← All topics

Guide · Data and hardware · Topic 21 of 32

Embodiments

The bodies and sensors a policy actually controls.

6 minute readBuilds on: Robotics fundamentals

Four robots you will meet

The methods in this guide run on machines like these. The bodies differ, and the recipe is the same: cameras in, action chunks out.

A Unitree G1 humanoid robot standing in an exhibition hall
Unitree G1. A compact humanoid: whole-body control, two arms and hands.
A white Franka robot arm at a workbench
Franka research arm. Seven joints with torque sensing. The standard lab arm, and the arm simulated in LIBERO.
A Flexiv Rizon robot arm polishing a surface while a person stands beside it
Flexiv Rizon. A seven-joint arm built around force control.
A NEURA MAiRA robot arm lit in purple against a dark background
NEURA MAiRA. A collaborative arm with built-in sensing, for work beside people.

Two more bodies

A yellow four-legged Boston Dynamics Spot robot standing on a forest path
Quadruped. Boston Dynamics Spot: 12 driven joints, three per leg.
A white PAL Robotics TIAGo robot with two arms and a head on a wheeled base
Mobile manipulator. PAL Robotics TIAGo: arms and a head on a wheeled base.
RobotTypeDriven jointsPayloadReach or sizeNotes
Franka Research 3Research arm73 kg855 mm reachA torque sensor in every joint
Universal Robots UR5eCollaborative arm65 kg850 mm reachRepeatability ±0.03 mm; force sensing at the tool flange
Boston Dynamics SpotQuadruped1214 kg32.5 kg; 1.6 m/sThree joints per leg; an arm is an option

These figures are from the makers’ published specifications. Two columns deserve a comment. Payload is what the arm can hold at full reach, and it includes the gripper, so a 1 kg hand on a 3 kg arm leaves 2 kg. Repeatability is how closely the arm returns to a position it has visited before. It is not accuracy, which is how close it gets to a commanded position it has never visited, and which is typically several times worse. Learned policies depend on neither as much as classical programs do, because they close the loop through the camera.

What the body decides

The same learning methods run on very different machines. The body sets the action space, the data you can collect and what can go wrong. Degrees of freedom below are typical ranges.

EmbodimentDegrees of freedomStrengthsChallenges for learning
Arms and collaborative robots6 or 7, plus gripperPrecise, mature, safe to work besideA fixed base limits reach and the tasks available
Parallel grippers1Simple, robust, easy to learnCannot re-orient an object in the hand
Dexterous hands12 to 20 or moreHuman-like manipulationHard to teleoperate, fragile, a large action space
Mobile manipulatorsAn arm on a moving baseLarge workspaceBase and arm must be coordinated
Quadrupeds12Stable on rough terrainLittle manipulation unless an arm is added
Humanoids20 to 50 or moreFit human spaces and tools; can learn from human videoBalance, whole-body control, cost and data collection

Most public manipulation data comes from single arms with parallel grippers, so that is where policies are strongest today.

Actuators: what moves the joints

A motor is fast and weak. A joint needs to be slow and strong. The gearbox in between decides how the robot behaves in contact, and it is the least discussed and most consequential hardware choice in robot learning.

A close-up of an assembled strain-wave gear: a ring of fine teeth around a thin flexible cup, with a ball bearing inside
A strain-wave gear, assembled. The fine teeth on the thin inner cup mesh with the outer ring.
The three parts of a strain-wave gear laid out: an elliptical ball bearing, a thin-walled toothed cup and a rigid toothed ring
Its three parts: an elliptical bearing, a thin flexible cup, and a rigid ring with two more teeth than the cup.

The gear in the photographs is the kind found in almost every robot arm. An elliptical bearing deforms a thin toothed cup so that it meshes with a rigid ring at two points. The ring has two more teeth than the cup, so one turn of the bearing advances the cup by two teeth. Ratios of 50 to 160 fit in a flat package with no backlash, which is why arms can be both precise and compact.

A gearbox of ratio nn multiplies the motor’s torque by nn and divides its speed by nn. It also multiplies the motor’s inertia, as felt at the joint, by n2n^2:

τjoint=n η τmotorωjoint=ωmotornIreflected=n2 Irotor\tau_{\text{joint}} = n\, \eta\, \tau_{\text{motor}} \qquad \omega_{\text{joint}} = \frac{\omega_{\text{motor}}}{n} \qquad I_{\text{reflected}} = n^2\, I_{\text{rotor}}
Two log-log plots against gear ratio from 1 to 160. Torque at the joint grows in proportion to the ratio, reaching 100 times at a ratio of 100. Motor inertia felt at the joint grows with the square, reaching 10,000 times. Low-geared designs for legs and hands sit below a ratio of 10, and strain-wave gears for arms between 50 and 160.
The two consequences of gearing. Torque grows in proportion to the ratio and the motor’s felt inertia with its square.

At a ratio of 100 the rotor’s inertia is multiplied by 10,000. The joint then behaves as though a heavy flywheel were attached. It cannot give way quickly when it hits something, and friction in the gearbox hides external forces from the motor, so the motor current says little about contact.

DesignGear ratioStrengthWeaknessWhere
Strain-wave geared50 to 160Compact, precise, strongHigh felt inertia; contact is invisible without a torque sensorArms; add a joint torque sensor to recover force control
Low-geared (quasi-direct drive)6 to 10Gives way on impact; motor current reveals contact forcesLarge, heavy motors for a given torque; runs hot holding a loadLegged robots, some hands
Series elasticHigh, with a spring in seriesThe spring measures force and absorbs shocksThe spring limits how fast force can changeSome humanoids and research arms
HydraulicNoneVery high power for the weightPumps, hoses, leaks, noiseLarge legged machines
Tendon-drivenVariesMotors can sit away from the jointStretch, friction and wearHands

This choice reaches all the way into learning. A stiff, highly geared arm under position control replays a demonstration exactly and is unforgiving in contact, so its policies must be precise. A low-geared or torque-controlled robot is compliant and tolerant of small errors, and it shows forces in its own motor currents, which gives the policy a sense of touch for free. Demonstrations collected on one kind do not transfer cleanly to the other.

Power and compute on board

Everything a mobile robot does runs from a battery. Motors dominate while it moves, and the computer dominates while it thinks: a GPU capable of running a large policy draws tens to a few hundred watts continuously, comparable to the motors of a walking humanoid at rest. That budget is why policies are compressed, why the largest model is often run off the robot, and why the split into slow and fast systems matters.

Heat is the companion limit. Motors that hold a load without moving, as a low-geared leg does when standing, turn their current into heat, and a robot can be limited by motor temperature long before its battery is empty.

Sensing

A policy can only use what its sensors provide, so choosing and placing sensors is part of designing the robot. Cameras are the main input: scene cameras for context, and wrist cameras for a close view of the grasp. Body state is always available. Force, touch and sound matter once contact begins. The note on senses beyond vision covers each of these, what its data looks like, and how a model takes it in.

A useful test for any sensor layout: sit a person at the teleoperation station with only the robot’s camera views. If they cannot do the task, the policy cannot either.

Common mistakes

  • Moving a camera after collecting data. The policy has learned the old viewpoint.
  • Expecting a policy to transfer between bodies for free. A different arm, gripper or camera layout usually needs new data.
  • Leaving out touch because data is scarce. It is under-used because tactile data is hard to collect, not because it is unimportant.

Terms

Degrees of freedom
The number of independent ways a robot can move.
End effector
The tool at the end of the arm, usually a gripper or hand.
Collaborative robot
An arm designed to work safely beside people.
Whole-body control
Coordinating legs, torso and arms together.
Gear ratio
How many turns of the motor make one turn of the joint.
Reflected inertia
The motor’s own inertia as felt at the joint, multiplied by the square of the gear ratio.
Backdrivable
Able to be moved by an external force acting on the joint.
Repeatability
How closely a robot returns to a position it has visited before.

Further reading