Boston Dynamics' Atlas ran on hydraulic actuators for nearly a decade — producing extraordinary athletic performance but requiring a tethered hydraulic power unit, complex sealing, and operating temperatures that made outdoor deployment nearly impossible. The 2023 pivot to a fully electric Atlas Gen 2 marked the industry's clearest signal: hydraulics are over for commercial humanoids.

The electric transition centers on three actuator archetypes now competing for design wins: quasi-direct-drive (QDD) motors with high-torque low-gear-ratio setups, series elastic actuators (SEA) that add compliance springs for safe human interaction, and harmonic drive gearboxes that achieve high reduction ratios in compact packages. Each involves fundamental tradeoffs in backdrivability, impact resistance, power density, and cost.

Strain wave gearing — commonly called "harmonic drives" after the original trademark — is the dominant gearbox technology in humanoid arm and wrist joints. Two companies control the majority of the global supply: Harmonic Drive AG (Germany/Japan) and Nabtesco Corporation (Japan). This creates a concentrated supply chain risk that every Western humanoid OEM is aware of but few have solved.

The SHD/SHF series from Harmonic Drive achieves zero-backlash, high-reduction-ratio gearing in a package small enough to fit inside a robot's forearm. A typical humanoid uses 12–20 harmonic drives — at $300–$1,200 each depending on size, this is one of the largest single BOM line items. Unitree's cost advantage partly comes from vertically integrating their own cycloidal gearboxes for leg joints and sourcing harmonic drives at scale through volume agreements unavailable to smaller competitors.

Figure 03 is Figure AI's third-generation humanoid, announced in October 2025. Figure has disclosed a limited set of hardware specifications publicly — the details below are drawn from official Figure announcements and verified third-party spec sheets. Figure vertically integrates their own actuators, batteries, sensors, structures, and electronics, per their own manufacturing announcement. Specific joint torques, gearbox suppliers, and internal control architecture have not been officially disclosed.

The robot weighs 61 kg — 9% lighter than Figure 02 — and uses integrated series-elastic actuators per the official specification. The hands include custom tactile fingertip sensors capable of detecting forces as small as 3 grams. The camera architecture delivers twice the frame rate and one-quarter the latency of Figure 02, with a 60% wider field of view per camera.

The humanoid robotics industry has split into two camps on perception — and the choice reflects deep assumptions about what AI can and cannot do reliably. LiDAR-equipped robots get precise 3D geometry data at 10–100m range regardless of lighting, but add weight, power draw, cost ($800–$3,000/sensor), and mechanical complexity. Vision-only systems bet that neural networks trained on massive datasets can infer depth, geometry, and object identity from stereo or monocular cameras alone — cheaper, lighter, but brittle in unusual lighting or novel environments.

Tesla's bet on vision-only for Optimus directly mirrors their FSD strategy: scale data, scale parameters, replace sensors with intelligence. Agility, Boston Dynamics, and most others carry at least one depth sensor — whether active stereo (Intel RealSense), structured light, or LiDAR — acknowledging that current vision models still fail in edge cases that geometric sensing handles trivially.

Camera-based manipulation works for rigid, predictable objects. It fails the moment an object is soft, deformable, reflective, or wet — because cameras cannot sense the contact forces, slip onset, and surface texture that human hands detect instinctively. The absence of tactile sensing is the single largest gap between current humanoid hands and human hands for manipulation tasks.

Current approaches to tactile sensing range from strain gauges at each fingertip (used by Figure, Sanctuary AI) to optical tactile sensors like the GelSight family (MIT-derived, used in research) that embed a camera inside a transparent gel finger pad to image surface deformation. Contactile's PapillArray — 3D force vector arrays across the finger surface — is entering commercial humanoid trials in 2026.

Every humanoid robot must run perception, state estimation, motion planning, and control — simultaneously, at millisecond latencies, on battery power. The compute stack is often the most power-hungry non-motor system on the robot. NVIDIA Jetson AGX Orin (275 TOPS, 15–60W) has become the de-facto standard for prototype and first-generation commercial humanoids, but is increasingly recognized as insufficient for next-generation vision-language-action (VLA) models.

The compute hierarchy in a modern humanoid typically has three layers: a high-level reasoning node (Jetson Orin or equivalent) running perception and AI inference at 20–100Hz, a mid-level motion controller (ARM Cortex-A or FPGA) running whole-body control at 500–1000Hz, and distributed joint controllers (STM32 MCUs) running individual joint FOC at 20–40kHz.

The average working humanoid is estimated to draw roughly 500W–2,000W depending on task intensity and locomotion speed — a commonly cited industry range, though no major OEM has published verified average power draw figures for their commercial robots. At that range, even a substantial battery pack provides only 1–5 hours of runtime, which is the most commonly cited deployment limitation.

Several architectural patterns are consistent across the industry: high-voltage battery packs reduce resistive losses in the motor drive chain, hot-swappable packs (confirmed in Digit and Apollo) allow near-continuous operation without plug-in downtime, and fast wireless charging is confirmed in Figure 03 at 2 kW. Specific voltage levels, battery chemistry, and charge times are not publicly disclosed by most OEMs.

The structural design of a humanoid robot is a continuous negotiation between weight, stiffness, cost, and manufacturability. Most OEMs do not publicly disclose their structural material choices in detail. What is confirmed: Boston Dynamics has publicly described 3D-printed titanium and aluminum components in the electric Atlas. Tesla confirms a 57 kg total weight using lightweight materials. Figure 03 uses soft textiles and foam for the outer surface. Unitree G1 achieves 35 kg with a compact frame.

The general engineering logic favoring carbon fiber composites for primary structure, aluminum for machined housings, and titanium for highest-load joints is well-established in aerospace and high-performance robotics — but specific material allocations per robot are proprietary and not independently verified for most platforms.