Why Robots Need Force Sensors

A robot arm that can see a bolt perfectly still can't feel how tight it should turn it. Force sensing — a robot's ability to measure the push, pull, and torque it experiences at the point of contact — is what tells a machine how hard it's gripping, pressing, or dragging against something, information a camera simply cannot supply. Without it, a robot is functionally blind the instant it touches anything, and touching things is where almost all real manufacturing, surgery, and human collaboration actually happens.

Vision systems answer the question "where is the object?" Force sensors answer a completely different question: "what is happening at the surface where my gripper meets that object right now?" Both questions matter, but only one of them prevents a robot from crushing an egg, snapping a connector pin, or injuring a person standing next to it on a factory floor. This article explains what force sensors actually measure, how they turn tiny mechanical flexes into usable data, why regulatory standards now require them for robots working near humans, and where the technology still falls short.

What a force sensor actually measures

Most industrial force sensing relies on a six-axis force/torque sensor mounted at the robot's wrist, between the arm and whatever tool or gripper is attached to the end. Internally, these sensors typically use strain gauges — thin metallic foil elements bonded to a deformable metal structure that flexes by a tiny, precise amount whenever a load is applied. That flex changes the electrical resistance of the foil in a predictable way, and onboard electronics convert the resistance change into a reading of force along three perpendicular axes (X, Y, Z) plus torque, or rotational force, around each of those same three axes. That's the "six-axis" part: three linear force measurements and three rotational ones, reported dozens or hundreds of times per second.

The deformable structure itself is usually a small metal spider or cross-beam design, engineered so that any combination of push, pull, twist, or shear produces a distinct, decodable pattern across the array of strain gauges bonded to it. Manufacturers calibrate each sensor against known reference loads before shipping, so the raw resistance changes can be converted into force values in newtons and torque values in newton-meters with a documented accuracy tolerance, typically in the low single-digit percent range for industrial-grade units.

This matters because a robot arm's own joint motor current is a poor substitute. Motor current reflects how hard each joint's motor is working to hold a commanded position, but that electrical signal is muddied by friction in the gearbox, backlash (the small amount of play between meshing gear teeth), and the arm's own inertia as it moves. A dedicated wrist sensor instead reads contact forces directly, right at the point where the tool actually touches the world. That's why precision assembly tasks, such as inserting a connector into a socket or threading a screw into a housing, depend on a purpose-built force/torque sensor rather than trying to back out contact force from motor current alone.

Force feedback instruments on Intuitive Surgical's da Vinci 5 system, which added true haptic force sensing in 2024, were associated with roughly a 20% reduction in average tissue force during procedures compared to earlier systems without it, according to a peer-reviewed analysis published in the *Journal of Robotic Surgery* (2025).

Why compliance, not rigidity, is the goal

A rigid robot arm that simply executes a pre-programmed path will crash destructively into any object that's even slightly out of position, because it has no mechanism to sense a collision until damage has already occurred. Force sensing enables what engineers call compliant control: the robot detects resistance as it happens and adjusts its motion in real time, rather than blindly continuing along a fixed trajectory. It's similar to how a person reaching for a glass in a dark room slows down and softens their grip the instant their fingertips make contact, adjusting continuously rather than shoving forward at full speed and hoping for the best.

Compliance also solves a subtler problem: the real world never matches the idealized model a robot was programmed against. Manufacturing tolerances mean a part might sit 0.3 millimeters off from where a CAD model says it should be. Vibration, thermal expansion, and cumulative wear all nudge things slightly out of place over the course of a shift. A rigid, position-only controller has no way to detect or correct for any of that; a compliant, force-aware controller feels the mismatch as resistance and nudges its path to accommodate it, closing the gap between the planned world and the actual one in milliseconds rather than requiring a human to reprogram the trajectory.

This is also the mechanism that lets collaborative robots, or "cobots," such as Universal Robots' UR series, work directly alongside human operators without the metal safety cages that traditional industrial robots require. Their joint torque sensors, built into each rotating joint rather than just the wrist, continuously monitor for unexpected contact forces and can stop or reverse motion within milliseconds of detecting an unplanned collision. That response speed depends entirely on having continuous force data feeding into the control loop, not just position data, because a controller that only knows where the arm is supposed to be has no way to notice that it has just pinched someone's hand.

How fast does a force-aware stop actually happen?

Collaborative robots built to power-and-force-limiting standards typically detect an unexpected contact and begin decelerating within single-digit milliseconds of the impact, well before a human's own reflexive withdrawal completes. The exact threshold force that triggers a stop is tuned per body region and per task, following the biomechanical limits published in the relevant safety standard rather than a single universal number.

The safety standard that made force sensing mandatory

Force sensing in cobots isn't just a nice engineering feature, it's the backbone of an actual international safety standard. ISO/TS 15066 is a technical specification published to accompany the broader ISO 10218 robot safety standards, and it defines four distinct modes collaborative robots can use to work safely near people. One of those four modes, called power and force limiting, is the one most tabletop cobots rely on, and it depends directly on force sensing to function (ISO/TS 15066 overview, Automate.org / Robotiq).

The standard's force and pressure limits weren't picked arbitrarily. They come from a pain-onset study conducted by researchers at the University of Mainz on behalf of the ISO committee, involving 100 human subjects, designed to establish the maximum force and pressure a robot can apply to 29 distinct areas of the human body before causing pain or injury (Engineering.com, "Standardizing Collaborative Robots"). A robot's fingertip pressing against a forearm has a much lower allowable force threshold than the same robot pressing against, say, the palm of a hand, because different tissues and bone structures tolerate contact differently. Manufacturers program these thresholds directly into the torque sensor's control logic, so the cobot never has to "decide" what's safe in the moment; it's already been configured against a table of biomechanically validated limits.

This is a genuinely useful thing to understand if you're evaluating cobots for a factory floor: a robot marketed as "collaborative" is not automatically safe to work beside without a documented risk assessment. ISO/TS 15066 requires that the specific combination of robot, end-effector tool, and task be evaluated together, because a blunt gripper and a sharp tool changer produce very different injury risks even at identical force thresholds (arXiv risk-assessment review, 2022).

Where force sensing changes outcomes

Not every robotics application needs force feedback in equal measure, but a handful of task categories are essentially impossible to do reliably without it.

Task category	What force sensing does	Typical sensor type	Failure mode without it
Precision assembly (connector insertion, screw threading)	Detects misalignment and adjusts approach angle mid-insertion	Six-axis wrist force/torque sensor	Cross-threading, bent pins, jammed parts
Surface finishing (sanding, polishing, deburring)	Maintains constant pressure against a curved, changing surface	Wrist force/torque sensor with compliant tool holder	Uneven finish, gouged material, unfinished spots
Fragile-object handling (glassware, produce, electronics)	Applies just enough grip force to prevent slipping without crushing	Tactile pressure arrays in the gripper fingers	Crushed or dropped items, inconsistent grip
Human-robot collaboration (cobots on shared floors)	Detects unexpected contact and halts motion within milliseconds	Joint torque sensors at each axis	Delayed stop, injury risk, safety cage requirement reintroduced
Surgical manipulation (robotic-assisted procedures)	Lets a surgeon feel tissue resistance transmitted through the instrument	Miniaturized strain-gauge or optical force sensors at the instrument tip	Excess tissue damage, reliance on visual cues alone

Force sensing solves a different problem in each row, but the common thread is the same: none of these tasks can be done safely or reliably by tracking position alone.

Each of these fails with vision alone, because a camera can locate an object with excellent precision and still tell you nothing about how much the gripper is currently squeezing it. A vision system watching a robot pick up a wine glass sees a hand closing around a shape; it has no way to know whether that grip is 2 newtons or 20 newtons, and the difference between those two numbers is the difference between a secure hold and a shattered glass.

Why doesn't a good camera system make this unnecessary?

Cameras measure geometry, not contact. Even sub-millimeter-accurate vision cannot detect how much force a gripper is applying to a surface, because force is a mechanical quantity generated at the point of contact, not a visual one. Manufacturing tolerances and part variation also mean the real-world position rarely matches a vision system's model exactly, which is precisely the gap force feedback closes.

Surgical robotics: the highest-stakes case for force feedback

Surgical robotics is arguably where the absence of force feedback has been most consequential, and where its recent arrival is best documented. For years, one of the most cited limitations of robotic-assisted minimally invasive surgery was that surgeons operating systems like Intuitive Surgical's da Vinci had no direct sense of how much force their instruments were applying to tissue. Surgeons compensated by training themselves to read visual cues, watching how tissue deforms and springs back under an instrument, but static forces and forces applied to tissue already under tension are notoriously difficult to judge by sight alone (PMC review of haptic feedback in robotic surgery).

Earlier research identified specific engineering obstacles to solving this: friction along the long, thin surgical instrument shaft distorts any force signal traveling from the tip back to a sensor, and reactionary forces from the patient's abdominal wall (in laparoscopic procedures) add noise that's hard to separate from genuine tissue contact (The Robot Report, "Challenges of building haptic feedback for surgical robots"). The da Vinci 5 system, which Intuitive Surgical launched in 2024, addressed this directly by incorporating force feedback (FFB) instruments that measure contact force at the tip and relay it back to the surgeon. A one-year peer-reviewed analysis of thoracic surgery cases using that system found force feedback instrument use at medium or high settings was associated with surgeons applying about 20% less average peak force compared to procedures without it (Journal of Robotic Surgery, 2025).

Separate controlled research on force reflection, feeding measured tissue force back to the operator through resistance in the controls rather than just a visual display, found it produced a statistically significant improvement in both palpation tasks (better detection of hidden tissue inclusions, such as tumors, by feel) and incision tasks (measurably less unintended tissue damage) (ScienceDirect, "Effects of visual force feedback on robot-assisted surgical task performance"). This is one of the clearest cases in robotics where force sensing isn't a convenience feature, it directly changes clinical outcomes.

What people get wrong about robot touch

A common assumption among people new to robotics is that a robot with high-quality cameras and accurate positioning doesn't need force sensing at all, since it already "knows" where everything is supposed to be. That assumption breaks down the moment a robot encounters the real world instead of a simulation. Manufacturing tolerances, part-to-part variation, thermal drift, and vibration all mean the actual position of an object rarely matches the planned position exactly, and force feedback is what closes that gap moment to moment, rather than relying on a perfect, static model that's already slightly wrong by the time the robot arrives.

Another misconception is that force sensors matter only for delicate, fragile tasks like handling glass or produce. In practice, they're just as important for tasks that require firm, consistent pressure. Polishing a car body panel is a good example: too little pressure leaves visible swirl marks and an unfinished surface, too much pressure burns through the clear coat, and the only way to hold a consistent middle ground across a curved, changing surface is to continuously measure and correct force in real time rather than following a fixed pressure setpoint.

A third misconception, more common among people evaluating cobots for a factory purchase, is that any robot labeled "collaborative" is automatically safe to place next to a human without further review. ISO/TS 15066 explicitly requires a documented, task-specific risk assessment covering the robot, its end-effector, and the specific job it's doing, because the same force threshold that's safe with a rounded plastic gripper can become dangerous with a sharp tool attached (arXiv, 2022). Force sensing makes safe collaboration possible; it doesn't make every configuration automatically safe by default.

Where force sensing still falls short

None of this means force sensing is a solved problem. Strain-gauge sensors add cost, typically several hundred to a few thousand dollars per unit for industrial-grade six-axis sensors, which is a meaningful line item on smaller automation projects. They also add a small amount of latency and require periodic recalibration, since mechanical wear and thermal cycling can shift a sensor's zero point over time.

Oversensitive force thresholds create their own problem: a cobot tuned too conservatively will stop or hesitate during completely normal, safe contact, such as a part settling into a fixture, which slows down production and frustrates operators. Tuning sensitivity to the specific task, rather than maximizing precision universally, is a real engineering trade-off, not a solved default. And in surgical robotics specifically, even the improved da Vinci 5 system's force feedback still travels through a long instrument shaft subject to friction, meaning the force a surgeon perceives is an estimate of tissue force, not a direct, friction-free measurement of it.

The takeaway

Force sensors give a robot a sense that vision cannot replace: the ability to feel resistance at the point of contact and adjust in the moment, rather than executing a fixed plan and hoping the real world matches it. Strain-gauge-based wrist sensors convert tiny mechanical deformations into precise force and torque readings, joint torque sensors let cobots meet documented safety thresholds like those in ISO/TS 15066, and even surgical robotics has only recently closed the loop with force feedback that measurably reduces tissue damage. None of that data lets a robot arm behave less like a machine executing a fixed script and more like a hand that actually reacts to what it touches.

---

Read next:

• Why Chiplets Are Changing Processor Design
• How Progressive Web Apps Work
• Why Thermal Throttling Slows Devices

FAQ

Q: What is a force/torque sensor made of?

Most wrist-mounted force/torque sensors use strain gauges, thin metallic foil elements bonded to a deformable metal structure, that change electrical resistance as the structure flexes under load. Onboard electronics convert that resistance change into force and torque readings across six axes: three linear and three rotational.

Q: Can a robot feel force without a dedicated sensor?

Robots can estimate force roughly from joint motor current, but that signal is distorted by friction, gearbox backlash, and inertia, so it lacks the precision needed for tasks like connector insertion, delicate gripping, or surface finishing where contact force must be controlled within a narrow range.

Q: Why do collaborative robots need force sensing specifically?

Cobots like Universal Robots' UR series use joint torque sensing to detect unexpected contact and stop motion within milliseconds, satisfying the power-and-force-limiting mode defined in the ISO/TS 15066 safety standard. That standard's force thresholds come from a University of Mainz pain-onset study covering 29 body regions, and force sensing is what lets a cobot enforce those limits in real time.

Q: Does surgical robotics actually use force feedback today?

Yes, though it arrived later than in industrial robotics. Intuitive Surgical's da Vinci 5 system, launched in 2024, added force feedback (FFB) instruments, and a 2025 peer-reviewed analysis found their use was associated with roughly 20% less average tissue force during thoracic procedures compared to instruments without it.

Q: Does more force sensing precision always help?

Not necessarily. Oversensitive force thresholds can cause a robot to stop or hesitate during normal, safe contact, slowing down production, so engineers tune sensitivity to the specific task rather than maximizing precision universally. In surgical applications, friction along the instrument shaft also means perceived force is an estimate, not a perfectly direct measurement.

Q: How do tactile sensor arrays differ from a wrist-mounted force/torque sensor?

A wrist-mounted six-axis sensor reports one combined force and torque reading for the entire tool or gripper, useful for tasks like insertion or surface finishing where the contact point is effectively a single interface. A tactile sensor array instead spreads many individual pressure-sensing elements across a fingertip or an entire gripper surface, so it can report a full pressure map showing exactly where contact is occurring and how force is distributed across it, not just a single aggregate number. Systems like the BioTac fingertip sensor, roughly the size of a human fingertip, combine that spatial pressure map with additional channels for micro-vibration and temperature, letting a robotic or prosthetic hand distinguish a slipping grip from a stable one and adjust before an object actually falls, a distinction a single wrist sensor further up the arm cannot make on its own (FUTEK, "Robotic Tactile Sensing"). This spatial resolution is precisely why fingertip-level tactile arrays, rather than wrist sensors alone, have become the focus of research into prosthetic hands that need to replicate the fine, distributed sense of touch a human hand relies on for everyday grasping.