Force Torque Sensing for Robots: Why It Matters and How to Use It
Vision tells a robot where things are. Force-torque sensing tells it how hard it is pushing, pulling, and squeezing. For contact-rich manipulation tasks, F/T sensing is not optional — it is what separates robots that can handle delicate assembly from robots that break everything they touch.
What Force-Torque Sensing Is
A force-torque (F/T) sensor measures six components of mechanical load at a point on the robot: three forces (Fx, Fy, Fz) along each axis and three torques (Tx, Ty, Tz) around each axis. Mounted at the robot's wrist — between the last arm link and the end-effector — an F/T sensor provides a complete picture of the mechanical interaction between the gripper and the environment: how much force the robot is applying, in what direction, and how much bending moment is being generated.
This information is invisible to cameras and cannot be reliably inferred from joint motor currents alone (though some modern arms, like the Franka, provide wrist-level joint torque estimates that give a coarse F/T-like signal). A dedicated F/T sensor at the wrist provides calibrated, high-bandwidth force data that can be used for compliant control, contact detection, grasp quality monitoring, and as an additional observation channel for learned manipulation policies.
When You Need F/T Sensing
F/T sensing is most valuable for contact-rich manipulation tasks: peg-in-hole insertion, mating connectors, opening jars, threading screws, sliding drawers, polishing surfaces, and assembly tasks where the robot must apply controlled forces rather than just following a position trajectory. In these tasks, small positioning errors cause large contact forces that a position-controlled robot cannot compensate for without force feedback — the robot either misses the target entirely or applies excessive force and damages the assembly.
For pure pick-and-place tasks with generous tolerances, dedicated F/T sensing is often unnecessary. Joint torque monitoring is sufficient for basic collision detection, and well-calibrated grasping with appropriate gripper stiffness handles most object handling without explicit force control. Add dedicated F/T sensing when your task involves precise insertion, contact-maintaining motions, or force-controlled surface following.
Hardware Options
The ATI Mini45 and Mini85 are the gold standard for research F/T sensing: high accuracy, high bandwidth (7 kHz), and extensive documentation. They are expensive ($3,000–$7,000 depending on measurement range and connector) but reliable and well-supported. The Robotiq FT 300-S is a more affordable option ($2,000–$3,000) with lower bandwidth but easier integration via USB and Ethernet and direct ROS support. The OnRobot Hex series integrates F/T sensing directly into the gripper-arm interface, simplifying wiring at the cost of less configurability.
For budget-constrained research, newer low-cost F/T sensors from companies like Bota Systems and Wacoh Tech have improved significantly and are now suitable for many research applications. When selecting a sensor, match the measurement range to your task: an overly sensitive sensor used for heavy payload grasping will saturate; an insensitive sensor used for fine assembly won't detect the contact events that matter.
Integration with Manipulation Policies
F/T data can be integrated into learned manipulation policies in two ways. The most common approach is to include the F/T readings as additional observation channels alongside images and joint states — the policy receives a 6-dimensional F/T vector at each timestep and learns to use it implicitly. This is straightforward to implement and works well when F/T data provides a useful signal that the policy can learn to exploit.
A more structured approach uses F/T data to parameterize a classical compliant control layer: the learned policy outputs a desired wrench or compliance parameter rather than a joint position, and a classical impedance controller translates this into joint commands while maintaining the specified compliance. This hybrid architecture separates the task planning (learned) from the contact dynamics handling (classical), often producing more reliable behavior on contact-rich tasks than pure end-to-end learning.
Contact-Rich Tasks and SVRC's Sensing Stack
SVRC has deployed F/T-equipped robot stations for data collection on contact-rich manipulation tasks including plug insertion, drawer opening, and assembly tasks requiring controlled force application. Our standard sensing stack includes an ATI Mini45 at the wrist of each robot arm, with F/T data logged synchronously alongside cameras and joint states at 500 Hz. This logging rate is sufficient to capture contact transients and is included in standard dataset exports.
For teams doing research on contact-rich manipulation, F/T-equipped stations at SVRC's Palo Alto facility are available through our robot leasing program. We also provide consultation on selecting and integrating F/T sensors for custom hardware setups. Contact the SVRC engineering team to discuss your specific contact-rich manipulation requirements, or browse our hardware catalog for available sensor configurations.