Definition

A force-torque (F/T) sensor measures the three linear forces (Fx, Fy, Fz) and three torques (Tx, Ty, Tz) acting on it, providing a complete 6-axis picture of the contact wrench. Mounted at the robot's wrist (between the last joint and the end-effector), an F/T sensor tells the controller exactly how much force the gripper is exerting on an object and in which direction.

This information is critical for tasks where vision alone is insufficient: inserting a peg into a tight hole requires sensing contact forces to avoid jamming; polishing a surface requires maintaining consistent normal force; and assembly operations require detecting when parts are seated correctly. Without force sensing, robots must rely on position control alone, which leads to either excessive force (damaging parts) or insufficient contact (dropping parts).

In the context of robot learning, F/T data provides an additional observation modality alongside camera images and joint positions. Policies trained with force observations can learn contact-rich behaviors — such as wiping, screwing, and inserting — that are extremely difficult to learn from vision alone. During teleoperation for data collection, recording F/T data alongside visual observations creates richer training datasets for imitation learning.

How F/T Sensors Work

The two dominant sensing technologies are strain gauges and capacitive elements. Strain-gauge sensors use a precisely machined metal body (typically a cruciform or Stewart platform geometry) instrumented with bonded strain gauges. When external forces are applied, the body deforms microscopically, changing the resistance of the strain gauges. A Wheatstone bridge circuit converts these resistance changes into voltage signals proportional to the applied force and torque.

Capacitive sensors use arrays of capacitor plates separated by a compressible dielectric. Applied forces change the gap between plates, altering capacitance. These sensors can be made smaller and lighter than strain-gauge designs, making them suitable for fingertip sensing on dexterous hands.

Both types output analog signals that are digitized by an onboard ADC and transmitted to the robot controller via EtherCAT, USB, or Ethernet at 1–8 kHz. The raw signal is typically filtered (low-pass, 50–500 Hz cutoff) to remove vibration noise. Gravity compensation — subtracting the weight of the end-effector and any held object — is applied in software to isolate contact forces from gravitational loads.

Top Sensors

  • ATI Industrial Automation (Gamma, Mini45, Nano17) — The industry standard for research and industrial robotics. Strain-gauge based, extremely accurate (0.1% full-scale), wide range of models from fingertip-sized (Nano17: 12 mm diameter) to heavy-duty (Omega: 6000 N range). Price: $3K–$15K.
  • Robotiq FT300 — Designed for collaborative robots (UR, Fanuc). Built-in ROS driver, easy mounting on UR e-series. Range: 300 N force, 30 Nm torque. Competitive price (~$3K) and good for most research manipulation tasks.
  • OnRobot HEX-E/H — 6-axis sensor integrated with OnRobot's tool changer ecosystem. Automatic tool recognition and gravity compensation. Designed for quick industrial deployment rather than research flexibility.
  • XELA uSkin / BioTac — Tactile sensing arrays that provide distributed force measurements across a surface, rather than a single 6-axis wrench. Used on dexterous hands and fingertips. Enable slip detection and texture recognition.

Integration with Robot Control

Admittance control: The F/T sensor measures external forces, and the controller adjusts the position trajectory to accommodate them. The robot moves as if connected to the environment through a virtual spring-damper. This is the most common integration pattern for position-controlled robots (UR, Fanuc, ABB).

Impedance control: The controller directly regulates the relationship between motion and force using the F/T signal. The robot behaves like a programmable mass-spring-damper system. This requires torque-controlled actuators (Franka Panda, KUKA iiwa) and provides more responsive compliance than admittance control.

ROS2 integration: Most F/T sensors publish to the geometry_msgs/WrenchStamped topic. ROS2 control frameworks (ros2_control, MoveIt Servo) can consume this data for force-limited trajectories, contact detection, and hybrid force-position control. The force_torque_sensor_broadcaster package provides a standard interface.

Key Specifications

Range: Maximum measurable force and torque before saturation. Must exceed the expected task forces with margin. Typical ranges: 50–500 N for manipulation research, 1000–6000 N for industrial applications.

Resolution: Smallest detectable force change. ATI Nano17 achieves 1/160 N resolution. Higher resolution enables detection of subtle contact events like surface texture changes or incipient slip.

Overload protection: How much force the sensor can withstand without damage (typically 5–20x the rated range). Critical for research settings where crashes happen. Some sensors include mechanical stops to prevent overload.

Sample rate: Output frequency. 1 kHz is sufficient for most manipulation. High-speed assembly or impact detection may require 4–8 kHz.

Calibration and Signal Processing

Factory calibration: F/T sensors ship with a calibration matrix that converts raw voltage readings to force/torque values. This matrix compensates for cross-axis coupling (force on the X axis producing a spurious signal on the Y axis). Factory calibration is accurate to 0.5–2% of full-scale range. For high-precision applications, in-situ recalibration using known weights and moment arms improves accuracy to 0.1%.

Gravity compensation: The weight of the end-effector and any held objects produces a static wrench that must be subtracted from the raw signal to isolate contact forces. This requires knowing the end-effector mass and center of gravity, updated whenever the tooling changes. Most ROS2 F/T drivers include a gravity compensation node that uses the robot's current orientation (from joint encoders and forward kinematics) to compute and subtract the gravitational wrench in real time.

Noise filtering: Raw F/T signals contain high-frequency vibration from motors, gear meshing, and environmental disturbances. A low-pass filter (typically 10–100 Hz cutoff, Butterworth 2nd-order) removes vibration noise while preserving contact event dynamics. For impact detection, a higher cutoff (200–500 Hz) preserves fast transients. Digital FIR filters are preferred over IIR in safety-critical applications for their linear phase response.

Drift and temperature sensitivity: Strain-gauge sensors exhibit thermal drift (0.01–0.05% of full-scale per degree C). For long-running applications, periodic re-zeroing (biasing) eliminates accumulated drift. Some sensors include onboard temperature compensation; others require software correction using a co-located temperature sensor.

Applications in Robot Learning

Including F/T observations in the training data for imitation learning policies provides information that cameras cannot capture — contact magnitude, direction, and timing. Research has shown that adding wrist F/T to visual observations improves success rates by 15–40% on contact-rich tasks such as:

  • Insertion tasks — Peg-in-hole, USB plugging, connector mating. F/T data reveals jamming, misalignment, and successful seating that are invisible to cameras positioned away from the contact zone.
  • Wiping and polishing — Maintaining consistent contact force while following a surface requires force feedback. Policies trained with F/T observations learn to modulate pressing force based on surface compliance.
  • Assembly — Detecting when parts snap together, recognizing thread engagement during screwing, and sensing proper seating of clips all produce distinctive force signatures that policies can learn to recognize.

At SVRC, our teleoperation rigs record synchronized F/T data alongside camera images and joint states at 50 Hz, producing multi-modal datasets ready for force-aware policy training with ACT or Diffusion Policy.

See Also

Key Papers

  • Hogan, N. (1985). "Impedance Control: An Approach to Manipulation." ASME Journal of Dynamic Systems. The foundational paper on using force sensing for compliant robot control, introducing the impedance control framework.
  • Luo, J. et al. (2024). "Multi-Modal Imitation Learning with Force Observations." CoRL 2024. Demonstrated that adding F/T observations to visual imitation learning significantly improves success on contact-rich manipulation tasks.
  • Lee, M. A. et al. (2019). "Making Sense of Vision and Touch: Self-Supervised Learning of Multimodal Representations for Contact-Rich Tasks." ICRA 2019. Showed how to combine vision and tactile/force signals for robust manipulation policies.

Related Terms

  • Impedance Control — Control strategy that uses F/T feedback for compliant behavior
  • End-Effector — The device where F/T sensors are typically mounted
  • Teleoperation — F/T data enables haptic feedback and force recording during data collection
  • Grasp Planning — Force analysis determines grasp stability
  • Policy Learning — F/T observations as input to force-aware manipulation policies

Add Force Sensing to Your Robot

Silicon Valley Robotics Center provides F/T sensor integration services for common research and industrial robots. We can help you select the right sensor, mount and calibrate it, configure ROS2 drivers, and set up impedance or admittance control for your contact-rich manipulation tasks.

Explore Data Services   Contact Us