 hour runtimes the industry actually publishes.
Why standing costs anything at all
A wheeled robot at rest transfers its weight through bearings into the floor and draws only what its electronics need, typically 20 W to 60 W. A legged robot supports its mass through actuators. Unless the joints are geared to be non-backdrivable, or fitted with brakes that can be engaged while standing, holding a pose means holding torque, and holding torque in an electric motor means current flowing into a stalled winding, which becomes heat rather than motion.
Design choices change the number substantially. High gear ratios reduce standing current but hurt backdrivability and impact tolerance. Quasi-direct-drive designs, favoured for dynamic balance and safe contact, are efficient in motion and expensive at rest. Most humanoids sit between the two, and their standing draw reflects it.
The energy budget of a working shift
| Activity | Share of time | Power | Energy per hour |
|---|---|---|---|
| Standing, waiting | 30 % | 280 W | 84 Wh |
| Walking unloaded | 25 % | 650 W | 163 Wh |
| Walking loaded, 15 kg | 20 % | 950 W | 190 Wh |
| Manipulating | 20 % | 520 W | 104 Wh |
| Compute and sensing | 100 % | 120 W | 120 Wh |
| Thermal management | 100 % | 45 W | 45 Wh |
| Average draw | ≈ 706 W | 706 Wh | |
| Runtime from 2.4 kWh | ≈ 3.4 h |
Two lines deserve attention. Compute and sensing at 120 W runs continuously and consumes 17 % of the budget, which is why on-robot inference hardware is chosen for efficiency rather than peak performance. And standing at 30 % of the time still consumes 12 % of the energy, so idle time is not free the way it is for a wheeled machine.
What actually extends useful uptime
- Opportunity charging. Docking during natural waits, exactly as mobile robot fleets do. It converts idle time into charge time and is the single most effective measure.
- Hot swap packs. A 60 to 120 second exchange restores full capacity, at the cost of spare packs and a handling step.
- Brakes for static poses. Engaging a holding brake instead of holding torque removes most of the standing draw, and it is a design decision rather than an operational one.
- Task sequencing. Grouping loaded walking into fewer, shorter trips reduces the most expensive activity in the budget.
- Efficient compute. Moving perception to a lower-power accelerator can reclaim 40 W to 80 W continuously, which is worth 10 to 20 minutes of runtime.
Why this does not improve quickly
Runtime is limited by cell energy density, which improves at roughly 3 % to 7 % per year across the industry. At that rate, doubling takes a decade or more. Efficiency gains in actuators and compute are faster in the short term, but the fundamental asymmetry with wheeled robots remains: legs pay a standing cost that wheels do not, and no battery chemistry on the horizon removes it.
The practical conclusion for anyone planning a deployment is to design around the charging pattern from the start, treating the robot as a machine with a duty cycle rather than as a worker who happens to need a break.
Frequently asked questions
How long does a humanoid robot run on one charge?
Two to five hours in published specifications. Compact research humanoids sit near two hours, full-size logistics machines around four, and one commercial model claims about five.
Why do humanoids use power while standing still?
Because the legs hold the body against gravity through joint torque rather than through bearings. Current flows into stalled motor windings and becomes heat, typically 100 W to 400 W depending on gearing and whether holding brakes are used.
Why not fit a larger battery?
At around 200 Wh/kg, doubling a 2.4 kWh pack adds roughly 12 kg that the legs must carry, which raises locomotion power in proportion to mass. A large share of the added capacity is consumed transporting itself.
How much power does the on-board computer use?
Commonly 80 W to 200 W continuously, which in a typical budget is 12 % to 20 % of total consumption. That is why efficient inference hardware matters more than peak throughput on a battery-powered robot.
Will battery improvements fix this?
Slowly. Cell energy density improves at roughly 3 % to 7 % a year, so doubling takes a decade or more. Near-term gains come from actuator efficiency, holding brakes and charging strategy rather than from chemistry.
Sources
- Unitree G1 entry in the ROBOTS guideIEEE Spectrum, published runtime and mass figures
- IEC 62619, safety requirements for secondary lithium cells for industrial applicationsInternational Electrotechnical Commission
- The Robot Report, humanoid hardware coverageReporting on published runtime and charging strategies