Humanoid Robotics: The Technology Landscape
The Technology Landscape
A read of the current commercial humanoid field along three axes: form factor, hardware, and intelligence architecture.
A humanoid robot is the product of three design layers. Each can be evaluated independently. Together they define what the platform can do, who can deploy it, and where it breaks.
The term humanoid robot covers more variation than is sometimes acknowledged. Two independent dimensions distinguish current commercial platforms.
How it moves (locomotion) and how it manipulates (end effectors).
Approximates human morphology. Mature on controlled surfaces. Open problems: novel terrain, fall recovery, reactive walking near humans.
Human-form upper body on a wheeled base. Cannot climb stairs. Trades leg locomotion for stability, efficiency, and uptime.
Two arms, no locomotion. Mounted in place or carried on an AMR. Focus entirely on dexterous manipulation.
Both legs and wheels. Climbs stairs when needed, runs on wheels on flat ground. Emerging hybrid form factor.
Five-digit, anthropomorphic. Grasps arbitrary objects, uses tools, performs fine motor tasks. Mechanically complex.
Simple grippers, suction, hinged effectors. Trades versatility for reliability, lower cost, predictable behavior. Tuned for known objects.
Task-specific effectors with optional dexterous attachments. A middle ground between purpose-built reliability and general-purpose dexterity.
Current humanoids share broad architectural patterns but differ in component choices that meaningfully affect capability and cost.
Each hardware layer is its own engineering problem with its own limit.
A full-size humanoid uses 28 to 44 actuators. Each is itself a small system. Their cost dominates the bill of materials, and their cost trajectory shapes humanoid pricing.
As humanoid AI moves from classical control with learned perception toward foundation models in the planning layer, the Protocol team’s read is that on-device compute is becoming the binding constraint on the next generation.
| Jetson AGX Orin | Jetson AGX Thor | |
|---|---|---|
| AI compute | up to 275 TOPS | 2,070 FP4 TFLOPS |
| Power draw | 15 to 60 W | 40 to 130 W |
| AI compute vs Orin | baseline | 7.5x higher |
| Energy efficiency vs Orin | baseline | 3.5x better |
| Sensor fusion latency | not specified | under 10 ms |
| Designed for | classical control with learned perception | many concurrent foundation models |
Current platforms cluster in a narrow capacity range. Hot-swap and inductive charging are the workarounds. Solid-state is the next inflection.
| Platform | Capacity | Runtime | Charging |
|---|---|---|---|
| Unitree H1 | 0.864 kWh | under 4 h static | standard |
| Tesla Optimus Gen 2 | 2.3 kWh | ~2 h dynamic | standard |
| Apptronik Apollo | per pack | 4 h per pack | hot-swappable |
Solid-state lithium-metal could theoretically double current energy density. A humanoid carrying half the weight, or running a full eight-hour shift on the same pack.
The software stack that decides what the robot does. Sits on top of the compute hardware. Three layers, three current patterns.