Compendium

Walkthrough — Design an Industrial Humanoid Robot Full Stack

16 min read

Walkthrough — Design an Industrial Humanoid Robot Full Stack

End-to-end design of a bipedal humanoid robot for industrial + commercial deployment: morphology + actuation + power electronics + battery + sensors + compute + ROS 2 + Nav2 + MoveIt 2 + GR00T-class foundation-model policies + safety + RaaS commercial model. Target customer: Tier-1 automotive (BMW, Mercedes-Benz), e-commerce 3PL (Amazon, GXO, DHL), light manufacturing (Spanx, Magna). Spec roughly matches Figure 02/03, Apollo (Apptronik), Digit (Agility), and Optimus Gen 2.

Adjacent walkthroughs: design-autonomous-warehouse for the facility that orchestrates humanoid + AMR fleets; design-autonomous-mobile-robot for the simpler wheeled-base counterpart.

1. What we’re building

A 1.7 m (5.6 ft), 60 kg (132 lb) electric bipedal humanoid with two 7-DoF arms, dexterous five-finger end-effectors, stereo vision + LiDAR + IMU sensing, on-board NVIDIA Jetson Thor compute, 2.5 kWh LFP battery for 4-hour continuous operation, designed for warehouse + light-manufacturing tasks: bin-to-bin transfer, kit assembly, machine tending, palletizing.

Target tasks (production-validated 2024–2026):

  • BMW Spartanburg + Mercedes-Benz Berlin: Figure + Apollo for sub-assembly transfer + parts kitting
  • Amazon Spanaway (WA): Digit (Agility Robotics) for tote movement in a fulfillment center
  • GXO + DHL: Stretch, Digit, Apollo for case-handling + truck-unload
  • Spanx (Sanctuary AI + Magna): Phoenix v9 for textile-handling cells
  • Hyundai (parent of Boston Dynamics): Atlas Electric for factory roles, post-2024 launch
  • Tesla Gigafactory Austin + Berlin: Optimus internal production support, 2024–2026 pilot scale
  • Kepler PRIME + EngineAI SE01 + UBTECH Walker S + Unitree H1/G1/H2 + 1X NEO Beta/Gamma: China + global pilots

Reference platforms in this class: Figure 02/03, Apollo (Apptronik), Atlas Electric (Boston Dynamics, Hyundai), Digit (Agility Robotics), Phoenix v9 (Sanctuary AI), Optimus Gen 2 (Tesla), NEO Beta/Gamma (1X), Unitree H1/G1/H2, Fourier GR-1/GR-2, Kepler K1/K2, Booster T1, EngineAI SE01, Galbot G1+, UBTECH Walker, MagicLab MagicBot, LimX Dynamics CL-1/P1, Astribot, Mentee, Vicarious Surgical (surgical humanoid offshoot).

2. Spec table

ParameterValueNotes
Height1.7 m (5.6 ft)adult-male median, fits human workspaces
Mass60 kg (132 lb)aggressive vs Atlas Electric 89 kg, Optimus 73 kg
Total DoF36 (2 legs × 6 + 2 arms × 7 + 2 hands × 5 + 1 torso + 2 head)minimum for industrial dexterity
Arm payload (per arm)15 kg (33 lb) at full reach, 25 kg (55 lb) close-inmatches industrial part-handling
Walking speed1.5 m/s (3.4 mph) sustained, 2.5 m/s peakhuman-pedestrian range
Stair capabilitystandard 18 cm rise × 28 cm treadOSHA stair geometry
Runtime4 hr continuous, 6 hr light dutyLFP, opportunity charge
Battery48 V × 50 Ah = 2.4 kWh LFPthermal runaway resistant
ComputeNVIDIA Jetson Thor 2025 (~2000 TOPS FP4)onboard inference + planning
SensorsStereo RGB-D × 3 + LiDAR + IMU + tactile fingers + F/T wristsredundant perception
Safety standardISO 10218-1/2 industrial + ISO/TS 15066 cobot + ISO 13482 personal-carelayered cert
Functional safetyISO 13849 PL d Cat 3 minimumbrakes, e-stop, force-limited collision
Ingress protectionIP44 (V1), IP55 (V2 outdoor)indoor-only initial
Noise<70 dBA @ 1 mhuman-coexistent
MSRP (purchase)$50–150k USDper-unit hardware cost
RaaS (lease)$3–7k/mo + $10–20k/yr servicetypical 3-year term
Operating envelope0–40 °C (32–104 °F), 10–90% RH non-cond.indoor industrial

The 1.7 m height + 60 kg mass + 4 hr runtime + ~36 DoF puts this design squarely in the 2025-generation industrial humanoid band. Aggressive cost targets ($50–150k) require quasi-direct-drive (QDD) actuation rather than the harmonic-drive + servo paradigm of Atlas Electric.

See humanoid-balance + legged-locomotion for the locomotion foundations; legged-morphologies for the morphology trade space.

3. Morphology + kinematics

The 36-DoF breakdown:

  • Legs: 6 DoF per leg × 2 = 12 DoF. Hip (3), knee (1), ankle (2). Some platforms use 5 (no ankle roll) for cost.
  • Arms: 7 DoF per arm × 2 = 14 DoF. Shoulder (3), elbow (1), wrist (3). Redundancy is essential for obstacle avoidance + null-space task control.
  • Hands: 5–15 DoF per hand. Underactuated tendon-driven (Schunk SVH 9-DoF, Allegro Hand 16-DoF, Shadow Dexterous 24-DoF, Sanctuary 21-DoF, Pisa-IIT SoftHand) or fully-actuated. 11+ DoF per hand for dexterous manipulation.
  • Torso: 1–2 DoF (waist yaw + optional pitch). Enables forward bending + reach without losing balance.
  • Head: 2 DoF (pan + tilt) carrying stereo cameras + LiDAR.
  • Wrist F/T: 6-axis F/T sensor at each wrist (ATI Mini45 or Robotous RFT) for force-controlled manipulation.

Total mass distribution target (60 kg total):

RegionMass
Pelvis + battery + main compute18 kg
Each leg7.5 kg × 2 = 15 kg
Torso + waist actuator10 kg
Each arm + hand6 kg × 2 = 12 kg
Head + sensors5 kg
Total60 kg

Center of mass at ~1.0 m above floor when standing. Stability margin: 0.05 m (50 mm) static; dynamic ZMP-based balance per humanoid-balance expands the dynamic envelope.

Kinematic frame conventions per kinematics-dh with DH parameters published in URDF format for ROS 2 consumption.

4. Actuation

The actuator family is the single largest cost + performance driver. Three paradigms in production:

  • Harmonic + frameless servo (Atlas Electric, Optimus): Harmonic Drive CSF/CSG strain wave gears at 50–160:1 reduction + Kollmorgen/Maxon frameless servos. Highest precision + backdriveability low; ~$2–5k per actuator.
  • Cycloidal (Nabtesco RV, Sumitomo Cyclo): high torque density, moderate efficiency, used in heavy-duty industrial cobots. Heavier than harmonic.
  • Quasi-direct-drive (QDD): lower-reduction (6–10:1) planetary + high-torque BLDC. Backdriveable, high-bandwidth, lighter, lower cost. Used by ANYbotics, Boston Dynamics Spot, Apptronik Apollo (custom QDD “Mover”), ETH spinoffs. ~$500–1500 per actuator.

We pick QDD for legs + arms (cost + backdriveability + impact tolerance) and harmonic drive for the high-precision wrist + hand mechanisms (low backlash for manipulation). Approximate actuator BOM:

  • Leg QDD (×12): $800 each = $9.6k
  • Arm QDD (×14): $700 each = $9.8k
  • Hand servos (×22 = 11 per hand): $200 each = $4.4k
  • Wrist F/T (×2): $3k each = $6k
  • Misc torso + head: $3k
  • Actuator subtotal: $33k

Reference QDD vendors: ANYdrive (ANYbotics), MyActuator RMD-X8 + RMD-L series (open + commodity QDD), T-Motor AK series, CubeMars AK10-9, Halodi Robotics REDcurrent, MJBots/Mr. Robot mj5208.

Custom QDD development: bigger players (Figure, Apptronik, Tesla) build their own QDD with in-house BLDC stators + custom planetary gears + integrated drive electronics + thermal management. Per-actuator BOM in volume drops to ~$200–400 with in-house production.

5. Power electronics

Per-actuator drive electronics:

  • Gate drivers: TI UCC21520, Infineon EiceDRIVER, ST L6388.
  • MOSFETs / IGBTs: Infineon CoolMOS / IRF / OptiMOS, ST STripFET, Onsemi NTHL — increasingly GaN (gallium nitride) for higher switching frequency (~200 kHz vs 20 kHz Si) and SiC (silicon carbide) for higher-voltage / lower-loss inverters. 48 V GaN inverter modules at 30 A continuous are commodity at ~$10–30 per axis.
  • Controllers: TI C2000 family (TMS320F28379D), STM32G4 / H7, NXP S32K, Microchip dsPIC33.
  • Current sense: Shunt + Allegro ACS780 or AMS Hall-effect inline; in-phase + sum-point shunts.
  • Encoders: AMS AS5048A / AS5147 magnetic absolute (14-bit) at the motor shaft + Renishaw AksIM2 absolute (20-bit) at the joint output for closed-loop joint feedback.

Total power-electronics BOM per actuator: ~$50–100. Across 36+ actuators: $2–4k per robot.

The drive boards are mounted integrated with each actuator (“smart actuator” pattern, like ANYdrive or T-Motor) to minimize wiring + EMC issues. Communication: EtherCAT (10 μs deterministic cycle, daisy-chain through the body) or EtherCAT-over-Coax-substitute for higher-noise environments. Some platforms use CAN-FD as a cost-down alternative at 1 ms cycle.

See motor-drive-electronics for the inverter taxonomy and passive-components for the EMC + filtering parts.

6. Battery + power management

Pack: 48 V nominal × 50 Ah = 2.4 kWh, ~16 kg, fitted in the pelvis cavity. Cell chemistry: LFP (LiFePO₄) for thermal runaway resistance, 3000+ cycle life, ~$140/kWh pack cost. Cell: EVE LF50K prismatic or CATL 50 Ah cylindrical.

Pack architecture: 15 series × 1 parallel (15S1P) of 50 Ah cells = 48 V nominal × 50 Ah = 2.4 kWh. Total pack 16 kg. BMS: custom 15S smart BMS with cell balancing + overcurrent + thermal protection + SOC estimation + isolation monitoring. Reference: Orion BMS 2, REC Active BMS, Daly Smart BMS 16S 200A ($300).

Power distribution:

  • 48 V bus to actuator drives (direct)
  • 48 V → 24 V DC-DC (Vicor) for cooling fans + LiDAR
  • 48 V → 12 V DC-DC for compute auxiliary
  • 48 V → 5 V for sensors + USB
  • Hot-swap connector at the pack for field battery exchange in <30 sec

Power budget at 50% locomotion duty + arm manipulation:

SubsystemContinuous W
Locomotion (6 active leg actuators @ 50 W)300
Arm (6 active arm actuators @ 30 W)180
Compute (Jetson Thor + carrier)80
Sensors (LiDAR + 3 cameras + IMU)25
DC-DC losses + auxiliaries25
Total~610 W

Runtime at full continuous: 2400 Wh / 610 W = 3.9 hr sustained. Light-duty (static stand + occasional manipulation) drops to ~200 W → 12 hr.

Battery shipping: UN 38.3 + IATA DGR Section II for ≤100 Wh single cell (irrelevant here) or full Class 9 hazardous-goods for the 2.4 kWh pack. Adds ~$200 to ship one robot internationally and requires UN 3480 + UN 3481 documentation. Operational thermal runaway prevention: cell-level fuses, pack-level pyro disconnect, ceramic fire-suppression barrier between cells.

7. Sensors

Perception stack — multi-modal redundancy is the norm in safety-critical humanoid deployments:

  • Stereo RGB-D × 3: head + chest + lower-front. Options: Intel RealSense D455 (90° FOV, 0.4–20 m, IMU integrated), D435, L515 (LiDAR-based depth, less common post-2023 discontinuation), ZED 2i (Stereolabs, neural stereo), Luxonis OAK-D + OAK-D Lite (integrated VPU), Orbbec Femto Bolt (low-cost ToF).
  • LiDAR: head-mounted Ouster OS1-32 (32-beam, 100 m, 1.5 kg) or Hesai PandarXT-32 or Velodyne Puck Lite (legacy). Some platforms (Figure, Optimus) skip LiDAR and rely entirely on vision — saves ~$3–8k.
  • IMU: pelvis-mounted, Bosch BMI270 (low cost, 9-DoF), Honeywell HG4930 (industrial, ~1°/hr bias stability), Analog Devices ADIS16475 (tactical-grade, ~0.01°/hr), ST ISM330 (cost-effective MEMS). Industrial cert needs tactical grade.
  • Joint encoders: RLS (Renishaw subsidiary), Bourns, Renishaw RESOLUTE, Avago AEAT-9000, iC-Haus iC-MU/MH. Dual encoders (motor-side + joint-side) for closed-loop control + safety cross-check.
  • F/T sensors: ATI Mini45 ($3k each, ±240 N / ±10 N·m, used on each wrist) or Robotous RFT alternates.
  • Tactile fingers: emerging — GelSight Mini, SynTouch BioTac, Contactile, Tacterion. Optical tactile (GelSight) is the production reference for 2026 designs.
  • Microphones: array of 4-6 MEMS mics for voice + sound localization.
  • Joint torque (in some platforms): strain gauge or current-derived per-joint torque estimate.

Total sensor BOM: ~$10–15k per robot at production volume.

See sensors-perception + perception-sensors + sensor-families for the catalog.

8. Compute

The single largest compute decision in 2024–2026 is Jetson AGX Orin vs Jetson Thor:

  • NVIDIA Jetson AGX Orin (2022): 275 TOPS INT8, 60 W typical, $1.5k module. The 2023–2024 baseline.
  • NVIDIA Jetson Thor (2025): ~2000 TOPS FP4 (LLM-optimized), 130 W, ~$3–5k. Designed for humanoid + autonomous-vehicle inference at scale. The 2025+ choice.
  • Apple M-series + custom: rare in production (Figure briefly used M2 + Jetson hybrid).
  • Qualcomm RB6 Robotics Platform: 15 TOPS at 7 W, edge-only. Suitable for small humanoids or auxiliary compute.
  • Hailo-8/15: 26 TOPS at 2.5 W, dedicated inference accelerator. Often paired with a Jetson for high-throughput perception offload.
  • Google Coral Edge TPU: 4 TOPS at 2 W; commodity edge inference.
  • Tenstorrent Grayskull e75/e150: 92–315 TFLOPS, RISC-V architecture, ~$800–2400. Emerging alternative for foundation-model inference.

Our design uses Jetson Thor primary (perception + planning + foundation-model policy inference) + Hailo-15 auxiliary (always-on tactile + audio + ambient-vision). Total compute BOM ~$5–8k.

Compute architecture pattern (post-2024):

  • Foundation-model policy (GR00T-class, see §10) runs on Jetson Thor at 30 Hz inference
  • Low-level whole-body controller (200 Hz) runs on a separate FPGA (Lattice MachXO5) or RT MCU (STM32H7)
  • Safety supervisor runs on a hardware-redundant pair of MCUs (cross-monitor, like in industrial PLC architecture)

See inference-optimization + llm-landscape for the foundation-model compute side.

9. Software stack

The 2025 humanoid software canon:

  • ROS 2 Jazzy (2024 LTS): middleware + node graph + lifecycle. See ros2-architecture.
  • Foxglove Studio: visualization + introspection.
  • MoveIt 2: motion planning for arms. See manipulator-design.
  • Nav2: navigation for wheeled-base systems; partially adapted for humanoid locomotion.
  • Open-RMF: multi-fleet orchestration (when fielded alongside AMRs in a warehouse — see design-autonomous-warehouse).
  • Drake (MIT): dynamics simulation + trajectory optimization, optimized for legged robots.
  • MuJoCo MPC: open-source MPC framework optimized for MuJoCo dynamics. Heavily used by DeepMind + Berkeley.
  • NVIDIA Isaac Sim + Isaac Lab: GPU-accelerated simulation + reinforcement-learning training environment. Sim-to-real pipeline.
  • NVIDIA Isaac ROS: GPU-accelerated perception nodes (VSLAM, depth estimation, object detection).
  • NVIDIA GR00T (2024): foundation-model platform for humanoid policies. Provides pre-trained backbones + fine-tuning infrastructure.
  • NVIDIA Cosmos (2025): world-model platform for synthetic data generation + closed-loop simulation.
  • Aloha + UMI (Stanford): imitation-learning toolchain.

The ROS 2 launch graph typically includes:

  • Perception nodes (camera drivers, LiDAR driver, VSLAM, object detector)
  • Whole-body controller (200 Hz)
  • Foundation-model policy (30 Hz, takes RGB + proprioception, outputs action distribution)
  • Safety supervisor (50 Hz watchdog)
  • Fleet client (VDA 5050-over-MQTT to fleet manager)
  • Diagnostics + telemetry export

10. Foundation-model policies (GR00T, Helix, π0, RT-X)

The 2024–2026 humanoid policy revolution: vision-language-action (VLA) foundation models trained on internet-scale video + teleop + sim data.

Major model families:

  • NVIDIA GR00T (2024): foundation backbone for humanoid policies, trained on internet video + human motion-capture + simulator rollouts. Open-weight reference: GR00T N1.
  • π0 (Pi-Zero, Physical Intelligence, 2024): 3B-parameter VLA model trained on 10k+ hours of teleop. π0.5 (2025) added open-world manipulation.
  • OpenVLA (Stanford + Berkeley + Toyota, 2024): 7B-parameter open-source VLA.
  • RT-2 + RT-X (Google DeepMind + Open X-Embodiment, 2023–2024): early VLA for tabletop manipulation. Cross-embodiment generalization.
  • Helix (Figure AI, 2025): bi-modal voice + vision-action policy on Figure 02. Production-deployed at BMW Spartanburg.
  • 1X Redwood (2024): autonomous voice + vision policy for NEO Beta.
  • Tesla Optimus learned: Tesla’s in-house imitation + RL pipeline, leveraging Optimus-internal teleop data.
  • Anthropic Claude Computer Use (2024–2025): adjacent — same VLA principles applied to desktop computing rather than embodied.
  • Boston Dynamics Atlas Electric DRL: Atlas’s whole-body deep-RL policy stack for parkour + manipulation.
  • Berkeley HumanPlus + Stanford ALOHA 2 + UMI: academic-foundation imitation-learning frameworks.
  • Sanctuary Carbon + Apptronik Apollo Voice: closed commercial VLA stacks.

Training pipeline: simulation pre-training in Isaac Lab (RL with domain randomization) + teleop fine-tuning (200–2000 hr per task) + online correction (reinforcement-learning-from-human-feedback, RLHF, in deployment).

Inference: 30 Hz policy + 200 Hz low-level controller. Latency budget end-to-end (camera shutter to action): ≤80 ms. See imitation-learning + rl-for-control + robot-learning-and-rl.

The shift from hand-engineered task scripts to general-purpose VLA policies is the single most important change between 2020-generation humanoids (DARPA Robotics Challenge era) and 2025-generation. Adoption velocity is gated by data: hours of expert teleop collected, quality of synthetic data generated in Isaac Lab, and the fine-tuning compute available.

11. Safety

The humanoid safety story is layered + still evolving. The relevant standards:

  • ISO 10218-1/2 (industrial robots) — applies to caged industrial deployment
  • ISO/TS 15066 (collaborative robots) — applies to human-coexistence cells; specifies max contact force + pressure limits
  • ISO 13482 (personal-care robots) — applies to home + retail-floor humanoid deployment; finalized 2014, scant updates
  • ISO 13849 (safety of machinery control systems) — PL d Cat 3 minimum
  • IEC 61508 (functional safety) — SIL 2 typical
  • EN 60204-1 (electrical equipment for machinery)
  • ANSI/RIA R15.06 (US industrial robots) + R15.08 (mobile + collaborative robots, 2020)
  • CE marking (Machinery Regulation EU 2023/1230, transition from Directive 2006/42/EC)
  • FCC Part 15 + FDA 510(k) if medical-adjacent + UL electrical safety

The safety architecture:

  • E-stop: dual-channel mushroom button on torso back, removes power to all actuators within 20 ms
  • Force limiting: whole-body torque limits set per ISO/TS 15066. Each actuator’s drive enforces a maximum continuous torque + a higher transient (collision-detection) threshold. On detection: immediate compliant retreat.
  • Collision-aware planning: predictive whole-body collision avoidance running at 200 Hz; if collision predicted, switches to compliant interaction mode.
  • Safety LiDAR or safety camera curtain: optional, for industrial cell deployment requiring Cat-3 protective stop independent of the main perception stack.
  • Mechanical brakes: spring-engaged, electrically-released on every active joint. Power loss → all joints lock in current position. Manual release for emergency retrieval.
  • Soft outer cover: foam-and-fabric overlay reduces peak contact pressure during incidental human contact. Per ISO/TS 15066 contact-force/pressure limits.
  • Functional safety classification: ISO 13849 PL d Cat 3 for the e-stop + brake chain; IEC 62304 Class C for the safety-supervisor software.

Emerging hazards that the standards don’t yet fully cover:

  • Falls from biped instability (Optimus tipped at Cybertron 2024 demo; Atlas falls in early demos)
  • Stair fall events (Digit + Apollo have published incident data)
  • Hand-finger pinch points during dexterous manipulation
  • Audio-induced false-positive activation (VLA models triggering on overheard speech)
  • Cybersecurity of the policy model itself (prompt-injection, adversarial-vision attacks)

The Embodied AI Safety Working Group + RIA + ISO TC 299 are actively drafting new humanoid-specific particular standards, expected 2026–2028.

12. Production economics + RaaS model

Per-unit BOM at volume 1000+:

SubsystemCost
Actuators (36 total, mixed QDD + harmonic)$33k
Power electronics (per-actuator drives)$3k
Battery + BMS + thermal$3k
Compute (Jetson Thor + Hailo aux)$6k
Sensors (cameras, LiDAR, IMU, F/T, tactile)$12k
Frame + skins + fasteners + brakes$5k
Final assembly + test + cal$8k
Software + license (per-unit allocation)$5k
BOM total~$75k

Retail/lease pricing:

  • Purchase: $50–150k depending on configuration + customer commitment volume. Tesla Optimus internal cost target ~$20–30k at million-unit scale (2027+ projection); current realistic per-unit cost $50–100k.
  • RaaS (robot-as-a-service): $3–7k/mo lease + $10–20k/yr service contract + insurance + training + integration support. Customer locks in 3-year terms. This is the dominant 2024–2026 commercial model (Agility Digit lease, Apptronik Apollo, Figure RaaS).
  • Integration cost: $50–200k per cell first-time integration (cell layout, safety analysis, task definition, fine-tuning data collection).

Customer economics: a humanoid replacing 1.5 FTE at $70k fully-loaded each = $105k/yr labor savings. RaaS at $60k/yr → 1.75 robot-FTE-equivalent net benefit. Payback on a $150k purchase: 1.4 years if running 2 shifts.

The race to scale in 2024–2026: Figure raised $1B+ (Microsoft + OpenAI + NVIDIA), Agility raised $400M, Apptronik raised $350M, Sanctuary raised $140M, 1X raised $240M, Tesla self-funded. Production targets: Tesla aims 10k Optimus/year by 2026, Figure aims 100k by 2029. Whether these targets are met depends on actuator cost-down + supply chain + foundation-model maturity + early-deployment safety record.

13. Production case studies

Active deployments through 2025–2026:

  • Figure + BMW Spartanburg (since 2024): Figure 02 on the X-series assembly line for sub-assembly transfer. Helix policy + voice integration.
  • Apptronik + Mercedes-Benz Berlin-Marienfelde (since 2024): Apollo for parts kitting and tool transport.
  • Agility Digit + Amazon Spanaway WA (since 2023): Digit for tote movement in fulfillment center. ~30 units deployed.
  • Agility Digit + GXO (2024+): Digit + Stretch for case handling.
  • DHL + Stretch + Digit: pilot deployments across DHL Supply Chain.
  • Sanctuary AI + Magna (Spanx textile): Phoenix v9 for fabric-handling cells.
  • Boston Dynamics Atlas Electric + Hyundai (post-2024): Atlas Electric for automotive assembly.
  • Tesla Optimus (internal): production support across Gigafactory Austin + Berlin, ~100s of units, 2024–2026.
  • Kepler PRIME + EngineAI SE01 + UBTECH Walker + MagicLab + LimX: China pilots across BYD, NIO, XPeng, JD.com, Foxconn.
  • Unitree H1/G1/H2: research + early-commercial, sold globally as a development platform.
  • 1X NEO Beta (2024) + NEO Gamma (2025): consumer-home pilots, voice-controlled household tasks.
  • Embodied AI safety incidents: an emerging discipline. Public incident registries are forming (Embodied AI Safety Initiative, similar to autonomous-vehicle SAE J3018 reporting).

14. Adjacent

End walkthrough — design-humanoid-full-stack.md.

Graph View

Backlinks