You’ve read that the Unitree H2 has 31 dof and 360 Nm torque. Here’s what that actually means — and why it matters.
The robots has spec tables for every humanoid on this site. They’re full of numbers: DOF, Nm, Hz, kg. Useful if you already know what they mean. Useless if you don’t.
This piece is the decoder ring.
I’ll walk through the five metrics that actually tell you something about a humanoid’s capability — what they measure, what good vs. bad looks like, and how the robots on this site compare.
Degrees of Freedom (DOF)
What it is: The number of independent axes of movement a robot can make. More DOF = more ways the robot can move.
The human baseline: A human has roughly 244 dof total — joints in fingers, spine, wrists, shoulders, hips, feet. But that’s misleading for robotics comparison. Most humanoid robots count “actuated DOF” — the joints with motors and active control.
A human body with just the major joints needed for general movement — shoulder, elbow, wrist, hip, knee, ankle, and enough hand function for grasping — needs roughly 30-40 actuated DOF to replicate general capability.
How to read it:
| DOF Count | What it signals |
|---|---|
| <15 | Basic locomotion. Limited or no arm function. Research-grade mobility, not dexterity. |
| 15–25 | Arms + legs functional. Coarse manipulation. Can carry objects, push buttons, basic grasping. |
| 25–35 | High-function arms. Moderate hand dexterity. Can handle most industrial manipulation tasks. |
| 35+ | Human-equivalent or beyond. Fine manipulation, complex assembly, multi-finger grasping. |
On this site:
- unitree-h1: 19 DOF — locomotion-focused, limited upper body
- unitree-g1: ~20 DOF — adequate for combat demos, not precision manipulation
- boston-dynamics-atlas: not publicly specified, but wide range of motion from video evidence
- unitree-h2: 31 DOF — high-function arms and legs, capable of meaningful manipulation
- Figure 02: 16-DOF hands alone — exceptional dexterity for industrial assembly
The Figure 02’s hand DOF is striking. 16 DOF per hand is approaching human-level finger articulation. Most humanoids have 1-3 DOF hands (pinch grippers, not fingers). That’s why Figure 02 can do bmw-group factory work where other humanoids can’t.
Torque (Nm)
What it is: torque measures rotational force at a joint — how hard a motor can push or pull against resistance. Measured in Newton-meters (Nm).
Why it matters: A robot lifting a heavy object, maintaining a push-up against gravity, or delivering a punch all require torque. Low torque = weak joints that buckle under load. High torque = strength, but also heavier motors and higher power consumption.
The human baseline: A human knee can produce roughly 150–200 Nm of torque during walking. A human elbow, maybe 80–100 Nm in resisted motion. Peak athletic torque at the knee (jumping, sprinting) can exceed 300 Nm briefly.
How to read it:
| Joint Torque | What it signals |
|---|---|
| <50 Nm | Lightweight, low-load tasks. Gesturing, light object handling. Fragile in collisions. |
| 50–150 Nm | Typical locomotion range. Can walk, climb stairs, handle moderate objects. |
| 150–300 Nm | Athletic range. Can run, jump, exert significant force. Relevant for combat. |
| 300+ Nm | Superhuman force at specific joints. Industrial lifting, high-impact tasks. |
On this site:
- unitree-g1: 120 Nm — adequate for boxing demos; can absorb impacts but not deliver maximum force
- unitree-h2: 360 Nm — significant. Exceeds typical human athletic torque. This enables sustained heavy manipulation.
- Boston Dynamics Atlas: “unmatched strength” claimed, specs not disclosed — [UNCERTAIN]
The H2’s 360 Nm number is meaningful. That’s not “can lift a box” torque. That’s “can exert meaningful force in adversarial situations” torque. Whether it’s wired for autonomous decision-making to use it is a different question.
Height and Weight
What it is: Physical dimensions. Self-explanatory.
Why it matters more than you think: Humanoids are designed to work in human environments — doorways, vehicles, shelving, workstations — sized for ~170cm, ~70kg humans. A robot that’s too short can’t reach standard shelving. Too tall and it can’t fit in standard doorways. Weight matters for floor load capacity in industrial settings, and for shipping, maintenance, and safe human interaction.
On this site:
| Robot | Height | Weight | Context |
|---|---|---|---|
| Unitree G1 | 130 cm | 35 kg | Small. Child-sized. Limited reach in human environments. |
| unitree-h1 | 150 cm | 47 kg | Mid-range. Most adult environments accessible. |
| unitree-h2 | 180 cm | ~70 kg | Human-standard. Fits most environments designed for adults. |
| Figure 02 | ~170 cm | ~70 kg | Human-standard. bmw-group factory-sized. |
| 1x-technologies-neo | 1X NEO]] | ~165 cm | ~70 kg |
| boston-dynamics-atlas | ~150 cm | ~89 kg | Dense and heavy for its size. Powerful but demanding on floors. |
The G1’s compact size is a feature for combat — smaller target, more agile — but it’s a limitation for real-world work where reaching standard heights matters.
Battery and Runtime
What it is: How long the robot operates between charges. Often expressed in minutes.
Why it matters: combat demonstrations are short (5–15 minutes). Real-world deployment needs hours. The Beijing Half Marathon (21 km) was won by robots with good thermal management AND battery management. It’s not just capacity — it’s efficiency across the mission.
The honest answer on most humanoids: Runtime specs are marketing. Real runtime depends on task, speed, payload, and environment. A robot rated for 90-minute runtime walking flat terrain might last 20 minutes if it’s carrying payload on uneven ground.
What to look for: Does the company give runtime under load, or unloaded? Unloaded runtime numbers are nearly useless for evaluating real-world deployment.
On this site, most humanoids don’t publish runtime specs under meaningful conditions. boston-dynamics-spot publishes 90 minutes — but that’s unloaded. figure-ai that halves under typical industrial payload.
Autonomy Level
This is the one that matters most for everything Robonomy covers.
Specs tell you what a robot can do mechanically. Autonomy level tells you who’s actually in control.
| Level | Definition | Example |
|---|---|---|
| teleoperated-vs-autonomous | Human controls every action in real-time | G1 at CES boxing, T-800 at URKL |
| Hybrid | Autonomous for routine tasks, human oversight for exceptions | Atlas in warehouses, figure-ai 02 at bmw-group |
| Autonomous | Robot makes decisions without real-time human input | darpa RACER vehicles, some UGVs in Ukraine |
Here’s the thing: a high-dof, high-torque humanoid that’s teleoperated is just a very expensive puppet. The mechanical specs are necessary but not sufficient. The autonomy stack is what determines whether a robot is doing work or being worked.
The gap between the most capable humanoid hardware (Atlas, Figure 02) and the highest autonomy level (DARPA RACER, Scout AI’s Fury) is enormous. We don’t yet have a humanoid that combines human-competitive mechanics with meaningful autonomous decision-making in real-world contexts.
That gap is the central story of autonomous robotics in 2026.
Putting It Together
When you see a new humanoid announced, here’s the checklist:
- dof: Is it above 25? If under 20, it’s locomotion-only — no meaningful manipulation.
- torque: Is it above 150 Nm at major joints? Below that, it can’t do physical work under load.
- Autonomy: Is it teleoperated, hybrid, or autonomous? High specs with teleoperated-vs-autonomous = impressive demo, not autonomous robot.
- Runtime under load: Does the company publish real-world runtime? If not, treat published specs as upper bound.
- Deployment: Is it in a factory, a lab, or a stage? Stage demos prove hardware survives demos. Factory deployments prove sustained real-world capability.
The robots on this site range from impressive lab demos to the actual operational frontier. The spec tables tell part of the story. The autonomy classification tells the rest.
Related
- robots — All specs, all platforms
- unitree-g1|Unitree G1]] — 120 Nm, compact, teleoperated-vs-autonomous combat demo
- unitree-h2 — 360 Nm, 31 dof, high-performance research
- Boston Dynamics Atlas — The dynamic capability benchmark
- figure-ai-02 — 16-DOF hands, real bmw-group factory deployment
- tech-autonomy — Full autonomy classification system
Sources
- unitree public documentation
- boston-dynamics public documentation
- Figure AI public documentation
Last updated: May 2026 | Spec data verified against company publications where available; [UNCERTAIN] noted where not | Sources: Unitree, Boston Dynamics, Figure AI public documentation