The whole robot

Every robot is a chain of rigid parts joined by motorised joints, a set of sensors, a computer that closes the loop between them, and a battery or a cable feeding all three. The shape of that chain, its body plan, decides what the robot can do, what it costs and how it breaks. Learn to read a body plan and you can estimate what a robot costs and where it will fail.

Seven body plans

Here are the seven, with the numbers that matter. Every count below is an actuated degree of freedom, meaning a joint with its own motor.

Two groups sit inside that table. The robot arm family (industrial arms and cobots) stays in one place and moves a tool with great precision. The mobile family (AMRs, tray robots, quadrupeds, humanoids) carries itself to the work. Wheeled robots are cheap, efficient and stopped by a single step. Biped robots climb stairs and fit human spaces, and they pay for it with a dozen leg joints that must balance the body hundreds of times a second. A wheeled humanoid or mobile manipulator splits the difference: human-shaped arms and torso on a wheeled base, which is why Galbot, AgiBot's G2 and Pudu's D7 all chose it for factory and retail work.

Links, joints and degrees of freedom

Strip any robot to its skeleton and you get two kinds of part. A link is a rigid piece: a forearm, a thigh, a chassis. A joint connects two links and lets them move relative to each other in one way. Almost every robot joint is one of two types. A revolute joint rotates, like your elbow. A prismatic joint slides in a straight line, like a drawer or the lifting column in Galbot's torso.

The part at the end of the chain that touches the world is the end effector: a gripper, a welding torch, a suction cup or a hand. Changing it changes the job. An arm with a torch welds; the same arm with a gripper packs boxes.

Workspace, reach, payload and repeatability

Four numbers on every arm's spec sheet decide whether it can do a job. The workspace is the volume the tool can reach, shown as the ring in FIG 01.2 and as a lumpy sphere on a real six-axis arm. Reach is the radius of that sphere. Payload is the mass the tool can carry, and it shrinks as the arm stretches, because the same joint torque has a longer lever to fight. Always ask at what reach and speed the maker measured the payload. Atlas quotes 30 kg sustained and 50 kg peak verified Sep 2026: two very different robots depending on which one your customer needs all shift.

Where a humanoid spends its joints

Humanoid makers quote a total DOF count as if it were horsepower. The total matters less than where the joints sit. Legs need about 12 for walking. Arms need 12 to 14 to reach and orient a hand. A waist and neck add 3 to 5. Hands swallow the rest, and they swallow a lot: on Tesla Optimus the two hands carry about 44 DOF against 28 for the entire rest of the body estimate. Each of those joints drags a motor, a reducer, sensors and a driver with it, which is why hands are about 17% of an Optimus bill of materials Morgan Stanley.

Sense, think, act

Two phrases describe the business around this loop. The robot body is the hardware: frame, joints, sensors, battery. A robot OEM is the firm that builds and sells that body, such as Unitree, UBTech or AgiBot. Embodied AI is the umbrella term for putting learned intelligence into a physical body, and it's the phrase on every pitch deck and government policy paper in China as of Sep 2026. The rest of Part 1 takes each piece of the loop apart: motors and joints, sensors, compute, wiring and power. Part 2 covers the thinking, from control loops to robot foundation models.

Three flows: power, data, force

Follow any robot from battery to fingertip and three things move through it. Power flows from the battery to the drivers, motors and computer. Data flows from sensors to the computer and from the computer to the drivers. Force flows from the motors through gears and links into the world, and the world pushes back. Every design decision trades one flow against another, and most field failures happen where two of them cross a moving joint.

The numbers put the trade-offs in view. Figure 03 carries a 2.3 kWh pack for about 5 hours at peak performance verified. Unitree's H2 carries 0.972 kWh for about 3 hours, and the small G1 about 2 hours on a 9,000 mAh pack verified Sep 2026. Every watt the computer burns is a watt the legs don't get, so a bigger onboard model costs runtime directly.

The crossing points are where robots break. Figure's own write-up of its BMW deployment named the forearm as the top hardware failure point on Figure 02: tight packaging, a three-DOF wrist, thermal limits, and a distribution board with cables flexing through the wrist to feed the motors. Figure 03 removed the board and the moving cables, so each wrist motor controller now talks straight to the main computer verified Nov 2025. Power, data and force all crossed the same moving joint, and that's where it failed.

From teleoperation to autonomy

How much of the work does the robot do by itself? The answer sits on a ladder, and most robots you'll see in 2026 sit lower on it than their videos suggest. Robotics has no agreed standard for these levels as of Sep 2026, unlike the SAE levels for cars, so the five below are this primer's own ladder. Each autonomy level is defined by what the human still does.

Two cautions. First, autonomy is always autonomy at a task. BellaBot is close to L4 at driving and at L0 for loading plates, because it has no arms. Second, lab success rates overstate field success. One industry report cites a drop from about 95% in the lab to about 60% in the field against a 99.9% industrial bar estimate. A robot at L2 still needs a paid human, so the business case depends on how many robots one person can supervise.

Where cost and failure concentrate

Money and breakdowns pile up in the same place: the joints. Actuators are 40 to 60% of a humanoid's bill of materials, about 56% on Optimus estimate. On Unitree's G1 the 23 joint modules cost about CNY 27,500 of a CNY 41,574 base BOM, 66% China Post Securities teardown. The computer, the part investors talk about, is 10 to 15%. That's why the next chapter, Motors and joints, is the longest in Part 1, and why the cost chapter keeps coming back to reducers and screws.

Failures follow the same map. On humanoids the public reports are overheated joint motors, broken small gears in wrists, and battery over-current trips when a new control policy demands a burst of torque. Two controllers on the same G1 differ by 45 to 60 °C at the ankle: Unitree's stock controller keeps ankles at 30 to 45 °C, while one research whole-body controller hit the 90 °C warning within 10 minutes of standing owner report. On wheeled service robots the faults cluster in cameras, sensors, calibration, lubrication, jammed joints and food debris on sensors verified. How robots fail and get fixed has the full picture.

Terms in this chapter