Humanoid Robot Batteries: Systems, Capacity & Runtime

Humanoid robots use lithium-based battery packs combined with battery management, protection and a charging or swapping arrangement designed for the robot. Many manufacturers do not disclose the precise cell chemistry. Public information more often covers capacity, runtime and energy replenishment. Choosing the “best battery” therefore starts with the job: walking, heavy lifting and stationary sorting place different demands on the same machine.

For factories, laboratories and homes, the useful question is how much productive work the robot completes between charging or replacement events. This guide connects battery fundamentals with a source-checked comparison of Figure, Unitree, Atlas, Walker S2, Apollo, Digit, NEO and Optimus.

Data checked: 9 September 2026. English edition of the substantially revised German article first published on 12 February 2026. Manufacturer claims, editorial interpretation and illustrative calculations are distinguished. No independent runtime tests were conducted.

Key findings

Capacity is not runtime: usable watt-hours and average power demand during the task determine endurance.
Cell chemistry is often undisclosed: a lithium-battery label or photograph does not establish NMC, LFP or a cell supplier.
Claimed hours are not a ranking: the product specifications reviewed do not share a common workload.
Swapping and charging extend operation: they do not increase the energy stored in one pack.
24/7 requires infrastructure: spare batteries, charging stations, routing and maintenance all matter.

What batteries do current humanoid robots use?

The comparison includes only specifications attributable to the linked manufacturer sources. “Not specified” means the reviewed source does not provide the figure. It is neither a zero nor proof that the information is inherently confidential. These are technical specifications for different product generations, not a list of immediately available products.

Battery capacity and runtime: manufacturer claims, checked 9 September 2026
Model / reference version	Published pack data	Claimed runtime	Qualification / source
Figure 03 / F.03 (2025)	2.3 kWh	5 h	Manufacturer claim; no common load test [1]
Unitree G1	13S lithium pack; 9,000 mAh	about 2 h	Do not derive Wh from charging voltage [3]
Unitree H1 / H1-2	15 Ah; 864 Wh; maximum 67.2 V	not established here	Maximum voltage is not nominal voltage [4]
Electric Atlas / sheet dated 23 Dec 2025	Energy not specified	4 h; 2 h with heavy lifting	Two different load conditions [5]
UBTECH Walker S2	Dual-battery system	Single-charge runtime not established here	24/7 refers to operation with replenishment [6]
Apptronik Apollo / 2023 introduction	Swappable pack; Wh not specified	4 h	Historical product version [7]
Apptronik Apollo 2	Swappable packs; Wh not specified	Not specified per charge	Do not transfer first-generation Apollo figures [8]
Agility Digit / innovation announcement	Wh not specified	up to 4 h	Separate runtime from recharging [9]
1X NEO / current product page	842 Wh	4 h	Not NEO Beta or Gamma [10]
Tesla Optimus / current AI page	Not specified	Not specified	Current-version battery data remain open [11]

What the comparison actually tells us

Disclosure is uneven. Some suppliers publish watt-hours, while others provide only operating time or a replacement method. A larger advertised battery does not by itself make a robot more efficient. Body size, payload, movement and computing hardware also matter. Missing data are relevant to procurement: the value needs to be established through the specific offer or acceptance test.

Figure 03 has separate battery-development and product-introduction sources. The early Apollo announcement must be distinguished from Apollo 2. The Atlas sheet concerns the electric platform. Older hydraulic demonstrators and previous Figure generations must not silently inherit those specifications.

Optimus needs particular care. Tesla’s current AI page contains no battery specification usable for this comparison. Historical presentation figures, vehicle cells and patent drawings cannot substitute for a current datasheet. The table therefore avoids an apparently precise estimate. [11]

What makes up a humanoid battery system?

A cell stores energy electrochemically. Cells are connected into groups and modules; the enclosure, interconnections, monitoring and protection turn them into a battery pack. Integration determines accessibility, heat removal and the mass the robot has to carry.

Cells / modulesPack + BMS + protectionPower electronicsActuators, sensors, computers

Simplified functional diagram. Charging and thermal management complement the energy path. The BMS monitors the pack; it is neither the main load nor a complete charger. This is a generic diagram, not the internal layout of a particular robot.

Series connections increase voltage; parallel connections increase charge capacity for otherwise equivalent cells. “13S” denotes 13 series-connected cell groups. It does not identify the number of parallel cells or their supplier. This distinction matters when specifications change between robot generations.

The battery management system, or BMS, monitors quantities including voltage, current and temperature. It supports protection, cell balancing and estimates of charge and battery health. Texas Instruments describes these functions in its humanoid-robot system overview. The semiconductor solutions shown there do not establish which components any particular robot manufacturer uses. [12]

Lithium-ion, NMC, LFP and solid-state: different levels of description

Lithium-ion is a family of rechargeable batteries. NMC, also written NCM, refers to a nickel-manganese-cobalt cathode; NCA to nickel-cobalt-aluminium; LFP to lithium iron phosphate. These are material choices. Cylindrical, prismatic and pouch describe cell formats. A flexible pouch is not evidence of a particular cathode chemistry.

Material selection involves trade-offs. Nickel-based variants are attractive for energy density; LFP is used partly for its safety and cost characteristics. The U.S. Department of Energy’s general assessment does not establish that any robot in the comparison uses LFP or NMC. That requires product-specific evidence. [13]

Silicon-rich anodes concern the other electrode and can be combined with different cathodes. Silicon and NMC are therefore not mutually exclusive categories, as Argonne research illustrates. Solid electrolytes concern the medium through which lithium ions move between electrodes. A solid-state battery can still be lithium-based. [14] [15]

Current capability, temperature behaviour, service life and complete pack design also matter. A high-energy cell may require additional cooling or protective mass once installed. Comparisons must distinguish cell energy density from pack energy density.

Why does lifting reduce battery life?

A humanoid moves its own mass, stabilises its body and performs tasks with its arms and hands. Acceleration, repeated standing up and heavy lifting change actuator loads. Sensors, communication and perception and control computers draw power too. A stationary robot is not necessarily a low-power robot: active holding and running computers may continue consuming energy.

The effect depends on the design and task. Atlas specifications separately list general operation and heavy lifting. That is more informative than a single figure, but still not a common cross-manufacturer test. [5]

Reading Wh, Ah and W correctly

Watt-hours (Wh) and kilowatt-hours (kWh) measure energy. Watts (W) and kilowatts (kW) measure power. Ampere-hours and milliampere-hours measure electrical charge. A rough conversion is nominal voltage multiplied by Ah, giving Wh. Actual deliverable energy also depends on the discharge curve and operating limits.

Maximum pack voltage and charger ratings cannot replace nominal voltage in that calculation. Multiplying the G1’s 9,000 mAh by charging voltage would create a misleadingly exact energy figure. Peak power is likewise different from average electrical consumption during a job.

A transparent runtime calculation

Runtime in hours ≈ usable energy in Wh ÷ average electrical pack power in W.

Consider an explicitly hypothetical 1,000 Wh pack. An 80% usable operating window provides 800 Wh. At an average 400 W, calculated runtime is two hours; at 800 W, it is one hour. None of these values describes a robot in the table. The simplified model does not separately account for temperature, ageing or load-dependent capacity changes.

More efficient actuators, suitable motions and better task scheduling can reduce demand. Software improvements nevertheless need assessment on the same task. Moving more slowly or carrying less does not automatically deliver more productive work per charge.

Battery replacement, hot swap and autonomous charging

A removable battery initially describes accessibility and servicing. Hot swap means replacement without fully shutting down the system functions designed to remain powered. “Autonomous” describes who performs the action. A robot may replace its own pack while interrupting particular functions. The terms answer different questions.

Charging and battery-swapping approaches
Approach	Example	Published information	Operational implication
Removable battery	Unitree G1; H1	G1: quick release; H1: quick replacement [3] [4]	Does not establish hot-swap capability
Autonomous battery swap	Atlas	3 min swap; 1.5 h charging [5]	Requires charged spare packs and a station
Autonomous hot swap	Walker S2	Two batteries; swap within 3 min [6]	Power continuity does not mean simultaneous productive work
Contact charging / docking	Digit; NEO	Digit docks; NEO plugs itself in [9] [10]	Charging access must fit the workflow
Inductive charging	Figure 03	Coils in the feet; 2 kW [2]	Charging periods must fit the task
Swapping and opportunity charging	Apollo 2	Swappable packs, opportunity charging, tethering [8]	Multiple supply options, not 22 h from one charge

Inductive charging avoids manually inserting a connector but requires appropriate alignment. Contact charging requires reliable mechanical and electrical connections. Swapping moves part of the charging time onto spare packs, with additional storage, charging capacity and handling requirements.

Fast-charge statements also need context. 1X’s six minutes of charging per hour of NEO runtime is not a verified zero-to-full charging time. Without initial state of charge, temperature and load conditions, it does not define a universal charge curve. [10]

Our Walker S2 battery-swap demo analysis examines what the visible demonstration establishes. A successful swap and reliably sustained shift operation require different evidence.

Can a humanoid work a full shift?

An operating plan can cover longer shifts even when one battery lasts less time. It must organise charging breaks, replacement or task handover to another robot. Whether that works depends on the entire process. A 24/7 claim is not 24 hours of work from a single battery.

A simplified energy-related working fraction is working time ÷ (working time + energy-related interruption). Hypothetical two-hour work periods followed by three-minute swaps give approximately 97.6%. Two hours of work followed by 90 minutes of charging give approximately 57.1%. These are illustrative calculations, not measurements of Atlas or Walker S2.

The swapping case assumes charged spare batteries are ready. Travel, queues, temperature checks, restart and failed attempts are excluded, along with maintenance or gripper faults. This is neither total plant availability nor an economic return calculation.

What fleet scheduling still has to solve

A process can continue when one robot charges and another takes over. But fleet size cannot be prescribed as two, five or ten robots. It depends on work cycles, recharge duration, handovers, reserves and required throughput. Extra robots are not a free substitute for efficient hardware.

A more useful operational metric is successfully completed tasks per kilowatt-hour and per shift. Longer endurance helps little if motion becomes substantially slower or people repeatedly need to intervene. Battery specifications therefore need to be evaluated alongside task completion and process reliability.

Safety, cooling and ageing are system properties

A moving robot exposes batteries to both electrical and mechanical stress. Protection has to address cell monitoring, electrical disconnection, interconnections, enclosure and heat removal. A BMS alone cannot prevent every form of mechanical damage. [12]

Figure describes measures to limit thermal-event propagation and flames escaping the F.03 pack. That is a specific manufacturer description of pack architecture, not proof of fire immunity under all conditions. [1]

Tests must also be distinguished by scope. UN 38.3 addresses transport testing for lithium cells and batteries. Such evidence does not constitute comprehensive approval of the robot for working beside people. The tested object, applicable edition and demonstrated status all matter. [16]

Ageing affects planning: a pack may deliver less energy and show greater voltage drop under load. Cycle-life figures need a charge window, temperature, load and remaining-capacity criterion. The reviewed information does not provide a uniform basis for comparing lifetime or replacement cost across these models.

What could new materials and solid-state batteries change?

New anodes and electrolytes can open design options. A successful cell experiment does not yet establish reliable charging, high-current delivery or mechanical resilience in a complete robot pack. Translating materials into a serviceable system is a separate engineering task. [14] [15]

A useful technology announcement identifies the stage reached: material experiment, sample cell, prototype pack, integration into a named robot or documented operation of delivered machines. A general “robotics” announcement does not prove deployment in Optimus, Atlas or Figure. The model sources reviewed here do not substantiate a specific solid-state installation.

Even better cells leave the system question intact. More usable energy, efficient motion and appropriately scheduled charging can reinforce one another. Where regular pauses are available, charging integration may matter more than a striking cell-level figure.

Frequently asked questions

Do all humanoid robots use the same battery?

No. Pack size, connections, protection and replenishment differ. Shared lithium-based technology does not establish identical chemistry or interchangeable batteries.

Which humanoid has the longest battery life?

The reviewed figures do not support a fair ranking. Manufacturer hours without matching loads, payloads and end-of-test criteria are not an efficiency comparison.

What battery does Tesla Optimus use?

The current Tesla AI page does not provide usable pack specifications for the current model. Historical claims must remain attached to their respective generations; this guide does not claim current figures.

Can a robot keep working during a swap?

A suitable hot-swap design can keep functions powered. Arms used to replace a battery cannot simultaneously perform a different task. Power continuity is not productive continuity.

Why are milliampere-hours insufficient?

They do not include voltage. Watt-hours are more suitable for energy comparison; runtime additionally depends on average electrical power demand.

Is there a universal humanoid battery standard?

The product sources reviewed identify no common interchangeable pack across manufacturers. Matching voltage is insufficient: mechanical interfaces, contacts, protection and communication must also match.

How much does electricity cost per shift?

It depends on energy drawn from the outlet and the tariff, including charging losses. Hypothetical consumption of 2 kWh at €0.30/kWh costs €0.60. These are assumed figures for electricity only, excluding ageing, infrastructure and robot servicing.

Will solid-state batteries solve runtime limitations?

They are a development approach, not a universal answer. Evidence must concern the full pack in the actual robot, including charging, safety, lifetime and workload.

Method, limitations and sources

Alpha Bionic reviewed the linked manufacturer pages, technical documents and research sources on 9 September 2026. Tables report manufacturer claims. No robots were independently tested and no manufacturer interviews were conducted. Missing data remain explicitly open; calculations use stated assumptions.

An independent runtime comparison would define task, payload, software, computing configuration, temperature, battery age, charge window and termination criteria. Successful cycles, outlet energy and human interventions should also be recorded. Fraunhofer IPA works on standardised assessments of application-relevant humanoid characteristics; this table does not report results from those tests. [17]

Revision note: The earlier German article focused on short battery life and fleet strategies. Those subjects are retained and clarified. Broad or insufficiently supported figures have been replaced with documented sources, and older generations are distinguished from current platforms.

For citations: Keep the model version, evidence status and limitations attached to each figure. Suggested attribution: Nico Nuss / Alpha Bionic, “What Batteries Do Humanoid Robots Use? Systems and Runtime”, data checked 9 September 2026, linking to this article. The primary sources below allow verification of the underlying specifications.

Bewerte den Beitrag hier!

[Total: 0 Average: 0]

Nico Nuss

Website

Author Nico Nuss has been working on mobile computing and automation software since 2001. Drawing on his experience and strong interest in future technologies, he focuses on robotics and AI.