Anatomy of a Modern Humanoid

A 2026 humanoid robot is not one machine. It is six or seven engineering disciplines bolted together — actuators, joint control, perception, on-board compute, batteries, thermals, and (sometimes as an afterthought) software. Most of the marketing photos show the chassis. This article is about what is inside it, and what the spec sheet does and does not tell you.

The reference set is five platforms you can actually buy today: Unitree G1, Unitree H1 / H1-2, Booster T1, EngineAI PM01, and EngineAI SE01. They span roughly $12k to $129k and three size classes. The differences matter.

What "degrees of freedom" actually means

Every spec sheet leads with a DOF number. It is the count of independent joint axes the robot can drive. More is not automatically better — it is more capability and more mass, more wiring, more controllers, more things to wear out.

A rough decomposition for a modern humanoid:

A 19-DOF platform (Unitree H1 base) is locomotion-first: legs walk well, arms are basically counterweights with grasp affordance. A 41-DOF platform (Booster T1 with dexterous hands) is a manipulation research platform — the legs are still there but a lot of the budget went to the fingers.

Cross-platform comparison, base configurations:

*PM01's waist rotates 320°, which is unusual and useful — most humanoids cap below 180°.

The honest reading: G1 and H1-2 are about evenly matched if you add EDU dex hands; T1 dex is the highest hand-DOF count in this set; SE01 is the biggest balanced platform; PM01 is the smallest serious one.

Actuators — what is actually moving the joints

This is where the marketing gets loose. Three actuator families dominate, and which one a vendor uses tells you more about the robot than the DOF count.

Quasi-direct-drive (QDD). A high-torque-density outer-rotor motor with a small reduction (typically 5:1 to 10:1) — either a low-ratio planetary or, in Unitree's QDD module, a harmonic reducer integrated into the case. QDD is backdrivable (you can push the joint by hand and feel it), tolerates impacts, recovers cheaply from a stumble, and is fast enough for dynamic gait. It is the right answer for legs.

The downside: limited static torque per unit mass at the high end. A QDD knee that hits 360 N·m peak (Unitree H1) is impressive; getting to 500+ N·m for heavy-duty industrial lift is hard without going to a different topology.

Harmonic / strain-wave drive. A high reduction ratio (100:1 and up) in a small package, with very low backlash. Used in industrial cobots (Universal Robots, Franka), and in arms or wrists of humanoids where positioning precision beats backdrivability. Harmonic drives are not friendly to impacts — they wear, they shock-load, they cost money. You see them in research-grade dexterous hands and high-precision wrist actuators.

Series-elastic actuator (SEA). A motor + gearbox with a deliberate compliant element (a spring or torsional element) between the output and the load. The spring acts as a torque sensor and as a shock absorber. Boston Dynamics' older Atlas hydraulic and some of Agility's Digit joints fall here in spirit. SEAs are good for human-contact tasks where you want soft response, but they are mechanically more complex and the spring is another wear part.

What the 2026 lineup actually uses:

Everyone in this price tier is QDD-first. The dexterous hand add-ons use small harmonic reducers because finger joints need precision more than backdrivability. If a vendor pitches "SEA" or "compliant actuation" on a sub-$50k humanoid, ask which joints — it is usually one or two, not the whole platform.

On-board compute — what the robot can think with

A modern humanoid carries two computers (logically, sometimes physically): a real-time controller that talks to the motors, and an application processor that runs perception, planning, and any on-device ML.

The real-time controller is usually an ARM Cortex-M or low-power Linux SoC running an RT kernel, closing the joint control loop at 500 Hz to 1 kHz. You will not see this in the spec sheet most of the time — it is a board behind the chest plate.

The application processor is what gets advertised. The two families in 2026:

Per-platform compute:

What an on-board Jetson AGX Orin can actually do:

• Run YOLOv8 or YOLOv11-s at 30+ FPS on the head camera.
• Host a 7B INT4 LLM (Qwen2.5-7B, Llama-3.1-8B) at 15–25 tok/s.
• Run a small vision-language model — Qwen2.5-VL 3B or 7B quantized — at usable but not fast rates.
• Local STT (Whisper small / distil) and a wake-word.

What it cannot do, and you should not pretend otherwise:

• A 70B+ LLM. Not even close. VRAM budget alone disqualifies it.
• A serious VLM (Qwen2.5-VL 72B, NVIDIA Cosmos). Same problem.
• Long-horizon planning with large context windows.
• Simulation or any training workload above LoRA-on-a-toy-model.

This is the split that motivates everything in the I01 edge-AI architecture article: the robot runs the closed-loop and the safety reflexes, the off-board GPU server runs the heavy thinking. The Jetson is enough to keep the robot useful when the network is gone, and not enough for the interesting models.

Sensors

A 2026 humanoid ships with more sensors than most engineers want to deal with. The standard complement:

LiDAR is the spec to read carefully. The G1 and SE01 ship with 360° solid-state LiDAR by default; the H1 base, T1, and PM01 lean on stereo depth and offer LiDAR as an upgrade. For an indoor mapping use case, LiDAR is almost mandatory. For a manipulation-only research platform, you can skip it.

Battery, runtime, hot-swap

The dirtiest number on the spec sheet. "2 hours runtime" assumes light walking. Sustained dynamic motion — running, lifting, carrying — pulls it down to 45–75 minutes. Standing still and thinking pulls it up.

Quick-swap is not the same as hot-swap. Most of these designs require the robot to be either powered down, in a safe pose, or on a docking stand for the change. True hot-swap (keep running while changing the pack) needs a parallel capacitor / second-pack arrangement and is uncommon in this price tier.

Plan for two batteries per robot at minimum, three if you are running sustained shifts. Plan for charger throughput too — a single charger that takes 90 minutes to refill a pack is the bottleneck once you have more than two robots.

Cooling

Almost every humanoid in this set cools the actuators and the chest electronics with a small fan or two pulling air through a chest cavity. It is not glamorous and it is the first thing to throttle under sustained load.

Realistic thermal behaviour:

• Light walking, 23 °C ambient: no throttle, indefinite runtime up to battery limit.
• Sustained running, lifting, manipulation: chest electronics climb 15–25 °C above ambient within 10–20 minutes. The Jetson will throttle from MAX_N power mode down to 30 W mode silently.
• Warm environment (28 °C+): runtime drops, perception latency increases as the Jetson clocks down.

The SE01 is the only one in this set that publicly cites "forced-air cooling at critical joints" — actuator cooling, not just chassis cooling. For sustained dynamic tasks (running, lifting), joint cooling is the difference between an hour of performance and ten minutes.

For an indoor lab at 22–24 °C, you will not hit thermal limits except during pushed demonstrations. For factory-floor or outdoor operation in summer, derate everything.

Communication

Wi-Fi 6E is the practical floor — the 6 GHz band is the only one with consistent low-latency behaviour around other consumer devices. Wi-Fi 6 (no 6E) works in a quiet environment and falls over the moment you have phones, laptops, and APs sharing 5 GHz.

A wired tether for development is worth it. The G1 EDU and Booster T1 both support tethered operation with a coiled umbilical for power-and-data; the H1 supports it less gracefully.

Price tiers, Q1 2026

These are list prices in USD, ex VAT, ex shipping, ex import duties. EUR equivalents at recent rates. Where vendors do not publish, the figure is the most-credible distributor list. Treat these as ballpark, not quotes.

The honest read. Below $20k buys you a developer toy with a Jetson — useful for software research, RL training, and SDK work, but not a tool you put on a job. The $30–60k tier is where research labs and serious integrators live. Above $90k is for performance work (H1's 3.3 m/s record) or arm-payload work (H1-2's 10 kg per arm). There is no "industrial humanoid" in this list — for that you are looking at Figure, Agility Digit, or Apptronik Apollo, which start north of $200k and are not cash-and-carry.

What differs across platforms

A short, opinionated map:

• Unitree G1 — best dexterity per dollar in the EDU configurations; smallest of the full-size set; tightest SDK story; the platform most third-party tutorials target. Default choice for academic robotics in 2026.
• Unitree H1 / H1-2 — full human-size, fastest locomotion in the lineup, highest payload (H1-2 = 10 kg per arm). Buy this if your task is "move fast" or "lift things at human height". Expensive.
• Booster T1 — championship-validated (RoboCup AdultSize 2025), strong on-board compute (AGX Orin + i7), good RL story. The pragmatic research platform if you do not want Unitree lock-in.
• EngineAI PM01 — the cheap one. 24 DOF, 320° waist, small (138 cm). For software work and student labs where breaking the robot is not a $50k mistake.
• EngineAI SE01 — full-size (170 cm, 55 kg), 32 DOF, the only one publicly citing forced-air joint cooling. The "we built this for actual sustained dynamic motion" platform.

If you want one sentence per platform: G1 = dexterity, H1 = power, T1 = compute, PM01 = price, SE01 = endurance.

Maintenance reality

A humanoid is a wear machine. The high-cycle parts in rough order:

1. Hand and wrist actuators. First to fail, especially the harmonic reducers in dex hands. Plan for one rebuild per active hand per year of moderate research use.
2. Cables and connectors at the joints. Vibration fatigues wires. Plan for one cable failure per quarter on a heavily used unit.
3. Knee and hip QDD modules. Robust but not eternal. The bearings, the harmonic gears, the brake mechanism — all wear. Expect a 2,000–4,000-hour service interval for serious dynamic use.
4. Battery packs. 300–500 full cycles before noticeable capacity loss. With 2 hr cycles, that is a year of moderate use.
5. Chest fans. Cheap, easy, but they fail silently and the result is throttle. Replace on a schedule, not on failure.
6. Foot pads / rubber. Especially on bipeds that walk on hard floors. Cheap, fast to replace.

Vendor parts availability is the differentiator. Unitree has the most mature spares pipeline; Booster is improving; EngineAI is still building out the channel in Europe. For a deployment in Prague or Brno, expect 2–4 weeks for non-trivial parts from any vendor.

Where Kentino fits

Most readers of this article are buyers, integrators, or research labs. Kentino is a Czech channel partner / co-sell on the compute side — we build the on-prem inference servers that the humanoid talks to (K-AI line, multi-GPU, EPYC or Xeon, RTX 5090 / Pro 6000 Blackwell / L40 / L4). We do not sell humanoids. The reason we write this article is so that the buyer who has decided on a G1 or a T1 can scope the GPU server correctly the first time.

The pairing question — which inference server matches which robot — is covered in I01 and will be deepened in I05 (the reference build).

---

This is part of the Kentino Wiki, a reference series on AI compute, robotics, and the systems that connect them. Comments and corrections welcome at info@kentino.com.