Can Humanoid Robots Work for 8 Hours? Battery Life, Heat and Charging Explained

Eight hours includes the stops

Can a humanoid robot work for eight hours? A shift can contain several battery cycles. None of the three specifications below establishes eight uninterrupted hours on one charge. Manufacturers describe charging or battery exchange for longer operation. Completed work still depends on the task, temperature and time spent recovering from failures. [1] [2] [3]

A powered robot waiting for a conveyor has runtime but no completed cycle. A second robot may cover a station while the first charges. These arrangements change the work delivered during the shift even when each machine has the same battery endurance.

G1, Atlas and Digit 5 specifications

Specifications checked on 8 October 2026. These are manufacturer records with different workloads and hardware stages. They are not an independent endurance ranking.

[1] [4] [2] [5] [3] [6]

Agility also advertises more than 20 productive hours per 24-hour day. Its launch announcement schedules Digit 5 early access for the first half of 2027 and general availability by the end of 2027. The company's existing Digit 4 customer operating records do not measure Digit 5 battery performance. [3]

The G1 charging figure comes from the 2024 manual. Its charger output differs from the product table. Confirm the supplied pack and charger revision. Atlas figures come from the sheet marked 23 December 2025; the accompanying product note identifies typical use. [4] [2] [5]

Where the electricity goes

Actuators

Joint motors convert electrical input into shaft motion. Mechanical power equals torque multiplied by angular speed. Electrical demand also includes winding resistance, bearings, gearing and drive losses. A peak joint torque rating alone cannot predict average battery consumption. [7]

Locomotion

Walking moves the body mass and swings the legs. Each step includes acceleration, deceleration and contact with the floor. Turning changes the motion of the body and feet. Climbing raises the centre of mass against gravity. A push can require extra steps and corrective joint torques. The route, speed and control policy therefore belong in an endurance test.

Posture also matters. Bent knees can require sustained joint torque even while the feet stay in place. Energy recovered during braking depends on the motor drive and battery system; assume no recovery credit without measurements.

Read how humanoid walking controllers manage contact and balance.

Manipulation

Lifting a box adds mechanical work. Holding it away from the torso increases the torque required at the shoulder. An electrically held gripper or joint can draw current while motionless. Mechanical brakes, springs and transmission design can change the holding demand. [7]

Computing

Control loops, perception and AI inference keep processors active. Hardware, model size and processing rate affect demand. The G1 product table lists an eight-core CPU; EDU configurations may add a computing module. Record the installed configuration when measuring runtime. [1]

Sensors

Cameras and LiDAR need electrical power while collecting data. IMUs report body motion and joint encoders report joint position. The G1 lists a depth camera and 3D LiDAR. Their individual power draw is not supplied in its product table. [1]

Cooling

Heat passes through motor housings and surrounding structures into the air. Fans or liquid pumps consume power where fitted. G1 specifies local air cooling. A cooling label alone does not establish a sustained workload at a given ambient temperature. [1] [8]

Calculate runtime from usable energy

Runtime (hours) = Usable battery energy (Wh) / Average electrical power (W)

1,000 Wh / 500 W = 2 hours
1,000 Wh / 750 W = 1.333 hours, or 80 minutes

Use energy available between the chosen start and stop limits, and average power measured at the same battery boundary. The 500 W figure must represent the whole intended job. Using a peak motor rating or a processor's power setting would leave other loads uncounted.

The G1 manual states 46.8 V nominal, 9,000 mAh, or 9 Ah, and 421.2 Wh. The arithmetic is 46.8 × 9 = 421.2. This is rated pack energy. It does not establish usable energy after discharge limits, temperature effects and ageing. Multiplying 9 Ah by charger output voltage would use the wrong voltage basis. [4]

Workload changes alter average draw. Temperature, discharge limits and electrical losses alter the energy that can reach the loads. Energy capacity also adds mass and occupies space. The March 2026 Energy Use perspective discusses these pack-level trade-offs and power-system architecture; it provides context rather than a new endurance experiment. [9]

Heat can limit the next lift

Current encounters resistance in a motor winding. Resistive heating follows P = I²R. Doubling current produces four times the resistive loss at unchanged resistance. A joint holding a box may have zero shaft speed while current continues to heat the winding. Friction in bearings and gears adds losses during movement. [7]

Picture repeated box lifts from a waist-height conveyor to a chest-height shelf. The first cycle starts with cooler hardware. Continued lifting can leave less time to reject heat before the next cycle. Winding temperature and the surrounding air affect the sustainable operating range. Peak torque is available only within the hardware's electrical and thermal limits. [8]

Thermal derating means reducing permitted output as temperature rises. ODrive documents a controller example that lowers its current limit between two temperature thresholds and stops closed-loop operation after further overheating. This is a documented controller behaviour, without implying that G1, Atlas or Digit uses ODrive. [10]

What the G1 arm experiment measured

A June 2026 preprint by Deniz and colleagues studied a physical G1 configuration with seven left-arm joints and a Dex3-1 hand. The robot stood with a safety tether and crane; the right arm rested. The authors fitted a power model to 897 filtered trajectories and tested it on 46 additional trajectories. Electrical readings arrived at about 1 Hz. [11]

Validation produced a root-mean-square prediction error of 3.58 W. The fitted model attributed 42.4% of predicted power to copper losses and 38.2% to viscous friction on that validation set. Those are model terms inferred from electrical measurements. They are not measured heat fractions for the whole robot. [11]

Charging and battery exchange

Manual replacement and powered swapping

A removable battery can be changed by a person. The G1 manual requires the battery to be off and removed for charging, and cooled to room temperature after operation. Cooling adds waiting time beyond the quoted charge duration. [4]

Hot swapping means the electrical system stays powered during replacement through a supported supply arrangement. A removable pack or an autonomous exchange does not, by itself, document that capability. Check the manufacturer's permitted procedure, retained state and restart requirements before assigning an exchange time.

Autonomous exchange and charging docks

Atlas is described as navigating to a station and replacing its own battery. Its approximately three-minute exchange leaves pack recharging to the charging system. A usable charged pack must be ready for the next visit. The published exchange time does not quantify customer availability over a complete shift. [2]

Digit 5 instead advertises autonomous dock charging. A 90-minute run followed by a nine-minute charge gives 90 / (90 + 9), or 90.9%, availability in a repeated ideal cycle. This calculation excludes travel, waiting, failed docking and other stops; it is not a measured task success rate. [3]

Contacts, retention and scheduling

Charging connections need correct seating and controlled power transfer. Boston Dynamics documents an interlock that closes before Spot Dock supplies power and disconnects as Spot rises. Spot is a quadruped. This example does not establish the construction of an Atlas battery station. [12]

For a battery exchange, ask how the system confirms the pack is retained and electrically connected before work resumes. The public Atlas sheet does not disclose its latch detection or contact sequence. The acceptance record should include failed insertion, incomplete locking and recovery procedures.

A fleet schedule also has to reserve chargers, spare packs and the route to the station. Two machines with equal runtime can lose different amounts of work when one has a nearby available dock and the other queues behind another robot. Include charger faults and staff response time in the availability log.

An eight-hour shift worked through

Assume an eight-hour shift, 1,000 Wh usable energy per full pack and a constant average battery draw of 500 W throughout powered periods. That gives 120 powered minutes per pack. Start with a full battery. Strategy A stops for 30 minutes to recharge. Strategy B exchanges a pack in three minutes, with a charged spare always ready. Both service intervals include all travel and handling in this simplified example.

The final powered interval ends at minute 480. Neither schedule includes a recharge after the shift. Strategy A uses 500 × 6.5 = 3,250 Wh at the battery boundary during powered operation. Strategy B uses 500 × 7.85 = 3,925 Wh. Charger input energy would also include charging losses, which are unspecified.

Strategy B assumes at least two compatible packs. With a 30-minute off-robot recharge, the depleted pack would finish before the next 120-minute powered interval ends, provided cooling and handling fit the stated service plan. Extra travel, a warm pack or a failed lock check reduces the availability shown.

Count accepted work after powered time

Now take Strategy A's 390 powered minutes. Reserve ten minutes for maintenance and twenty for recovery and operator intervention. These pauses stay inside the assumed 500 W powered budget. That leaves 360 minutes for attempts. Assume every attempt, including a failed attempt, takes 60 seconds.

Attempt time = 390 - 10 - 20 = 360 minutes
Attempts = 360 / 1 = 360
Assumed success rate = 95%
Expected accepted cycles = 360 × 0.95 = 342
Good-cycle-equivalent time = 342 × 1 minute = 5 hours 42 minutes
Expected accepted throughput = 342 / 8 = 42.75 cycles per shift-hour

The 95% rate is an assumption, and 342 is an expected count. The twenty recovery minutes are additional to failed-attempt time. Report actual intervention counts and completed cycles in a site trial. A slower cycle, extra cooling pause or lower success rate changes the output even with unchanged battery runtime.

Download the interval-by-interval schedule as CSV.

Download the calculation script for Node.js. It prints the assumptions, intervals and results without external packages.

Questions for a task acceptance trial

Use the same box, route, lifting height, pace and stopping rule across repeated trials. Record ambient temperature, starting charge, pack revision, software, installed computers and payload. Ask for the full run log alongside the quoted duration.

• What average and peak battery power were measured during this exact task? Which electrical boundary did the meter use?
• What starting charge, discharge cutoff, payload and motion cycle produced the advertised runtime?
• How much time was spent cooling, travelling, docking, charging and returning to the work area?
• How long did a complete battery exchange take, including lock checks, restart and failed attempts?
• Which task has been evaluated at a customer site on this exact hardware generation? What were the sample size and trial duration?
• How many attempts produced accepted work, and how many required a person to intervene?
• What temperatures and current limits were reached after repeated lifts? Did cycle time or available torque change?
• How many compatible spare packs and chargers are required for the proposed shift schedule?

Read the guide to humanoid machine safety and task risk assessment.

Read the Atlas hardware and control record.