Skip to main content
Robotics & Embodied AI

⏱ About 15 min15 XP

Touch, Force, and Proprioception

Pick up a raw egg. You know instantly whether you are gripping it firmly enough to hold it without dropping it, and gently enough not to crack it. That feedback loop between your fingers and your brain, happening entirely without visual input, is one of the most sophisticated sensory systems in biology. Replicating it in a robot is one of the hardest open problems in embodied AI.

Touch Sensors: Detecting Contact

A touch sensor detects whether physical contact is occurring and often measures how hard the contact is. The simplest version is a binary switch: either touching (1) or not touching (0). More sophisticated tactile sensors — named by analogy with human skin — use arrays of tiny pressure-measuring elements called taxels (tactile pixels). A tactile array can tell not just that something is being touched but where on the sensor surface the contact is, and how the pressure is distributed across that area. Researchers at MIT, Stanford, and robotics companies have developed tactile skins — flexible sheets of sensor material that can be wrapped around a robot finger or palm. These skins measure thousands of taxels simultaneously, giving the robot a rich spatial map of exactly which part of its hand is in contact with an object and with what force.

Taxel

A taxel is a single pressure-measuring element in a tactile array — the touch equivalent of a pixel in an image. An array of taxels produces a 2D pressure map of a contacted surface.

Force-Torque Sensing

A force-torque sensor measures the full set of mechanical loads acting on a robot joint or end-effector — the tool at the end of a robot arm. Force has three components (push/pull in each of three directions), and torque has three components (rotation around each of three axes). Together these six measurements, called a wrench, tell the robot exactly how hard and in what direction it is being pushed or twisted. Force-torque sensors are small devices typically mounted between the robot's wrist and its gripper. They allow the robot to perform sensitive tasks like inserting a peg into a hole with tight tolerances, screwing in a bolt, or polishing a surface with a precise amount of pressure. Without force feedback, a stiff industrial robot would either not push hard enough to complete the task or push so hard it breaks the part.

Proprioception: Knowing Your Own Body

Proprioception — from the Latin proprius, meaning 'one's own' — is the sense of the position and movement of your own body parts. Close your eyes and touch your nose. You can do this precisely because proprioceptors throughout your muscles and joints tell your brain exactly where your arm is at every moment. In robots, proprioception is typically achieved through encoders — sensors attached to each motor that count exactly how many degrees the motor shaft has rotated. By knowing the starting position and adding up every small rotation, the robot's computer always knows the current angle of every joint. For a robot arm with six joints, six encoder readings combine through a mathematical process called forward kinematics to determine the precise 3D position of the gripper.

Encoders are highly accurate under normal conditions, but they can accumulate error over time — a problem called drift. If a joint slips mechanically, or if the robot is powered off and moved, the encoder count no longer reflects reality. For this reason, many robots periodically calibrate by moving to known reference positions. Legged robots and humanoids face extra challenges because their feet contact uncertain terrain, making proprioception alone insufficient — they combine it with inertial measurements.

Match each sensing concept to its unique function in a robot body.

Terms

Tactile skin
Force-torque sensor
Encoder
Forward kinematics
Taxel

Definitions

Computing the 3D position of the gripper from all the joint angles
A single pressure-sensing element within a larger tactile array
Measuring the full six-component mechanical load at a robot's wrist
Counting motor shaft rotations to track the precise angle of each joint
Providing a spatial map of contact pressure across a robot's finger or palm surface

Drag terms onto their definitions, or click a term then click a definition to match.

Why These Senses Matter for Manipulation

Vision is great for understanding a scene before contact, but once a robot's gripper closes around an object, vision is often blocked. At that point, touch and force become the primary feedback channels. A robot sorting fragile medical samples, assembling delicate electronics, or handling produce must rely on tactile and force sensing to avoid crushing what it holds. The combination of proprioception (knowing where the arm is) and touch and force sensing (knowing what the arm is feeling) enables compliant manipulation — the ability to yield and adapt in response to unexpected forces. A compliant robot arm that senses resistance will stop pushing; a rigid arm without force feedback may snap a part or injure a nearby person.

What does proprioception allow a robot to do?

Why is force-torque sensing important for assembly tasks like inserting a peg into a hole?

A tactile array is made of individual elements that measure pressure across a surface. Knowing the position of your own body parts is called . In robots, joint count motor shaft rotations to provide this sense.

Blind Grasping Challenge

  1. Step 1: Gather five objects of different sizes, textures, and weights — a pencil, a crumpled paper ball, a smooth stone, a rubber eraser, and a small water bottle.
  2. Step 2: Close your eyes and have a partner hand you each object one at a time. Without looking, describe: Is it soft or rigid? Smooth or rough? Light or heavy? Where exactly are your fingers touching it?
  3. Step 3: Write down which sensory cues (pressure distribution, weight, texture) told you the most about each object.
  4. Step 4: Now imagine you are a robot with only binary touch sensors on each fingertip (on/off). Which of the cues you just used would be lost? Which would remain?
  5. Step 5: Propose a sensor design improvement that would restore the most useful lost cue.