Humanoid Robotics: The Technology Landscape

The term “humanoid robot” covers more variation than is sometimes acknowledged. Two independent design dimensions distinguish current commercial platforms: how the robot moves (locomotion) and how the robot manipulates objects (end effectors).

Locomotion

The lower body determines how the robot moves through space. Four approaches are present in current commercial platforms.

1. Full bipedal humanoids walk on two legs and have two arms, approximating human morphology. Bipedal walking on controlled indoor surfaces is a mature control problem with well-established solutions including Zero Moment Point control. Robust walking across novel terrain, fall recovery from arbitrary configurations, locomotion while performing significant manipulation, and reactive walking in dynamically changing human-occupied environments remain active research problems.
2. Wheeled humanoids combine a human-form upper body with a wheeled mobile base, sacrificing leg locomotion for dramatically improved stability, energy efficiency, and operational uptime. Wheeled humanoids cannot climb stairs but can operate continuously without the energy cost of bipedal balance. They are well-suited to flat industrial environments.
3. Bimanual manipulation platforms without locomotion focus on dexterous two-armed manipulation while remaining stationary or mounted on autonomous mobile robot (AMR) platforms. The focus is entirely on manipulation rather than mobility.
4. Wheeled bipedal humanoids are an innovative form factor combining the strengths of both full bipedal humanoids and wheeled humanoids. This form factor can navigate through complex landscapes such as stairs while preserving the speed and predictability of wheels for flat ground.

End Effectors

The upper body’s terminal manipulators determine what objects the humanoid can interact with.

Multi-finger dexterous hands approximate human manipulation capability. Five-digit designs are capable of grasping arbitrary objects, using tools, and performing fine motor tasks. Dexterous hands are mechanically complex, with 11 to 22 degrees of freedom per hand, and require sophisticated tactile sensing and control to operate reliably. Reliable dexterous manipulation across diverse objects, force-controlled grasping of fragile items, and tactile feedback at human-level resolution remain active research problems.

Task-specific end effectors are designed for narrow operational contexts rather than general-purpose manipulation. Simple grippers, suction cups, and hinged effectors optimized for specific tasks like tote movement sacrifice manipulation versatility for reliability, lower cost, and predictable behavior. Task-specific effectors are well-suited to repetitive logistics work where the object profile is known and stable.

Hybrid configurations combine task-specific end effectors with optional dexterous-hand attachments, giving operators a choice between reliability for the primary deployment context and dexterity for occasional tasks requiring it. This form factor is emerging as a practical middle ground between purpose-built simplicity and full general-purpose dexterity.

The current generation of humanoid robots shares broad architectural patterns while differing in component choices that meaningfully affect capability and cost.

Actuators

Actuators are the muscles that execute every action. Modern humanoid actuation has converged on electric drives, with hydraulic actuation largely abandoned (Boston Dynamics’ Atlas transitioned from hydraulic to electric in its 2024 redesign). Joint count ranges from approximately 20 degrees of freedom in simpler designs to over 50 in the most dexterous platforms.

Actuators are the single largest cost driver. A full-size humanoid uses 28 to 44 actuators, each combining a motor, gearbox, encoder, and controller.

• Motor: Converts electrical energy into rotational force, typically a brushless DC motor spinning fast at low torque.
• Gearbox: Trades the motor’s high speed for the high torque the joint actually needs, using harmonic, planetary, or cycloidal designs.
• Encoder: Measures the joint’s actual position and feeds it back to the controller so motion can be corrected in real time.
• Controller: A small dedicated chip at the actuator that runs the high-frequency control loop, closing the gap between commanded and actual position.

According to Robozaps, high-torque actuators for major joints cost $500 to $2,000 each at current production volumes, producing total actuator subsystem costs of $16,000 to $40,000 per robot and accounting for 40 to 50 percent of total manufacturing cost. Mass production at Tesla-scale targets is projected to drop per-actuator cost to $100 to $300, making actuators the central driver of humanoid pricing trends.

Sensors

Sensors are the perception layer that lets the humanoid see, feel, and balance. The sensor stack determines what the robot can perceive, and perception failures cascade directly into action failures, making this layer disproportionately important for safe deployment.

Current humanoids combine multiple sensing modalities, each suited to a different aspect of the world.

• RGBD cameras: Provide color 2D images and depth representations such as depth map and point cloud supporting grasping, obstacle avoidance, and 3D scene reconstruction.
• LiDAR: Uses laser beams to measure distances and map 3D spaces. LiDAR excels at long-range and is more robust against sunlight than camera-only vision, but adds cost.
• Inertial measurement units (IMUs): Track motion: acceleration and angular velocity to support balance and locomotion control. An IMU sits on the robot’s body (typically chest or pelvis) and reports two things:
  • Linear acceleration along three axes (how fast it is speeding up or slowing down)
  • Angular velocity around three axes (how fast it is rotating)
• Force-torque sensors: measures two kinds of contact forces between the robot’s wrists, ankles, and the outside world:
  • Force along three axes (how hard something is pushing or pulling)
  • Torque around three axes (how much rotational pressure is being applied)
• Tactile sensors: Detect pressure and contact at the fingertips, supporting dexterous grasping.

Compute

Compute is the brain that runs the perception, planning, and control models in real time. The compute layer determines what AI models the robot can run on-device, how fast it can respond to sensor input, and how much power and heat are spent on AI relative to the rest of the system. As humanoid AI shifts from classical control to foundation-model-driven architectures, the Protocol team predicts that compute is becoming the primary constraint on what the next generation of robots can do.

Current humanoids rely primarily on NVIDIA Jetson modules, with two generations in active deployment.

• NVIDIA Jetson AGX Orin: The current workhorse module, delivering up to 275 TOPS of AI performance at 15 to 60 W of power, used by Agility Robotics, Boston Dynamics, Figure AI, and most other commercial humanoid platforms.
• NVIDIA Jetson AGX Thor: Launched in August 2025, designed specifically for humanoid robotics. Delivers 2,070 FP4 TFLOPS of AI compute at 40 to 130 W, with 7.5x higher AI compute and 3.5x better energy efficiency than AGX Orin.

The transition from Orin to Thor reflects the broader shift in humanoid AI architectures. Orin is sufficient for classical control with learned perception and policy components. Thor is designed for running multiple foundation models concurrently, including vision-language-action models and NVIDIA’s Project GR00T humanoid foundation model. Thor’s under 10-millisecond sensor fusion latency and ability to process synchronized data from LiDAR, cameras, and microphones at high speed addresses the real-time control requirements that limited earlier compute generations.

Battery and Power

Battery and power is the bottleneck that determines how long the humanoid can actually work before stopping. Energy is shared across locomotion, sensing, compute, and actuation, and the trade-off shapes everything from the robot’s weight to its deployment economics. The Protocol team deduces that battery technology is the largest constraint holding back continuous humanoid operation today.

Current humanoids use lithium-ion battery packs, with most platforms clustered in a narrow capacity range.

• Typical capacity: Under 2 kWh for most current platforms, roughly equivalent to a high-end electric bicycle. For comparison, a Tesla Model 3 Long Range carries 82 kWh.
• Typical runtime: 2 to 4 hours per charge under real workload conditions, with most platforms delivering closer to 2 hours during heavy operation. Unitree H1 lists 0.864 kWh and under 4 hours static. Tesla Optimus Gen 2 at 2.3 kWh runs approximately 2 hours dynamic. Apptronik Apollo offers 4 hours per pack.
• Chemistry: Most platforms use high-nickel ternary lithium (NMC/NCA) chemistries for high energy density and high discharge rate, abandoning cheaper Lithium Iron Phosphate (LFP) batteries due to weight and space constraints. Off-the-shelf EV batteries cannot be directly repurposed for humanoids due to different discharge profiles and thermal demands.
• Charging strategies: Two approaches dominate. Hot-swappable battery packs replaced in under five minutes (Apptronik Apollo, UBTECH Walker S2 with autonomous 3-minute swap) enable near-continuous 24/7 operation through battery rotation. Autonomous charging, including Figure 03’s inductive charging coils integrated into its feet delivering 2 kW wireless power, removes the need for human battery handling.

The industry is awaiting solid-state battery commercialization as the next inflection point. TrendForce projects demand for humanoid solid-state batteries to surpass 74 GWh by 2035, over 1,000 times current demand. Solid-state lithium-metal batteries could theoretically deliver energy densities above 500 Wh/kg, roughly double current standards, allowing a humanoid to operate a full eight-hour shift on a battery weighing half as much as today’s equivalent.

The intelligence architecture is the software stack that governs how the humanoid perceives, reasons, and acts. It sits on top of the compute hardware (Jetson AGX Orin or Thor) and ranges from classical control with hand-engineered equations to foundation-model-driven architectures with billions of learned parameters. The current humanoid field shows real differentiation across vendors at this layer, and the architectural choice has direct consequences for capability, predictability, and safety.

The Three Layers

The intelligence stack typically decomposes into three layers, each of which can be implemented with classical or learned approaches.

• Perception layer: Processes raw sensor data into structured understanding of the world. Cameras, LiDAR, IMUs, and force-torque sensors all feed in; the perception layer outputs object detections, 3D scene reconstructions, semantic labels, and pose estimates.
• Planning and reasoning layer: Decides what to do given a goal and the perceived state. Translates “pick up the cup on the table” into a sequence of intermediate actions. This is the layer where foundation models are most actively being deployed. Vision-language-action (VLA) models, large language models adapted for robotics, and humanoid foundation models like NVIDIA’s GR00T all sit here.
• Control layer: Converts planned actions into specific motor commands at high frequency. Maintains balance, executes locomotion, drives manipulation motions. Predominantly classical control (model predictive control, ZMP-based methods, inverse kinematics) in current humanoids because real-time stability and provable safety matter most at this layer. Increasingly augmented with learned components.

Three Architectural Patterns

Classical-control-dominated stack. Earlier humanoid platforms and conservative current platforms use hand-engineered planning and control with learned components limited to perception. The robot identifies objects and scenes using neural networks but decides what to do and executes motions using classical engineering. Predictable and safety-arguable but limited in generalization.

Foundation-model-driven stack. The frontier architecture. A large foundation model takes vision and language input and outputs action plans, with classical control retained primarily for low-level motor execution. Figure 03 with Helix VLA, Apptronik Apollo with NVIDIA Project GR00T integration, and various humanoid foundation models in active development represent this direction. These platforms are betting that general-purpose intelligence transfers better across tasks than purpose-built models.

Teleoperated stack. A human operator’s perception, reasoning, and motor actions are mapped onto the robot through a control interface usually via a VR headset. The robot is the operator’s body, not an autonomous agent. Many commercial humanoid pilots include teleoperated phases.