Humanoid Robots

Humanoid robots use a body plan that resembles the human form, usually including a torso, arms, and legs. The practical argument for that form is compatibility. Much of the built world, including doors, stairs, shelves, tools, and workstations, was designed around human reach and mobility.

Why the form matters

A humanoid can potentially use spaces without rebuilding them around a special machine. That flexibility comes at a cost. Balance, fall safety, power consumption, dexterous manipulation, and reliability are difficult engineering problems. A wheeled base or fixed arm is often simpler when the task and environment permit it.

The capability stack

Useful humanoids combine locomotion, whole-body control, perception, manipulation, task planning, and safe interaction. The weakest layer can limit the entire system. Strong walking does not guarantee useful hands; a capable model does not remove payload or battery constraints.

How to evaluate a demonstration

Use the robot demo evidence checklist to inspect task scope, continuity, attempt counts, control modes, resets, timing, variation, failures, and operating evidence. Distinguish a research demonstration from a pilot and a repeatable deployment. Each can be meaningful, but they answer different questions.

The Boston Dynamics Atlas status audit applies that distinction to the product generation introduced in 2026. It keeps manufacturing, committed partner allocations, Hyundai field testing, and scheduled fleet shipments separate from recurring customer operation.

Randy Levine’s contributed editorial, Beyond the Demo, extends this framework across demonstrated capability, pilots, contracts, deployments, and recurring commercial operation.

Latest evidence boundary

XPENG now says its IRON manufacturing line is commissioned, more than 80% of core processes are automated, and one assembled unit walked off the line autonomously. This is a manufacturing milestone, not proof of mass production or customer deployment. XPENG still schedules mass production for the end of 2026, followed by launch and deliveries in 2027. The humanoid availability tracker separates the commissioned line from undisclosed capacity, yield, order terms, delivered units, and field reliability (CnEVPost).

BMW now confirms that Figure 02 worked in production for ten months and supported production of more than 30,000 BMW X3 vehicles. That customer-side evidence is stronger than a vendor demonstration alone. BMW describes Figure 03 as a separate next logistics-sequencing project, while Figure calls the published workflow its first demonstration; recurring Figure 03 operating metrics remain unpublished. Our Figure at BMW evidence audit keeps the retired generation’s production record separate from the successor’s current status (BMW Group, Figure).

Mitsubishi Motors and Highlanders have also signed an MOU to explore humanoid development, use at Mitsubishi facilities, and possible production at the Kyoto Plant. The early-2027 timing is under study, so this remains an announcement and feasibility plan, not an active pilot or manufacturing start (Mitsubishi Motors).

Agility Robotics now describes a three-stage deployment process for Digit: controlled solution setup, a cordoned customer-site proof of concept, and 90 days of production measurement before ongoing service. This is process evidence, not a published customer result. The humanoid Robots-as-a-Service buyer’s guide explains what the measurement window and contract still need to disclose (Agility Robotics).

Agility’s September 2026 S-4 adds more than 65,000 reported Digit v4 operating hours, nine facilities described as deployed or committed, planned Digit v5 pricing and cost assumptions, and historical financial statements. The Digit economics evidence audit explains why those figures need separate evidence labels and fleet denominators.

Those hours are not directly comparable with Figure 02’s 1,250-plus runtime hours or AGIBOT G2’s bounded factory tests. The humanoid robot operating-hours guide separates runtime, uptime, utilization, productive work, human assistance, and fleet-summed hours before they are used in a reliability or economic claim.

Where to continue

Start with the evidence-based comparison Humanoid Robots in 2026: What Is Actually Available, Piloting, or Deployed?. It separates announcements and demonstrations from customer pilots, commercial ordering, and documented production work across Digit, Apollo, Atlas, Figure, Phoenix, NEO, G1, and H1.

Then read What Is Physical AI?, compare the company directory, and use the glossary for terms such as whole-body control, teleoperation, and degrees of freedom.

Developers evaluating open training and rollout tooling can also read the LeRobot v0.6.0 practical guide, including its support for Unitree G1 data workflows, simulation evaluation, and intervention capture.

For hardware developers, the open-source humanoid reproducibility audit uses Bridge to separate a paper and video release from versioned CAD, bill of materials, calibration, control artifacts, licenses, physical trial denominators, and an outside build.