Every robot on Earth is the same five layers — structure, actuation, sensing, compute, power — wired through a control loop. Pick an archetype, snap the real parts together, and the animation shows the machine come alive. No brand secrets, no copyrighted CAD — just how robots actually work.
~28 DOF. Whole-body MPC at 500 Hz keeps ZMP inside the support polygon.
Aluminium extrusion, CF composite, or 3D-printed nylon. First rule of robot design: pick the stiffest structure you can afford, because every millimetre of flex becomes a control error the software has to fight forever.
Brushed / brushless motors, smart servos, harmonic-geared servos, hydraulic cylinders, or pneumatics. Modern robots default to BLDC + FOC (field-oriented control) because it's backdrivable, silent, and force-controllable.
Proprioception (encoders, IMU, F/T) tells the robot where its body is. Exteroception (cameras, LiDAR, depth, mics) tells it what the world is. You cannot skip either — a robot with no encoders is a rag doll, one with no cameras is blind.
An MCU (STM32 / Teensy) runs the deterministic control loop at 1 kHz — safety-critical, cannot miss a deadline. An SBC (Jetson / Pi / NUC) runs ROS 2, perception, planning and any AI model. They talk over UART, CAN or Ethernet.
Battery Wh divided by average draw = runtime. Motors dominate mobile robots (~70%), compute dominates AI robots (~30-50%). A BMS is mandatory: cell imbalance in Li-ion causes fires, and no CV of a demo is worth that.
sense → estimate → plan → control → act. Every robot runs this loop, differing only in what fills each stage. A cobot's plan is inverse kinematics; a humanoid's plan is a whole-body MPC; a swarm robot's plan is 'match my neighbours'.
A 6-DOF arm can reach any pose in its workspace. A 7-DOF arm avoids singularities (positions where a joint velocity blows up). Bipeds need 6+ DOF per leg; hands need 20+ DOF to grasp the way we do — most cheat with underactuation.
Simultaneous Localization And Mapping fuses LiDAR/camera + IMU + odometry to answer two questions at once: where am I, and what does the world look like? Modern stacks (ORB-SLAM3, LIO-SAM, Cartographer) run in real time on an SBC.
Classical robots run scripts (move here, grip, place). Autonomous robots run policies — functions that map sensors to actions. Modern ones use learned policies: RL, imitation learning, or Vision-Language-Action models like OpenVLA and RT-2.
ISO 10218 (industrial) and ISO/TS 15066 (collaborative) define allowed contact forces. Cat-3 PLe safety controllers cut motor power in <10 ms. E-stops are dual-channel and mechanically latching. If the software can override the safety, it isn't safety.
Between joints: CAN or EtherCAT — deterministic, hard real-time. Between processes on the SBC: ROS 2 DDS pub/sub. To the operator: Wi-Fi 6 / 5G / mesh. Never mix real-time and non-real-time traffic on the same bus.
Pick an archetype (arm, wheeled, quadruped). Buy an open kit: LeRobot SO-100 arm, Unitree Go2 edu, TurtleBot 4. Install ROS 2 Jazzy. Wire ODrive/moteus. Run the demos. Only then write your own node. 90% of robotics is plumbing, not brains.
Open sources only. Content is synthesised from open standards (ROS 2, ISO 10218 / TS 15066, IEC 61508), open-source robotics stacks (Nav2, MoveIt2, ODrive, moteus, Betaflight, LeRobot) under Apache-2.0 / BSD / MIT, public research papers (Cassie, Kilobot, OpenVLA, RT-2) and Wikipedia (CC BY-SA). No proprietary CAD, firmware, or internal documents are used or redistributed.
Trademarks. Unitree, Boston Dynamics, Tesla, TurtleBot and all other product / company names are trademarks of their respective owners, mentioned only as editorial and educational references (nominative fair use). This site is not affiliated with, endorsed by, or sponsored by any of them.