Force-Torque Sensing Explained
A practical guide to force torque sensing: sensor types, where to place them, and why contact force data is the highest-value signal in manipulation.
A robot can watch its own hand close around a mug in crisp detail. What it cannot see is the instant the glass begins to crack. Vision goes blind at the moment of contact, exactly when a manipulation task lives or dies. Force is the sense that does not blink. It reports what is happening at the interface, in real time, whether or not a camera can see it.
This is why force-torque sensing keeps showing up in the datasets that produce reliable contact-rich policies, and why it is missing from most of the cheap ones. A camera is easy to add. A well-calibrated force channel is not. That gap is where a lot of robot learning quietly succeeds or fails.
What a force-torque sensor actually measures
A six-axis force-torque sensor measures a full wrench: three forces and three torques. The forces are Fx, Fy, Fz, the push or pull along each axis, in newtons. The torques are Tx, Ty, Tz, the twist about each axis, in newton-meters. Together they describe the complete mechanical interaction between the tool and the world at one point.
Why six numbers and not one? Because contact is directional. Pressing a peg straight down into a hole feels different from pressing while it is slightly tilted, and the difference lives in the side forces and the torques, not the downward push. A policy that only knows total force cannot tell the two apart. A policy that reads the full wrench can feel the tilt and correct it before the peg jams.
Force also unlocks a different way to move. With a wrench in hand, a controller can be compliant. Instead of commanding a rigid position, it commands a target force and lets the arm yield when it meets resistance. That is how a robot slides a drawer shut without knowing its exact position, or seats a connector by feel. None of it works without a clean force reading underneath.
Vision tells a robot where things are. Force tells it what is happening the instant they touch. Only one of those two survives the moment of contact.
Where to put it: wrist, fingertip, or joint
Placement is a design decision with real consequences. The sensor can sit at the wrist, in the fingertips, or be inferred from the joints, and each choice trades coverage against fidelity and fragility.
| Placement | What it measures | Main tradeoff |
|---|---|---|
| Wrist six-axis sensor | Net wrench at the tool | Robust and clean, but cannot say which finger felt what |
| Fingertip tactile | Local contact and pressure map | Rich and local, but fragile and hard to wire |
| Joint torque sensing | Torque at every joint | Whole-arm compliance, coarse for fine manipulation |
| Motor current proxy | Estimated load from current draw | Cheap, no extra hardware, noisy and indirect |
Most manipulation rigs start with a wrist sensor because it is the best ratio of signal to trouble. Fingertip tactile arrays add the localization a wrist sensor lacks, which is what dexterous, in-hand tasks demand, and it is a modality groups like Physical Intelligence treat as central to fine manipulation. Joint-torque and current estimates are the fallback when adding dedicated sensors is not an option.
Why contact data is high-value and scarce
Here is the awkward economics of it. The tasks that matter most in real deployments are contact-rich: insertion, wiping, tool use, assembly. These are exactly the tasks where force carries information vision cannot. Yet the largest robot datasets are built from what is cheap to record, which is video, so force is the channel most often left out.
Datasets that do include it stand out. RH20T records force-torque alongside vision, tactile, and audio, which is part of why it is useful for contact-heavy skills. Model efforts like NVIDIA Isaac GR00T and the large behavior models from Toyota Research Institute care about this modality because a policy that never saw force during training has no way to use it at test time. Work coming out of the Berkeley BAIR group has repeatedly shown that contact feedback changes what a manipulation policy can even attempt.
There is a harder problem underneath. Force is difficult to capture from human demonstration, because you cannot easily measure the forces a person applies without instrumenting their hands. A passive video of someone assembling a part records the motion but not the newtons. This is one of the sharpest limits on learning contact skills from ordinary human video, and it is why instrumented capture that records force directly is worth so much more than its raw hours suggest.
Reading force well: bandwidth, drift, and calibration
Owning a force sensor is not the same as having good force data. Three things separate a usable channel from a decorative one.
Bandwidth. Contact events are fast. The spike when a gripper first touches a surface can be over in a few milliseconds, so a sensor sampled too slowly smears the very transient you wanted to capture. Contact-rich work often wants sampling in the hundreds of hertz or higher.
Drift and bias. Force sensors wander with temperature and load history. Without periodic re-zeroing, a channel that read zero this morning reads a steady false offset by afternoon, and the policy learns a phantom force that is not there.
Gravity and inertia compensation. A wrist sensor feels the weight of the tool and the accelerations of the arm, not just the contact. Unless you subtract the known tool weight and motion, the contact signal you actually want is buried under the robot moving its own hand.
The robot-learning field spent years treating manipulation as a vision problem, because vision was the data that was lying around. The tasks that pay in the physical world are the ones where seeing is not enough. A robot that can feel the tilt of a peg, the give of a clip, or the drag of a rag is a robot working from a sense the camera never had. Capturing that sense well, at the right bandwidth, with the bias removed, is one of the quietest hard problems in building useful robot data.