8 Best Humanoid Robots in 2026: Price, Tasks and Availability Compared

Agility Digit is the strongest overall choice when the ranking rewards work completed in live commercial environments. Agility reports active deployments with GXO, Schaeffler, Toyota Motor Manufacturing Canada and Mercado Libre. Its June 2026 transaction announcement states that Digit had accumulated more than 65,000 hours of operation across deployment commitments at nine customer facilities. In July, the company reported that Digit had moved more than 100,000 totes at GXO. These are company-reported figures, but they describe sustained work rather than a single filmed sequence.

Figure 03 is the strongest general-purpose contender. Figure publishes a complete headline specification, says it has produced more than 350 third-generation robots and demonstrated a one-robot-per-hour manufacturing cycle. Its June 30 BMW demonstration combined part picking, placement, stepping and cart pulling under Helix 02 whole-body control. The main limitation is access: Figure does not publish a price or open sales channel.

Boston Dynamics Atlas leads the hardware specification. Its 56 degrees of freedom, continuous joint rotation, 30 kg sustained payload, 50 kg instantaneous capacity, four-hour battery and IP67 rating are unusually detailed for a humanoid product. Atlas ranks below Digit and Figure because its new product version entered customer deployment in 2026 and does not yet have the same published operational history.

The ranking uses six criteria. Deployment evidence carries the most weight, followed by manipulation, locomotion, documented autonomy, availability and documentation quality. A machine receives no credit for a specification copied from an earlier generation. A production target does not count as production. A customer agreement does not count as a completed deployment unless the source describes robots operating at the site.

Autonomy is scored task by task. A platform can navigate autonomously while relying on teleoperation for manipulation. A policy trained from human demonstrations can execute autonomously later, so teleoperated data collection does not make every later run teleoperated. The strongest evidence is an uncut run with the control mode, duration, resets, intervention count and task conditions disclosed.

Price is treated as a practical access signal, not a direct measure of capability. A $13,500 base robot without dexterous hands, development compute, service and safety equipment cannot be compared directly with a full enterprise contract. Quote-only systems may include integration, fleet software, support and maintenance that are absent from a store listing.

Digit is designed around material movement in warehouses, distribution centers and factories. Agility publishes a 35-pound carrying capacity and battery operation of up to four hours. The 2025 update added autonomous docking, a Category 1 stop, safety PLC, on-robot emergency stop, wireless teach pendant and Functional Safety over EtherCAT. Those details matter because a useful industrial humanoid needs predictable stopping behavior, fleet supervision and service support as much as dynamic walking.

The strongest evidence comes from GXO. Digit unloads totes from autonomous mobile robots and places them on a conveyor inside a live fulfillment operation. Agility describes the arrangement as a multi-year Robots-as-a-Service deployment. In 2026, Toyota Motor Manufacturing Canada signed a commercial agreement after a pilot. Agility also identifies Schaeffler and Mercado Libre as active customers.

Digit does not win every technical category. Its published payload is far below Atlas's instantaneous 50 kg figure. Public videos still cover a narrow set of material-handling workflows. Agility has not released a complete independent reliability study with failure rates, mean time between interventions and customer-verified ROI. Its lead comes from operating history and commercial integration, not from claiming general human-level work.

Figure 03 stands 5 feet 8 inches tall, weighs 61 kg and has a stated 20 kg payload, five-hour runtime and 1.2 m/s speed. The third-generation design adds a revised camera system, palm cameras, soft exterior materials, wireless charging and a 2.3 kWh battery. Figure says the camera architecture delivers twice the frame rate, one-quarter the latency and a 60 percent wider field of view per camera than the previous design.

The robot's control stack is Helix. Figure describes Helix 02 as a vision-language-action system that controls hands, arms, torso and feet from visual input. In the BMW sequencing workflow published June 30, Figure 03 grasped thin parts, stepped to adjust reach, placed parts into slots and pulled a wheeled metal cart. The demonstration is more informative than a stationary pick-and-place clip because it combines manipulation with balance and repositioning.

Figure also publishes unusually specific manufacturing information. On April 29, the company said BotQ had delivered more than 350 Figure 03 robots, raised throughput from one robot per day to one per hour and produced more than 9,000 actuators. The cycle time shows manufacturing capability, not demand or customer deployment. Figure 02's earlier BMW results, including 1,250 runtime hours and more than 90,000 parts loaded, cannot be assigned to Figure 03.

Figure ranks second because its current model has limited public field history. The BMW sequencing video confirms a physical workflow at the plant, but Figure has not published a randomized task benchmark, intervention rate or shift-long Figure 03 performance table. Pricing and procurement terms remain private.

The product version of electric Atlas is 1.9 m tall and weighs 90 kg. Boston Dynamics publishes 56 degrees of freedom, a 2.3 m reach, continuous joint rotation, tactile sensing, a 360-degree camera view and an IP67 rating. It is rated for a 30 kg sustained load and a 50 kg instantaneous load. The stated operating range is minus 20 to 40 degrees Celsius.

Atlas is designed for continuous industrial work. A typical battery lasts four hours and the robot can swap its own battery in less than three minutes. Boston Dynamics says Atlas can run autonomously, under direct teleoperation or through a tablet steering interface. Orbit connects the fleet to manufacturing and warehouse systems, including MES and WMS software.

Boston Dynamics began manufacturing the product version in January 2026. The company said all 2026 deployments were committed, with fleets scheduled for Hyundai's Robotics Metaplant Application Center and Google DeepMind. That is stronger than an undated prototype promise, but it is still an early deployment stage. Public pricing, fleet size and completed customer task metrics are not available.

Atlas ranks behind Digit and Figure because published hardware is not the same as operating history. Its mobility and load specifications are exceptional, yet the new generation needs public evidence on cycle time, uptime, intervention frequency and maintenance in customer production.

Unitree G1 is the easiest robot in this ranking to price and order. The official base price is $13,500 before tax and shipping. The standard robot stands 1.32 m tall, weighs about 35 kg, has 23 degrees of freedom and runs for about two hours. Its sensing package includes a depth camera, 3D LiDAR, four-microphone array and speaker.

The G1 EDU configuration expands the platform to as many as 43 degrees of freedom. Options include additional waist joints, wrist joints, higher onboard compute and the Dex3-1 three-finger hand. Unitree states that each optional hand has seven active degrees of freedom and can include tactile sensor arrays. The EDU price is quote-only.

G1 is a development platform rather than a finished autonomous worker. A customer may run a remote controller, scripted sequence, reinforcement-learning locomotion policy, imitation-learned behavior or a higher-level autonomy stack. A viral boxing, dancing or tool-use clip does not identify the control method unless the publisher discloses it.

The main buying risk is configuration mismatch. The public $13,500 price is not a reliable budget for a complete manipulation laboratory. Researchers must add hands, compute, spare batteries, shipping, import duties, fall protection, safety procedures, integration time and local support.

Walker S2 is built for industrial material handling. UBTECH states that the robot can lift 15 kg, work across a 0 to 1.8 m vertical workspace and rotate its waist through plus or minus 162 degrees. The head uses binocular RGB cameras with learned stereo depth estimation rather than a listed LiDAR.

Its defining feature is energy management. Walker S2 uses two batteries and can continue operating on one while removing and replacing the other. UBTECH says the complete autonomous swap takes less than three minutes. The robot can choose between charging and swapping according to task priority and battery state.

UBTECH reports that mass production and delivery began in November 2025. The company also describes automotive factory programs, but the Walker S2 product page does not publish customer-level uptime, task success, intervention rate or a public price. The battery swap is mechanically important, yet continuous power availability does not guarantee continuous productive work.

NEO is the only robot in this ranking with a clear consumer ownership offer. Early Access costs $20,000 and includes ownership, a three-year warranty, premium support and priority delivery. The alternative subscription is $499 per month. A $200 refundable deposit secures an order. 1X says U.S. deliveries began in 2026 and other markets are planned from 2027.

The control model is explicit. NEO arrives with basic autonomy and uses 1X's Redwood generalist model to learn and repeat tasks. For a chore the robot does not know, a 1X expert can remotely supervise its actions during a scheduled session. This Expert Mode may complete useful work and generate demonstrations for later learning, but it is human assistance rather than independent autonomy.

The home design uses a soft exterior, voice interface, mobile application and self-charging behavior. Those features make NEO easier to imagine around people than an exposed industrial machine. They do not answer the key reliability questions: how often an expert is required, how long chores take, what the failure recovery process looks like and which rooms or objects remain outside the operating envelope.

NEO ranks sixth because it is a genuine product program with pricing and delivery terms, while household performance remains early. Buyers should review privacy controls, camera access, remote-session scheduling and the exact capabilities included at delivery.

Apptronik introduced Apollo 2 on June 30, 2026 in bipedal and wheeled-base configurations. The company uses fleets at its Austin Robot Park, other Robot Park locations and customer sites to collect task data. The program is being developed with Google DeepMind to train future humanoid intelligence.

This is a credible route to broader skills because repeated data collection can expose a robot to different objects, placements and operator strategies. It is not the same as a commercial deployment benchmark. Apptronik has not published a current Apollo 2 specification sheet with height, weight, payload, runtime, hand configuration, intervention rate or sale price.

Apollo 2 ranks seventh because the company has a strong hardware and AI partnership base but less public task-level evidence than the leaders. The wheeled version may be practical for flat industrial floors where stability, runtime and data throughput matter more than stairs.

NEURA lists 4NE1 Gen 3.5 at 1.80 m tall and 80 kg, with a stated speed of 5 km/h. The platform includes high-dexterity hands, Wi-Fi 6, Ethernet, Python, ROS 2 and C++ interfaces. Buyers can access teleoperation tools, a digital twin and NEURA Gym training.

The commercial terms are unusually specific for a full-size European humanoid. NEURA estimates €98,000 per unit for orders of one to nineteen and €60,000 per unit for orders of twenty or more, excluding tax and shipping. A refundable €100 reservation fee secures a place in the queue. Availability is expected at the end of 2026.

A reservation and estimated price do not prove a final production configuration. NEURA states that hardware and aesthetics may change before release. Broad claims about multimodal interaction, transportation and manipulation need task success rates, operating duration and customer deployment evidence before they can support a higher ranking.

Tesla Optimus remains one of the most visible humanoid programs, and Tesla's official AI page describes the goal as a general-purpose autonomous biped for unsafe, repetitive or boring tasks. The same page identifies balance, navigation, perception and physical interaction as active engineering problems.

A ranking needs a named current generation, specifications and evidence. Tesla does not publish a current Optimus product page with height, weight, payload, runtime, degrees of freedom, sale price or delivery terms. Public clips also do not consistently state whether each task was autonomous, supervised, teleoperated or selected from multiple attempts.

Optimus may become competitive through Tesla's manufacturing, batteries, inference hardware and vision software. Those advantages remain a thesis until Tesla publishes current hardware data and comparable operating results. Excluding Optimus from the numerical ranking avoids turning production goals or executive statements into measured performance.

A robot lifting one box does not establish its sustained payload. Running in a laboratory does not prove safe navigation around workers. Folding one shirt does not establish a general laundry capability. A continuous video is more useful than a montage, yet it still needs the control mode, trial count and object setup.

Teleoperation signs must be interpreted carefully. Slow precise movement, latency, a VR headset, operator gloves, a camera feed, repeated trajectories or cuts can suggest human involvement. None is conclusive alone. Learned autonomous policies are often trained with teleoperated demonstrations, and remote supervision can intervene only after a failure.

The missing metric across most humanoid programs is intervention rate. Buyers need tasks completed per hour, failure categories, recovery time, uptime, battery and charging losses, maintenance intervals and the number of human minutes required per robot hour. Without those values, polished demos remain engineering evidence rather than an ROI case.

Request the exact model and configuration in writing. Confirm hands, wrists, onboard compute, cameras, LiDAR, batteries, charger, SDK access, ROS support, warranty, spare parts, remote assistance, safety functions and delivery territory. Ask whether payload means one arm, two arms, instantaneous lift or sustained carrying while walking.

For a pilot, define one workflow and measure it before adding more tasks. Record success rate, cycle time, interventions, resets, falls, damaged objects, charging time, maintenance and operator training. A robot that completes a narrow task for an entire shift may be more valuable than a more athletic machine that needs frequent recovery.