Comprehensive field database covering companies, market economics, technology, dexterous hands, and investment landscape
| Company ↕ | Country ↕ | Robot Model(s) | Segment | Website |
|---|
Scored 1-10 across technology advantage, commercial momentum, AI capabilities, hardware IP, manufacturing, funding, talent, patents, gov support, supply chain, and cost competitiveness.
| Company | Country | Safety | Tech | Commercial | AI | HW IP | Mfg | Funding | Talent | Patents | Gov | Supply | Cost | Total |
|---|
≈ US $1,000,000
≈ US $150,000
≈ US $5,600
≈ US $5,900
≈ US $10,000
| Metric | Human Picker (Night Shift) | Humanoid (RaaS) |
|---|---|---|
| Annual Cost | US $49,000 | US $21,600 |
| Available Hours/Year | 1,900 | 7,000 |
| Cost per Productive Hour | US $26 | US $3 |
| Break-even | < 9 months at premium lease rates | |
Raised: US $1.5B | Post-money: ~$39.5B
Nvidia, SoftBank, Tiger Global
Raised: US $1B | Post-money: ~$5B
OpenAI, NVIDIA, EQT Ventures, Tiger Global
Raised: US $150M | Private
DCVC, Amazon, Sony Innovation Fund
Latest hardware, social-media bragging rights. Willing to pay $50K+.
Tangible labour savings (cleaning, errands). Budget ≤$50K.
Car-level pricing (~$20K), reliability over novelty.
Need convincing use-cases and proven track record.
| Task | Tech Readiness 2025 | Blocking Gaps | Likely Timeline |
|---|---|---|---|
| Floor vacuum/mop | Medium | Biped balance; battery life | Pilots 2026-27 |
| Laundry handling | Low | Reliable cloth grasp & fold | 2028-30 |
| Basic cooking | Low | Heat-safe manipulators; food-safe | 2029-32 |
| Garden weeding/mowing | Low | Weather sealing; outdoor localisation | 2027-29 |
| Child/elder companionship | High (AI speech) | Safety certification; trust frameworks | Incremental via OTA |
| Industry | Near-term Use Cases | Pain-point Fit | Barrier / Enabler |
|---|---|---|---|
| E-commerce & 3PL | Trailer unloading, aisle cleaning | High | ROI obvious; unions watch safety |
| Manufacturing | Line-side material delivery | High | Needs CE/UL safety sign-off |
| Facilities Mgmt | Night-shift cleaning, security | Medium | Lift-integration & BMS APIs |
| Healthcare | Pharmacy runs, waste removal | Medium | Strict hygiene, patient trust |
| Hospitality & Retail | Room service, shelf restock | Medium | Image-sensitive; robot must be quiet |
| Oil & Gas | Inspection, valve turning | High | EX certification; harsh environment |
| Construction | Material ferry, site scanning | Low | Outdoor durability, rough terrain |
Tote handling & trailer unloading. Multi-year paid contract, KPI ≥ 800 cases/hr. Live Q4-2024.
Assembly-line material shuttle. 500-unit pre-order; 16h/day uptime target.
In-store marketing mascot. US $16K unit cost; remote-piloted to satisfy insurance.
End-of-line inspection. Proof-of-concept; part of funding rationale.
| Approach | Examples | Pros | Cons |
|---|---|---|---|
| Actuators in Palm (Direct Drive) | Figure AI, Unitree (Dex5-1) | Simple mechanics, high precision, local sensor ease | Heavier hand, heat issues, bulkier design |
| Wire-Driven (Tendon/Cable) | Tesla Optimus, 1X (NEO) | Lightweight, backdriveable, biomimetic efficiency | Backlash/friction, maintenance, control complexity |
| Pneumatic (McKibben muscles) | Clone Robotics | Better embodiment of nature, compliant | Compliance issues, hard to control precisely |
High design freedom, brakes/clutches/springs as fluid circuit. But: leaks unacceptable indoors, hard to miniaturize. "Persian rug problem."
Quasi-direct drives: high-torque PM motor + low-ratio transmission + sensing. Default topology for humanoids. Backdriveable, efficient.
Fine at slow speeds. At faster walking/running: heat up, fatigue, fail. Trend is direct actuation with lowest gear ratio possible.
Cycloidals: backdriveable but hard to make light, tough to miniaturize. Harmonics: dominant for hips/knees/shoulders, compact, low backlash.
Endurance under cyclic loads is non-negotiable. Wire-driven: cable fatigue. Palm actuators: overheating.
High DoF and fluid motion. Wire-driven approaches often lead due to lighter construction.
Balanced grip force (8-40 kg) without excess power draw.
Allows compliance for safety. Inherent in wire systems. Prevents damage from external forces.
SensoBright: 400+ data points per sensor, light-based, -40°C to +90°C, waterproof.
| Bottleneck | Status 2025 | Risk if Unresolved | Mitigation |
|---|---|---|---|
| Manipulation reliability | ≈70% simple grasps | Task failures kill ROI | Multi-modal AI + force sensors |
| Battery energy density | 3h shift limit | Frequent charges → downtime | Fast-swap packs, solid-state cells |
| Fail-safe perception | Good lighting only | Accidents, insurance push-back | LiDAR + RGB-D fusion + ISO cert |
| Hand ↕ | Company | Country | DoF | Weight | Force | Actuation | Tactile | Self-Lock | Price | Status |
|---|
Sharpa Robotics (Singapore/Shanghai) — CES 2026 Innovation Award
22 DoF via 22 proprietary multi-mode power modules. 1:1 human scale using golden ratio (0.618) phalangeal proportions — critical for ML as it allows direct mapping of human mocap data without lossy retargeting. Weight: 1,300g including all sensing electronics.
DTA Tactile: 1,000+ pixels/fingertip, sub-mm spatial resolution, 6-axis force (Fx,Fy,Fz,Mx,My,Mz), 0.005N–30N range, 6,000 pressure levels, 180 FPS at 20ms latency. Survives 100K+ press cycles at 16N.
Comms: 1000Base-T Ethernet, 500Hz control. Sim: ROS 2, Isaac Gym/Lab, MuJoCo. Gesture speed: >4 Hz full open/close.
Wuji Robotics (Shanghai) — Masterclass in electromechanical actuation
20 DoF, all independently actuated. Abandons tendons entirely for self-locking rotary direct-drive micro-joint architecture. Motors, encoders, and gear reducers housed directly within fingers and palm. No mechanical coupling between joints — enables complex motions like piano playing. Skeleton: <600g.
IP Joints: Micro-worm gears (self-locking — grip holds indefinitely without power). MCP Joints: Four-bar linkage with palm motors delivering 3x finger torque.
Control: 1kHz FOC on all 20 axes, dual encoders (input + output) per joint. Power: 12-20V DC, 1A static / 35A peak. Comms: USB-C / RS485 / EtherCAT. SDK: wujihandpy (MIT license), ROS 2, URDF. Durability: 300K+ QC cycles, survives 80cm drop.
Shadow Robot Company (UK) — Research gold standard since 2004
24 joints, 20 DoF. Tendon-driven with pneumatic muscle or electric motor options. 1:1 human scale. ~4,000g. >10N/tip force. The most widely used research dexterous hand in academia with hundreds of published papers.
Tactile: Optional BioTac (SynTouch) multimodal sensors. Comms: EtherCAT. Sim: ROS, Gazebo, MuJoCo. Extensive 3rd-party sim-to-real RL research base.
Psyonic (USA) — FDA-cleared prosthetic, adaptable for robotics
6 DoF (6 independent actuators). Extremely fast: closes in 200ms. Originally designed as myoelectric prosthesis — world's fastest multi-articulating prosthetic. Built-in pressure sensors on each fingertip. Rugged construction for daily prosthetic use. Touch feedback via sensory substitution.
Wonik Robotics (South Korea) — Popular RL research platform
16 DoF (4 fingers x 4 joints). Torque-controlled via servo motors. 1.1 kg. Widely used in reinforcement learning and teleoperation research. Well-supported sim models in MuJoCo and Isaac Gym. Position, velocity, and torque control modes.
Inspire Robots (China) — Cost-effective research/industrial hand
6-12 DoF depending on model (Dexterous / RH56 series). Servo-actuated with position feedback. Models from basic grippers to multi-finger dexterous variants. Used on multiple Chinese humanoid platforms. ROS support, affordable entry point for researchers.
Robotiq (Canada) — Industry-standard adaptive grippers
2-3 finger designs (2F-85, 2F-140, Hand-E). Not full dexterous hands but widely deployed adaptive grippers. Plug-and-play with Universal Robots. Force-controlled, position feedback. 5kg+ payload. Production-hardened with >10M cycle life.
Unitree Robotics (China) — First-party dexterous hand for H1/H2/G1
20 DoF (16 active + 4 passive coupled joints). Uses proprietary hollow-cup (coreless) motors with self-developed micro force-controlled composite transmission joints. 12 composite transmission joints + 4 gear transmission joints. All joints are smoothly backdrivable — a deliberate design choice for safe RL training. Micro-gap joint design places rotation axes close to finger surfaces, preventing object snagging.
Force: 10N/fingertip, 3.5kg grip (palm down) / 4.5kg (sideways). Weight: ~1,100g. Dims: 217x128x72mm. Control: 1kHz via USB 2.0. Voltage: 24-60V DC. Temp: -20C to 60C. Positioning: ±1mm fingertip repeat accuracy.
Dex5-1P (tactile variant, +$4K): 94 force sensors distributed across palm (10), fingertips (6 each), phalanges (6 each), finger roots (6 each). Detection: 10g–2.5kg range, 20kg max load. Reports both pressure and temperature.
Sim: MuJoCo (unitree_mujoco), Isaac Lab (unitree_sim_isaaclab), ROS/ROS 2 (URDF), Gazebo. SDK: unitree_sdk2 (C++), unitree_sdk2_python. Teleop: Apple Vision Pro, PICO 4, Meta Quest 3 via xr_teleoperate. Modular: Individual fingers replaceable.
Compatible: Unitree H1, H1-2, H2, G1 EDU (U1)+, and third-party via USB.
BrainCo (China/USA, Harvard Innovation Lab origin) — Prosthetic heritage, class-leading grip-to-weight
11 DoF (6 active + 5 passive). Only 383g — lightest production dexterous hand available. Despite minimal weight, delivers 50N grip force, 15N pinch force, and lifts 20kg single-hand — a grip-to-weight ratio of 52.6. Aerospace-grade aluminum frame with food-grade polymer contact surfaces. <50dB operation. 9-64V input range.
Three variants: Basic ($5K, RS-485 + CAN-FD) | Pro ($6.7K, + EtherCAT + adaptive impedance) | Touch ($7.5-8.4K, + 3D multimodal tactile: pressure, force direction, hardness, texture, proximity detection).
Precision: 0.1° repeatability, 0.1mm positioning. Speed: ≤0.65s full open-close. Control: ~1kHz. Comms: RS-485, CAN-FD, EtherCAT (Pro/Touch), Protobuf, DDS.
SDK: Python (3.8-3.12), C/C++, Linux/macOS/Windows. ROS 1 & ROS 2 (URDF via revo2_description, Apache 2.0). Unitree DDS bridge via brainco_hand_service.
BCI Heritage: BrainCo's prosthetic hand (TIME 100 Best Inventions 2019) uses 8-channel surface EMG for individual finger control. EMG armband teleoperation documented for Revo 2. Nature Communications publication on EEG-based BCI for real-time robotic hand control.
Compatible: Unitree H2, G1-D (Standard & Ultimate editions), R1 Edu Pro D, 6-DoF robotic arms, general platforms via flange interface.
Custom direct-drive in-palm actuated hands with integrated sensing. Demonstrated folding shirts, sorting items, and delicate manipulation. Harmonic drive actuators for rotary joints. Undisclosed DoF but high dexterity shown in demos.
22 DoF (11 per hand). Wire/tendon-driven with actuators in forearm. Demonstrated threading needles, handling eggs, playing jenga. High sim-to-real gap due to tendon nonlinearities — active area of development with Grok integration.
Tendon-driven with actuators in forearm/torso. Biomimetic muscle routing for natural compliance. Designed for household tasks — soft, compliant interaction. World model approach (1XWM) for manipulation learning.
20 DoF (16+4) dexterous hand for H1/H2/G1. Hollow-cup motors with micro gear transmission, fully backdrivable. $25K base / $29K with 94 tactile sensors. 1kHz control, MuJoCo + Isaac Lab + ROS 2. See detailed card above.
11 DoF (6+5), only 383g but lifts 20kg (grip-to-weight 52.6). Three variants: Basic ($5K), Pro ($6.7K), Touch ($8.4K with 3D multimodal tactile). Harvard BCI heritage. Ships on Unitree H2, G1-D, R1. See detailed card above.
Custom tendon-driven hand designed for general-purpose manipulation. 20+ DoF. Core IP of Sanctuary AI — designed specifically for their "Carbon" AI system. Demonstrated retail tasks at Mark's stores.
Biomimetic musculoskeletal hands using McKibben-style pneumatic artificial muscles. Replicates human tendon routing. Highly compliant but difficult to control precisely. Unique approach in the field.
| Hand ↕ | Institution | DoF | Actuation | BOM ↕ | Sim | Tactile | License | Key Feature |
|---|
Software Ecosystem Leader — Deepak Pathak & Kenneth Shaw
20 DoF with hybrid rigid-soft architecture: 3D-printed rigid bones wrapped in compliant TPU exterior mimicking human skeletal structure. Powered palm articulations (one spanning four fingers, one near thumb) for natural palmar arch flexibility. BOM ~$5,080 with recommended CNC metal palm upgrade.
Software: Python/C++ APIs, tuned Isaac Gym/Lab environments, Apple Vision Pro and Manus glove teleoperation. Most mature open-source hand software stack available.
Supply chain risk: Dynamixel servo dependency creates risk for strict domestic manufacturing.
Closest Wuji Analog — Direct-drive BLDC, perception-first
20 DoF, fully direct-driven. Physical architecture very similar to Wuji Hand. Primary contribution: co-optimized hardware-level perception framework that stably integrates wrist-mounted vision, fingertip tactile sensing, and proprioception with sub-7ms latency and perfect spatial alignment.
Perception integration has demonstrated superior performance in training diffusion policies compared to prior baselines. The most relevant open-source reference for replicating Wuji's actuation approach.
Durability Champion — Crash-proof poppable joints
17 DoF, tendon-driven with integrated FSR tactile sensors. Unique "popping joint" design: under excessive load, joints safely dislocate rather than breaking, protecting actuators. Survives 10,000+ continuous cycles (~20 hours non-stop manipulation) without hardware failure.
Auto-calibration via low-friction tendons routed through center of rotation. BOM under $2,200. Ideal for early RL training where crashes are frequent.
16 DoF. Direct-drive Dynamixel servos. BOM ~$2,000. MIT license. Strong simulation support with Isaac Lab and MuJoCo. Predecessor to v2 — simpler design, widely replicated. No tactile sensing in base design.
15 DoF. Tendon-driven with 11 Dynamixel servos. BOM $1,300. MuJoCo simulation support. Compact design optimized for research on in-hand manipulation. Open-source CAD and firmware.
16 DoF (7 independently actuated). Tendon-driven. BOM only $314 — most affordable multi-finger hand. CC BY-SA license. MuJoCo support. Tendon force sensing. Excellent entry point for researchers on a budget.
16+5 DoF (21 total). Cable-and-pulley actuation. BOM ~$400. MIT license. ROS 2 support. Built-in force sensors. Designed for bimanual dexterous manipulation research. Excellent value for dual-hand setups at ~$800 total.
At each interphalangeal joint, the BLDC motor acts as the structural bone itself. A 90-degree micro-worm gear converts high-speed rotation into powerful flexion/extension. Worm gears are inherently self-locking: when power is cut, thread geometry prevents back-driving. The hand maintains massive grip force indefinitely without consuming power or generating heat.
For surgical instruments requiring 30-60 min continuous retraction, self-locking is transformative vs. standard planetary/direct-drive requiring continuous motor current.
At metacarpophalangeal joints (primary knuckles), a four-bar linkage mechanism allows a significantly larger, more powerful motor housed in the palm base, generating up to 3x the torque of finger joints. This dual architecture — worm gears for fingers, linkage-amplified motors for knuckles — explains how Wuji achieves 15N fingertip force and 10-20 kg full-hand grasp with human-scale motors.
Tendon-driven hands (Shadow at $100K+, Tesla Optimus) suffer from a critical sim-to-real gap: cable tension varies dynamically with friction, stretching, temperature, and wear, making accurate simulation extremely difficult. Wuji's in-joint actuation eliminates these nonlinearities.
Note: Direct-drive reduces only the actuation modeling gap. Contact dynamics, sensor noise, environmental variation, and object properties still require domain randomization during policy training.
No open-source true BLDC direct-drive hand with in-phalanx worm gears exists today.
The RAPID Hand is the closest analog, but replicating Wuji's self-locking worm gear mechanism at IP joints requires custom motor-transmission integration from scratch. This represents the key engineering challenge for any custom build.
Vision-Based Tactile Sensors (VBTS) — "feel by seeing" — use a clear silicone elastomer pad with reflective membrane, RGB LEDs, and miniature CMOS camera to produce 1,000+ tactile pixels matching camera resolution, achieving SharpaWave-level sensitivity without complex wiring.
| Sensor | Type | Cost | Source | Key Feature | Resolution |
|---|
| Transmission | Best For | Advantages | Disadvantages | US Source | Cost/Joint |
|---|
Domestic suppliers for manufacturing dexterous hands without relying on foreign smart servos (Dynamixels) or Chinese micro-motors. The US hosts a sophisticated aerospace/medical motor base pivotable to humanoid robotics.
| Supplier | Products | Relevance | Key Advantage |
|---|---|---|---|
| Allied Motion / Allient | PerformeX slotless BLDC, HT/Megaflux frameless torque motors | Hand finger motors (15-20mm) | 2x performance of conventional motors at equivalent size; patented slotless design eliminates cogging |
| Portescap | Brushless slotless mini motors | IP joint actuation | Originally designed for surgical power tools — withstands sterilization and continuous operation; ideal thermal management |
| Kollmorgen | Frameless servo motors | Palm-housed MCP motors | Aerospace/satellite heritage; failure-proof operation; exceptional torque density with minimal EMI |
| Teknic | ClearPath integrated servos | MCP four-bar linkage actuation | Built-in encoders and drivers; slightly larger form factor ideal for palm housing |
| mjbots | moteus-c1 controller ($69) | FOC motor control for any BLDC | Open-source HW/FW; 38x38mm; CAN-FD; pairs with any BLDC motor; MIT license |
| Material | Application | Properties | US Supplier |
|---|---|---|---|
| CF-Nylon (PA12+CF) | Hand skeleton, phalanges | Carbon fiber reinforced nylon; extreme rigidity + impact resistance; lighter than metal | 3DXTech (Michigan), MatterHackers (CA) |
| PEEK | Gear components, motor mounts | Mechanical properties rivaling aerospace aluminum; chemical resistance; 134C autoclave safe | 3DXTech (Michigan) |
| TPU | Compliant finger exteriors | Passive rubber-like compliance for safe grasping; hybrid rigid-soft (proven by LEAP v2) | MatterHackers (CA) |
| CF-Nylon SLS/MJF | Production structural parts | Near-isotropic ±0.2mm tolerance; no support structures; complex pre-assembled linkages | Protolabs, RP America |
| Al 6061-T6 | Joint pins, actuator housings, wrist | General-purpose CNC aluminum for structural components | Xometry, Protolabs |
| Ti Grade 5 | Surgical biocompatible components | ISO 10993-1 biocompatibility for surgical applications | Xometry, Protolabs |
Fast Path (5-6 months) — LEAP v2 Advanced architecture + GelSight fingertips + ORCA poppable joints
| Dynamixel XC330 (x16) + XM430 (x4 MCP) | $2,760 |
| DIY GelSight fingertip sensors (x5) | $300 |
| MJF CF-Nylon structural parts | $500 |
| CNC aluminum joint pins + wrist | $600 |
| U2D2 + Power Hub + wiring | $200 |
| TPU compliant exteriors | $50 |
| Fasteners, tendons, misc | $150 |
| TOTAL (Prototype) | $4,560 |
| 50-Unit Cost | $3,060 |
Production Path (7-10 months) — US BLDC + worm gears (IP) + strain wave (MCP)
| US/Chinese BLDC motors 12-18mm (x20) | $1,500 |
| Micro-worm gears IP joints (x16) | $500 |
| Strain wave gears MCP joints (x4) | $600 |
| AS5048A encoders input+output (x40) | $200 |
| Custom FOC driver PCBs TMC4671 (x20) | $700 |
| STM32H7 central controller | $150 |
| CNC + SLS CF-Nylon structural | $2,200 |
| DIY GelSight sensors (x5) | $300 |
| Bearings, fasteners, wiring, TPU | $400 |
| TOTAL (Prototype) | $6,550 |
| 50-Unit Cost | $2,900 |
| NRE (one-time) | $35-55K |
| Fully-loaded unit @50 | $3,600-4,000 |
| Component | Tool | Notes |
|---|---|---|
| Training Framework | NVIDIA Isaac Lab v2.3 | GPU-native, PyTorch tensor interface, automatic domain randomization |
| Physics Engine | NVIDIA Newton / PhysX 5 | 313x faster than MJX for manipulation tasks |
| Hand Sim Model | LEAP Hand Isaac Lab | github.com/leap-hand/LEAP_Hand_Isaac_Lab |
| Tactile Simulation | TacSL (NVIDIA) | GelSight in Isaac Lab, 200x faster than prior methods |
| Tactile Sim Alt. | TACTO (Meta AI) | Renders DIGIT output in physics engines at 100s FPS |
| Policy Learning | 3D-ViTac methodology | Fuses visual + tactile into unified 3D point cloud for diffusion policy |
| Surgical Tasks | ORBIT-Surgical | 14 benchmark tasks: needle grasp, suture, handover |
| Teleoperation | Apple Vision Pro / Manus gloves | Demonstration collection for imitation learning |
| Motor Control | SimpleFOC + STM32G4 | Open-source FOC framework; dual AS5048A encoders per joint |
| Communication | CAN-FD (5 Mbps) | 20 motors @1kHz = 2.048 Mbps; ~$1/node via MCP2562FD |
| Phase | Actuator | Per-Unit | 20-Axis Total | Force | Sim Gap | Notes |
|---|---|---|---|---|---|---|
| Phase 1 | Dynamixel XC330 | ~$90 | $1,800 | 8-12N | Moderate | Fastest start; LEAP ecosystem works out-of-box |
| Phase 1+ | Dynamixel XM430 (MCP) | ~$330 | 4x = $1,320 | >20N | Moderate | Force upgrade for MCP joints only |
| Phase 2 | US BLDC + worm gear | ~$150-300 | $3,000-6,000 | 15-25N | Low | Self-locking, surgical-grade, domestic supply |
| Budget | Chinese BLDC + moteus | ~$100-150 | $2,000-3,000 | 15-25N | Low | Best value for QDD path |
github.com/leap-hand — CAD, API, Isaac Lab envs, teleoperation
github.com/facebookresearch/digit-design — PCB layouts, firmware, 3D models, BOM
github.com/facebookresearch/tacto — Vision-based tactile simulation in PyBullet
orca.ethz.ch — Poppable joint design, auto-calibration
arxiv.org/abs/2506.07490 — 20-DoF direct-drive, perception framework
binghao-huang.github.io/3D-ViTac — Unified visuo-tactile policy learning
simplefoc.com — Open-source FOC motor control library
github.com/mjbots/moteus — Open-source HW/FW BLDC servo controller ($69)
orbit-surgical.github.io — 14 benchmark surgical tasks in Isaac Lab
hackaday.io/project/165599 — 30:1 ratio harmonic drives, <$20/joint
Northwestern, CMU, MIT, Florida A&M, Texas A&M — $26M+ center for industry collaboration. PI: Ed Colgate / Kevin Lynch
any-skin.github.io — Plug-and-play replaceable magnetic tactile skins
Dominant or most likely gear/actuation families per vendor. Confidence: High = public disclosure; Medium = inferred from teardowns/patents; Low = best-read assumption.
| Company | Gear / Actuation Type | Confidence |
|---|
| Layer | Typical Players | Strategy |
|---|---|---|
| "Brains" — Multimodal LLMs | OpenAI, xAI, Google DeepMind, Anthropic, NVIDIA, Meta, Microsoft | Scale frontier models; retune from text-only to VLA control; license/open-source interfaces for robot makers |
| "Bodies" — Humanoid OEMs | Tesla, Figure AI, Agility, Apptronik, 1X, Sanctuary AI, Fourier, Unitree, UBTECH | Prove reliable mechatronics; raise massive rounds; sign first commercial pilots |
Led 1X Series A & Figure Series B. Training GPT-4o-Vision + Voice on humanoid sensor data.
Grok 4 co-developed with Optimus Gen-3. Vision-language-action control into real-time sensor fusion.
RT-2 → RT-X → Gemini Robotics-ER (2025). Partnered with Apptronik for humanoid trials.
GR00T N1 open VLA model on Fourier GR-1. Pull-through for Thor & Jetson hardware. Isaac Sim ecosystem.
Azure Robotics Platform. Invested in Figure ($675M round) and Sanctuary research MoU.
Reality Labs Robotics division. LLaMA-3 VLA models. Early talks with Unitree & Figure.
| Axis | 1X World Model (1XWM) | Figure Helix |
|---|---|---|
| Purpose | Generative simulator ("dreaming engine") — predicts future sensor frames | Real-time VLA policy — converts instruction to 200Hz whole-body control |
| Architecture | Diffusion-based video generator + action conditioning | System-2 (7B VLM @ 7-9Hz) + System-1 (80M visuomotor @ 200Hz) |
| Data | 1000s hrs on-board RGB-D, joint states, proprioceptive in homes/offices | ~500 hrs teleop data achieved broad generalisation |
| Latency | Offline (GPU-intensive) | 5-10ms end-to-end on-device |
| Key Advantage | Ground-truth physics; synthetic data flywheel for edge cases | On-device privacy; meets ISO 10218-2; runs under 50W |
LLM vendors publish certified "robot brains" APIs; OEMs compete on price, dexterity, and vertical-market integrations.
One or two hyperscalers bundle compute + model + reference robot and commoditise the rest.
Breakthrough in cheap actuators lets OEMs detach from cloud models and train small on-device VLAs.
Dual-use export controls choke cross-border supply chains; adoption tilts to regulated domestic pilots only.
| Standard | Domain | Relevance to Humanoids |
|---|---|---|
| ISO 10218-1/-2 | Industrial | Machinery safety for articulated industrial robots |
| ISO/TS 15066 | Collaborative | Permissible contact pressure for human-robot collaboration |
| IEC 61508 / ISO 13849 | Functional Safety | SIL/PL targets for control systems (torque-limited joints, brakes) |
| UL 4600 Ed. 3 | Autonomy | Safety case for autonomous mobile operations |
| ISO 13482:2014 | Personal Care | Force/speed caps for personal-care robots in homes |
| IEC 60335-1 | Domestic Electrical | Clearances, creepage, over-current for household devices |
| ISO/WD 25785-1 | Mobile Industrial | Safety for industrial mobile robots with actively controlled stability |
| ISO/IEC 27001 | Cybersecurity | Certifiable framework for corporate security policies |
| IEC 62443 | ICS Security | Cybersecurity for industrial automation control systems |
| GDPR | Data Privacy | Consent, data minimization, encryption for robot-collected data |
| EU AI Act | AI Regulation | High-risk classification; risk management, transparency, human oversight |
| NIST CSF | Cyber Framework | Identify, Protect, Detect, Respond, Recover for robot deployments |
Lightweight CFRP skeleton, joint travel hard-stops, energy-absorbing skins. ISO 12100.
Series-elastic actuators, velocity & torque monitoring, zero-power brakes. ISO/TS 15066.
Multi-modal depth cameras + LiDAR for 360° occupancy; tactile skins. UL 4600.
Double insulation, PELV 48 VDC buses, surge suppression. IEC 60204-1.
Secure boot, OTA signing, GDPR-compliant data handling. ISO/IEC 27001.
Predictable motion, intuitive intent signalling (LED, speech), teachable modes. ISO 9241-210.
Attackers gain control via network or malicious firmware. Demonstrated on SoftBank NAO.
Cameras/mics coerced into capturing personal data. Unauthorized sensor access.
Poisoned training data makes robots learn unsafe behaviors.
Crafted visual/audio inputs fool perception systems into misclassifying objects.
| Policy Lever | Evidence | Intended Effect |
|---|---|---|
| Industrial Plan | MIIT guideline: "realise mass production by 2025" | Fast-track R&D, rally local governments & SOEs |
| Capital Mobilisation | 83 funding events in H1-2025 raised ≈¥6.9B (~$950M) | Flood sector with patient money |
| Champions & IPOs | Unitree preparing 2025 domestic IPO | Anchor supply chains and talent |
| Domestic Compute | Huawei Ascend 910 chips; courting buyers abroad | Reduce dependence on Nvidia/AMD |
| Track | Action | Strategic Aim |
|---|---|---|
| Export Controls | New BIS rules (Apr 2025) stop Nvidia H20 to China (~$8B revenue hit) | Deny China most efficient training/inference silicon |
| Outbound Investment | Treasury rule under EO 14105 restricts U.S. capital in China's AI | Prevent "know-how by cheque" |
| Domestic Incentives | CHIPS Act + DoD AI Rapid Capabilities Cell (Nov 2024) | Keep advanced pipelines onshore |
| Allied Market Access | BIS eases restrictions for trusted partners | Lock U.S.-centric supply chain |
Two incompatible HW/SW ecosystems; export bans expand to servomotors and Li-ion packs.
U.S. chips in Chinese humanoids under license caps; JVs in neutral markets (Mideast logistics).
Tit-for-tat curbs hamper global supply, slowing overall adoption; EU/Japan becomes "swing supplier."
Both powers relax on civilian lines (elder-care, retail) while cordoning off military uses.
| # | Company | Humanoid Project | Ticker | Key Notes |
|---|---|---|---|---|
| 1 | Tesla Inc. | Optimus production robots | NASDAQ: TSLA | Highest-volume ticker; targeting 1000s of units in 2025 |
| 2 | Xiaomi Corp. | CyberOne (Robotics Lab) | HKEX: 1810 | Low entry price, deep trading pool; consumer-electronics synergies |
| 3 | Hyundai Motor | 80% owner of Boston Dynamics | Korea SE: 005380 | Only pure industrial play with direct BD stake |
| 4 | SoftBank Group | Pepper v2 & Vision Fund stakes | OTC: SFTBY / Tokyo: 9984 | Diversified holding; liquidity best via ADR |
| 5 | Toyota Motor | T-HR3 & "Musashi" R&D | NYSE: TM | Partner-robot platform for healthcare & mobility |
Figure 01 general-purpose bot. Highest flow on Hiive among robotics names.
NEO home-service bot. OpenAI-backed.
Phoenix dexterous workspace bot. Canadian leader; regular secondary blocks.
Apollo manufacturing assistant. Mercedes-Benz pilot in 2025.
Digit warehouse & logistics bot. Frequent bids/asks; recently raised Series C.
| Vendor | Platform | Price | License | GitHub Stars | Discord | Units/Field | 2025 Milestones |
|---|---|---|---|---|---|---|---|
| Pollen Robotics (🇫🇷) | Reachy 2 | €29K; 100+ shipped | GPL v3 / CERN-OHL | 2,100 | 4,000 | 100+ | CES 2025 launch; Hugging Face acquisition (Apr 2025) |
| K-Scale Labs (🇺🇸) | K-Bot / Z-Bot | $8,999; Nov 2025 | MIT + CERN-OHL-S | 144 | 2,300 | Pre-orders | First YC-backed humanoid; full RL locomotion + VLA alpha |
| UC Berkeley (🇺🇸) | Berkeley Humanoid Lite | <$5K self-build | BSD-3-Clause | 820 | 1,100 | 200+ BOM downloads | RSS 2025: zero-shot sim-to-real walking + cube manipulation |
| InMoov Community | InMoov (1.8m) | ~$1,500 DIY | CC-BY-NC | 5,800 | — | 10,000+ partial | World's largest DIY humanoid community |
| Poppy Project (🇫🇷) | Poppy Humanoid | €8K kit | GPL v3 / CC-BY-SA | — | — | 60+ research labs | ROS 2 port Feb 2025; €1.6M EU Horizon grant |
| NimbRo (Univ. Bonn 🇩🇪) | NimbRo-OP2X | €35K research | GPL v3 | — | — | Research | AGILOped redesign for RoboCup AdultSize |
All vendors ship MuJoCo or Isaac Gym environments. Millions of control iterations before hardware moves.
NVIDIA GR00T N1 (open weights, Mar 2025) and Hugging Face twin humanoids (May 2025).
ROS 2 Foxy/Humble messaging backbone. Real-time control migrating to Rust for safety (K-Scale K-OS).
Dual-stack copyleft pattern dominates. BSD/MIT for academic spin-outs.
| Phase | Unit Volumes | Price Band (USD) | Typical Buyer | Success Metric |
|---|---|---|---|---|
| 2025-27 Pilot Era | Hundreds → low thousands | $50-150K | Enterprise pilots | Reliability ≥ 95% |
| 2028-32 Early Scale | Tens of thousands/yr | $20-50K | Large-fleet industrial + premium homes | 2-yr payback |
| 2033+ Mass Uptake | Millions/yr | < $15K | Mainstream SMEs & households | Human-parity dexterity |
Pilot fleets (>1,000 units); 95% uptime; ISO draft standard.
BoM ≈ $10K; consumer beta units for cleaning & laundry.
BoM < $5K; multi-million installed base; services > hardware revenue.
Human-parity dexterity; solid-state batteries; trillion-dollar services economy.
Atlas-era $1M robots → sub-$6K production units. >90% drop in 6 years; $20K feasible before 2030.
Logistics contracts already fund multi-thousand-unit pipelines. Consumer volumes lag ~5 years.
Cleaning, laundry, and garden chores top every wish-list. Companionship is the only non-chore request.
98% grasp reliability outranks back-flip videos. Manipulation mastery is the king-maker capability.
Early movers in safety and privacy regulation will lock in advantage. Ecosystem consolidation inevitable.
| Stakeholder | Do Now (2025-27) | Prepare For (2028-32) |
|---|---|---|
| Investors | Diversify across brain-body-integrator; insist on ≤24-mo pilot payback | Double down on emerging winners; safe + capable home robots |
| Robot OEMs | Nail one paid task at 99% uptime; lock actuator supply | Scale volume; open app store; drive BoM < $20K |
| Component Providers | Co-design standard actuators/sensors with top OEMs | Offer integrated sensor bundles |
| AI/Software Providers | Collaborate on perception stack and task autonomy | License perception-as-a-service; integrate cloud & edge |
| Enterprise Adopters | Run 3-6 mo pilots; capture downtime & safety KPIs | Negotiate fleet leases; integrate with WMS/ERP |
| Policy-Makers | Draft baseline mobile-service-robot rules; fund re-skilling | Establish certification labs; monitor labour displacement |
| Positive Trigger | Negative Trigger |
|---|---|
| Sub-$30K retail from Tier-1 OEM | High-profile home accident → moratorium |
| First pure-play humanoid IPO | Actuator/battery supply crunch |
| ISO humanoid safety standard issued | Insurance premiums spike after injury lawsuit |
| Multi-modal AI masters reliable folding & cooking | Home robot privacy breach |
| National "robot driver-licence" framework passes | Actuator supply crunch delays shipments >12m |