Humanoid Robot Companies Compared: Tesla, Figure, Boston Dynamics, Unitree, 1X, and More

Humanoid robots in October 2026 split into three groups: hardware you can order, hardware you can only pilot, and hardware you can only watch. Specs matter less than which group a platform sits in. A 56-DoF robot you cannot buy teaches you less than a 23-DoF robot on your bench. This guide compares seven companies by availability, actuation, control stack, and integration path, so you can match a platform to a real project.

Quick Takeaways

• Only Unitree and 1X have public price lists. Unitree’s G1 is orderable at $13,500. 1X NEO costs $20,000 or $499/month, with a $200 refundable deposit.
• Atlas, Figure 03, and Apptronik Apollo are enterprise-only. Access runs through pilot agreements, not a store.
• Tesla Optimus has no public specs or orders. Treat every Optimus number as a company target, not a measurement.
• For ROS2 and RL research, buy Unitree. For factory pilots, talk to Figure, Boston Dynamics, Agility, or Apptronik.

Master Comparison Table

Specs below are approximate and drawn from manufacturer claims. Verify against current datasheets before you design around them.

Company-by-Company Analysis

Unitree: The Buyable Platform

Unitree shipped more than 5,500 humanoids in 2025, a 32.4% share of unit shipments according to its prospectus. Its store lists the G1 at $13,500 and orderable, the R1 from $4,900, and the H2 at $29,900 with checkout disabled. The H2 routes buyers through sales, so confirm delivery before you budget around it.

The company is also now public, trading on the Shanghai Stock Exchange since August 2026. For a buyer, that adds financial disclosure and some assurance the vendor will survive.

For engineers, the draw is the SDK. The G1 exposes low-level joint control over CycloneDDS, so you can bypass the stock locomotion controller and run your own policy at 500Hz or faster. The tradeoff is payload. The G1 is a research chassis, not a tote mover.

Boston Dynamics: Atlas

The electric Atlas product version debuted at CES in January 2026. Boston Dynamics said it would begin production at its Boston headquarters immediately, with 2026 deployments scheduled for Hyundai and Google DeepMind. The robot has 56 DoF and the highest published payload on this list.

Deployment is still early. Boston Dynamics has opened the Robotics Metaplant Application Center at Hyundai’s Georgia plant to train Atlas on automotive tasks. Hyundai’s timeline puts initial Atlas deployment at Metaplant America in 2028, starting with parts sequencing. The long-range target is 25,000 Atlas robots across Hyundai and Kia plants, backed by a facility building 30,000 robots a year.

Atlas has the strongest mechanical claims. Its AI stack is the least proven in public, though the DeepMind partnership targets that gap.

Figure AI: Figure 03 and Helix

Figure 03 pairs in-house hardware with the Helix vision-language-action model. Figure documented a demonstration at BMW, and Figure leads the group on valuation at roughly $39B and runs paid factory pilots. Treat third-party fleet counts and hourly billing figures with caution. Several aggregator pages repeat them without primary sourcing.

There is no public home program, price, or order page for Figure 03. Access means an enterprise conversation.

Tesla: Optimus

Optimus is the largest gap between narrative and public evidence. As of September 2026, Tesla has not revealed Optimus generation 3, and Musk said on the April 2026 earnings call that it would be unveiled closer to production start. Some blogs describe V3 as already shipping at volume. None of that is verifiable from Tesla’s own disclosures. Use Optimus numbers as targets only.

Tesla’s real advantages are vertical integration, in-house actuators, and automotive-grade manufacturing. Those matter at scale. They tell you nothing about what you can integrate this year.

1X: NEO

1X NEO is the only humanoid built home-first with public terms. Early Access costs $20,000, with a $200 deposit, or $499 per month by subscription. 1X announced full-scale production at its Hayward, California factory on April 30, 2026. Customer delivery status has been murky. As of mid-July 2026, no customer delivery had been independently verified. Recheck before ordering.

Autonomy is partial by design. 1X says NEO arrives with basic autonomy, while a remote expert can supervise tasks it does not yet know. That means camera data from your home leaves the house. Read the privacy terms before you commit.

Agility Robotics and Apptronik

Agility Digit has a commercial logistics deployment, per one 2026 industry roundup. Its design is deliberately narrow: a bipedal tote mover. Apptronik Apollo targets similar work, with pilot and task-data work at Mercedes-Benz and GXO sites.

Control Stack Foundations

Every platform above solves the same problem: keep a floating-base, underactuated body upright while the arms do useful work. The approaches differ in where they put the learning.

Balance Models

Most legged controllers start from the Linear Inverted Pendulum Model (LIPM). With constant CoM height z_c and gravity g:

omega       = sqrt(g / z_c)               # natural frequency (rad/s)
x_ddot_com  = omega^2 * (x_com - x_zmp)   # CoM dynamics
xi          = x_com + x_dot_com / omega   # capture point (m)

To stop, place the ZMP (zero moment point) at the capture point xi. For z_c = 0.7 m, omega is about 3.74 rad/s, so a 0.5 m/s CoM velocity puts the capture point roughly 0.13 m ahead of the CoM.

Joint-Level Control

At the lowest layer, most research humanoids run an impedance-style PD law with feedforward torque:

tau = Kp * (q_des - q) + Kd * (qd_des - qd) + tau_ff

Kp and Kd are per-joint gains. An RL policy typically outputs q_des at 50Hz, while the joint loop runs at 500Hz–1kHz.

Control Philosophy Comparison

Production systems layer these. A VLA proposes hand targets, a whole-body controller tracks them, and a fast joint loop enforces limits.

Integrating a Unitree G1 with ROS2

If you buy one humanoid for development, it is probably the G1. The unitree_ros2 package bridges the robot’s DDS topics into ROS2 topics. The node below reads the IMU and flags dangerous torso tilt. Confirm message and field names against your installed unitree_hg package, because they have changed between SDK releases.

import math
import rclpy
from rclpy.node import Node
from std_msgs.msg import Bool
from unitree_hg.msg import LowState  # message type for G1-class robots; verify in your install


class TiltMonitor(Node):
    """Publishes True on /fall_risk when torso tilt exceeds a threshold."""

    def __init__(self):
        super().__init__("tilt_monitor")
        self.declare_parameter("max_tilt_deg", 25.0)  # trip threshold in degrees
        self.max_tilt = math.radians(self.get_parameter("max_tilt_deg").value)
        self.pub = self.create_publisher(Bool, "fall_risk", 10)
        # 'lowstate' carries IMU + motor state at ~500 Hz on the robot's DDS network
        self.sub = self.create_subscription(LowState, "lowstate", self.on_state, 10)

    def on_state(self, msg: LowState):
        # Quaternion order on Unitree IMUs is [w, x, y, z]
        w, x, y, z = msg.imu_state.quaternion
        # Roll (rotation about x) and pitch (rotation about y) from quaternion
        roll = math.atan2(2.0 * (w * x + y * z), 1.0 - 2.0 * (x * x + y * y))
        pitch = math.asin(max(-1.0, min(1.0, 2.0 * (w * y - z * x))))
        # Combined tilt magnitude: small-angle approximation is fine for a safety trip
        tilt = math.hypot(roll, pitch)
        self.pub.publish(Bool(data=tilt > self.max_tilt))


def main():
    rclpy.init()
    rclpy.spin(TiltMonitor())


if __name__ == "__main__":
    main()

Launch it with:

# Match the DDS network interface to the Ethernet port wired to the robot
export RMW_IMPLEMENTATION=rmw_cyclonedds_cpp
ros2 run my_pkg tilt_monitor --ros-args -p max_tilt_deg:=20.0

Set the interface in your CycloneDDS XML config. A wrong interface is the most common reason ros2 topic list shows nothing.

Real-World Workflow: Scoping a Humanoid Pilot for Tote Handling

Suppose you must move 12 kg totes between a conveyor and a rack at 3 picks per minute. Work through these steps before you request a quote.

1. Check the payload envelope. Include the tote plus gripper and a 1.5x dynamic margin: 12 kg × 1.5 = 18 kg. That rules out the G1 and fits Figure 03, Digit, and Apollo only at the edge of their claimed ratings.

2. Budget the energy. Runtime equals battery energy divided by average power:

runtime_h = E_battery / P_avg
example:    0.5 kWh / 250 W = 2.0 h

A 4-hour vendor claim usually assumes light duty. Ask for a measured figure on your payload and duty cycle.

3. Plan for charging. Compute shift coverage. An 8-hour shift with a 2-hour runtime needs either swappable batteries or four charging windows.

4. Test the integration surface. Ask each vendor three questions:

• Does the API expose joint-level or only Cartesian commands?
• Can you run a custom policy, or only vendor-trained skills?
• Where does inference run, on the robot or in the cloud?

5. Run a safety review. Humanoids are dynamically stable only while powered. Define the protective stop behavior and the fall envelope. Check them against ISO 10218 and the emerging humanoid-specific standards work.

Which Platform Fits Which Job

Risks to Weigh Before You Buy

• Teleoperation dependence. Many “autonomous” demos hide a human operator. Ask what fraction of tasks run unassisted.
• Cycle life of actuators. A 4,000-hour actuator lifetime sounds long until you compute it: at 16 hours a day, that is about 8 months.
• Software lock-in. Vendor-trained skills may not transfer if you switch hardware.
• Delivery slippage. Humanoid timelines have slipped repeatedly across the industry. Contract on delivered units, not announced dates.
• Export and data rules. Chinese hardware may face procurement limits in some sectors, and cloud-connected robots raise data residency questions.

FAQ

Which humanoid robot can you actually buy in 2026?

The Unitree G1 is the clearest answer. It is orderable at $13,500 from the official store. 1X NEO takes preorders at $20,000, but verify delivery status first. Atlas, Figure 03, and Optimus are not open to general purchase.

Is Tesla Optimus available to buy?

No. As of September 2026, Tesla has not yet revealed Optimus generation 3. Tesla has not opened public orders or published verified specifications.

Which humanoid has the highest payload?

Boston Dynamics Atlas has the strongest published industrial specifications, with a peak payload near 50 kg. Figure 03 and Optimus are both cited near 20 kg. Peak payload differs from sustained payload, so ask each vendor for sustained numbers at full reach.

Which humanoid is best for ROS2 and robotics research?

Unitree platforms. They give low-level access over DDS, a published SDK, and the largest community of researchers. For deeper control work, pair the robot with a simulator such as MuJoCo or Isaac Sim, and train policies there before deploying to hardware.