Humanoid Robots
Humanoid Robot Technology
Hand Dexterity Comparison
DOF = degrees of freedom, the number of independent joints in each hand. More DOF means finer manipulation (picking up a coin, threading a needle). Tesla Gen 3 and 1X NEO tie at 22, Fourier GR-2 at 12; a human hand has 27 as the benchmark.
Supply Chain Geographic Concentration
Where humanoid robot parts are actually made. China dominates rare earth processing (90%), battery manufacturing (86%), and cell assembly (77%); Japan controls 85% of precision reducers (the high-tolerance gearboxes inside each joint).
| Component | Dominant Country | Global Share | Key Companies | Risk |
|---|---|---|---|---|
| Rare Earth Processing | China | ~90% | Northern Rare Earth, China Minmetals | Critical |
| Rare Earth Mining | China | ~69% | Various state-owned | High |
| Battery Manufacturing | China | 86% of global capacity | CATL (38.1%), BYD (16.7%) | Extreme |
| Battery Cell Assembly | China | ~77% | CATL, BYD, CALB, EVE Energy | High |
| Cobalt Mining | DRC (Congo) | 76% | Glencore, CMOC (Chinese-owned) | Critical |
| Cobalt Refining | China | ~80% | Huayou Cobalt, GEM Co. | Critical |
| Lithium Mining | Australia | ~47% (86,000 MT) | Pilbara Minerals, Allkem | Moderate |
| Lithium Refining | China | ~65% | Ganfeng, Tianqi Lithium | High |
| Harmonic Drives | Japan | ~85% (Nabtesco + HD Systems) | Nabtesco, Harmonic Drive Systems | Very High |
| RV Reducers | Japan | ~60% | Nabtesco | High |
| Precision Motors (coreless) | Germany/Switzerland | ~70% combined | Maxon (CH), Faulhaber (DE), Portescap (US/CH) | Moderate |
| AI Compute Chips (design) | USA | ~90% (design) | NVIDIA (dominant), Qualcomm (emerging) | Moderate |
| AI Compute Chips (fabrication) | Taiwan | ~90% (advanced nodes) | TSMC | Critical |
| Force/Torque Sensors | USA/Denmark | ~60% combined | ATI Industrial (USA), OnRobot (Denmark) | Moderate |
| Depth/Vision Sensors | USA/France | ~50% | Intel RealSense (USA), Stereolabs (FR), Luxonis (USA) | Moderate |
| Carbon Fiber | Japan | ~50% | Toray, Teijin, Mitsubishi Chemical | Moderate |
| Critical Mineral Refining (avg) | China | ~70% (19 of 20 minerals) | Various state-owned + private | N/A |
Actuator BOM Detail
Cost split inside each actuator (joint motor) by type. Rotary actuators are driven by harmonic-drive gearboxes (36% of cost), linear actuators by roller screws (64%), and dexterous hands by tiny coreless motors (50%).
Rotary Actuator
Linear Actuator
Dexterous Hand
Onboard Compute Platforms
Published edge-compute products that humanoid stacks name. TOPS figures are vendor datasheet numbers under a stated precision (INT8 vs FP4) and are not interchangeable. Tesla Optimus onboard TOPS is not a public datasheet.
| Platform | Vendor | Published compute | Role | Disclosure |
|---|---|---|---|---|
| NVIDIA Jetson AGX Orin | NVIDIA | 275 TOPS (INT8, AGX Orin 64GB class) | Current edge inference on many research and commercial humanoids | Datasheet |
| NVIDIA Jetson Thor (AGX Thor) | NVIDIA | Up to ~2,070 FP4 TOPS (Thor class, vendor page) | Target runtime for GR00T-class VLA models | Datasheet / GTC 2025 |
| NVIDIA Isaac GR00T N1 | NVIDIA | Trained on DGX-class GPUs; inferred on Orin or Thor | Open humanoid VLA foundation model (Mar 2025) | NVIDIA newsroom |
| Tesla custom inference (Optimus) | Tesla | Not disclosed as a TOPS datasheet | Captive Optimus stack | Not a public module spec |
| Qualcomm robotics SoCs | Qualcomm | Varies by RB / Dragonwing SKU | Some China and research platforms | SKU-specific; not universal |
Humanoid Foundation and VLA Models
Announced vision-language-action models with a dated primary source. Closed models (Figure Helix, Tesla) have no public weights. An announcement is not a factory utilization rate.
| Model | Lab | Announced | Access | Embodiment | Note |
|---|---|---|---|---|---|
| Isaac GR00T N1 | NVIDIA | Mar 2025 | Open weights | Humanoid (Fourier GR-1, 1X demos) | Dual-system VLM + diffusion action head. GTC 2025. |
| Helix | Figure AI | Feb 2025 | Closed | Figure 02 | On-robot VLA. No public weights or independent task audit. |
| π0 (pi0) | Physical Intelligence | 2024 | Research / open materials | Multi-robot | Generalist VLA. Not a humanoid OEM stack. |
| RT-2 / Gemini Robotics line | Google DeepMind | 2023-2025 | Research | Lab manipulators and robots | Research papers, not a shippable humanoid SKU. |
| OpenVLA | Academic consortium | 2024 | Open | Multi-robot | Open VLA baseline used in comparisons. |
| Tesla Optimus net | Tesla | Ongoing | Closed | Optimus | End-to-end claims at AI events. No public eval suite. |
Actuator Architectures
Qualitative map of joint types. No invented per-robot dollar costs. Harmonic Drive Systems FY2025 reducer sales are company-wide, not a BOM line.
| Architecture | Where used | Supplier examples | What is (and is not) sourced |
|---|---|---|---|
| Strain-wave (harmonic) rotary | Arms, waist, many research humanoids | Harmonic Drive Systems, Leaderdrive, Green Harmonic | HDS FY2025 reducer sales 42.3B yen company-wide, not a BOM line. |
| RV / cycloidal rotary | Heavier joints, industrial arms, some legs | Nabtesco, Sumitomo | Incumbent industrial-robot reducer. Share figures are industry-level, not humanoid-specific. |
| Quasi-direct-drive / planetary | Unitree and several China bipeds | In-house plus China planetary vendors | Lower ratio, more backdrivable. List prices of finished robots are known; joint BOM is not. |
| Linear (roller / ball screw) | Some legs and Tesla-style linear joints | SKF, THK, NSK | Existing BOM file assigns roller screws a high share of linear-actuator cost. That is a category split, not a dollar invoice. |
| Hydraulic | Legacy Atlas, Sanctuary Phoenix hands | In-house hydraulic packs | High peak force. Electric Atlas moved off full hydraulics. |
| Tendon / cable hands | Tesla Optimus hands, 1X, some dexterous hands | In-house plus Maxon-class coreless motors | Do not invent a hand dollar cost. Count DOF from datasheets instead. |
Safety and Policy Standards
ISO 10218 (industrial robots), ISO 13482 (personal care), and ISO 25785-1 as a 2025 working draft for bipedal industrial robots. The draft is not a finished mandatory code. EU AI Act high-risk duties phase in through 2027.
| Instrument | Jurisdiction | Status | Applies to | Date |
|---|---|---|---|---|
| ISO 10218-1:2025 | International | Published standard | Industrial robots, including some factory humanoids used as industrial robots | 2025 |
| ISO 13482:2014 | International | Published standard | Personal care robots (mobile, physical-assistance, person-carrier) | 2014 |
| ISO 25785-1 WD | International | Working draft, not a finished standard | Bipedal industrial robots, dynamically stable platforms | May 2025 draft |
| EU AI Act | EU | In force, phased duties | High-risk AI systems; autonomous workplace robots likely in scope | High-risk duties Aug 2026; embedded product rules Aug 2027 |
| EU Machinery Regulation 2023/1230 | EU | Applies 2027 | Machinery including robots placed on the EU market | 2027 application |
| China MIIT humanoid guidelines | China | Industrial policy, not a safety test method | Domestic humanoid industry build-out (2025 readiness, 2027 lead language) | Nov 2023 |
| ANSI/RIA R15.06 | United States | US industrial robot safety (aligned with ISO 10218) | Industrial robot cells in US plants | Current US reference |
Published Perception and Tactile Suites
What operators disclose about cameras, force, and tactile sensing. Missing cells mean not published, not that the robot has no sensors.
| Model | Vision | Force / torque | Tactile | Note |
|---|---|---|---|---|
| Figure 02 | Onboard cameras (Helix) | Wrist sensing (company) | Yes (company) | Closed stack |
| Optimus Gen 3 | Tesla vision | Not a public module list | Company: more sensitive Gen 3 skin | No datasheet TOPS or sensor SKU list |
| NEO (1X) | Onboard | Not fully specified | Yes; IP68 hands | Consumer SKU, 2026 delivery |
| Phoenix Gen 8 | Onboard | Hydraulic hands | 5 mN claim (company) | Company tactile spec |
| A2 (AgiBot) | Visual fingertip sensors (company) | Not fully specified | Fingertip vision | 19 DOF hands |
| GR-2 (Fourier) | Onboard | Not fully specified | 6 array sensors per hand (company) | 12 DOF hands |
| Digit | Onboard | Force sensors for compliance | Gripper, not a 5-finger hand | Logistics end-effector |
| Walker S2 | Onboard | Not fully specified | Not fully specified | Battery-swap industrial SKU |
Published Robot Learning Datasets
Trajectory counts from papers and company dataset cards. Open X-Embodiment is multi-robot, not humanoid-only. AgiBot World 1 million is a company claim. GR00T mix has no single public integer and is omitted from the bars.
Charge vs Hot-Swap vs Self-Swap
How each platform says it stays on shift. Runtime hours and swap times are manufacturer figures. This is not a measured factory uptime study.
| Model | Runtime | Pack | Method | Shift implication |
|---|---|---|---|---|
| Walker S2 | Not a single-charge figure | Not disclosed | Autonomous battery swap (~3 min, company) | Designed for 24/7 line stories. Swap, not a 10-hour pack. |
| Optimus | 10-12 (company all-day target) | 2.3 (compilation) | Charge (no public hot-swap SKU) | All-day claim is a target/spec, not a third-party duty-cycle test. |
| Apollo | 4 | Not disclosed | Hot-swap <5 min (company) | Shift coverage depends on spare packs, not one pack. |
| Atlas Electric | 4 | Not disclosed | Self-swap (company) | Demo'd pack exchange. Not a disclosed MTBF. |
| Figure 02 | 5 | 2.25 | Charge / pack (company) | BMW pilot was not a 24/7 unmanned cell. |
| NEO | 4 | 0.84 | Charge; 6 min per hour claim (company) | Home SKU, not a factory shift robot. |
| Digit | 4 | Not disclosed | Shift-based charge | Designed around warehouse shifts. |
| Galbot G1 | 10 | Not disclosed | Charge (wheeled) | Wheeled base, not bipedal locomotion energy. |
Prefer Sterling in Google
Add sterlingcharts.com as a preferred source so our charts can appear with a preferred badge in Google Top Stories, AI Mode, and AI Overviews.