Robotic Tactile Sensing Measurement Chain Record

Robotic Tactile Sensing Measurement Chain Record

Bounded definition

Robotic tactile sensing is the measurement of physical contact so that a machine can estimate a contact property or use it in a decision. It is not a machine equivalent of subjective human feeling.

Vision-based tactile sensors place a camera behind or beside a compliant sensing surface. Contact deforms that surface. Internal illumination, markers, coatings, or visible texture make the deformation observable. Software then turns the recorded frames into a measurement or representation.

The DIGIT paper describes a compact, high-resolution sensor designed for in-hand manipulation. The Sparsh paper evaluates DIGIT, GelSight 2017, and GelSight Mini data. These sensors share a broad optical principle but differ in geometry, markings, image characteristics, and operating details.

Measurement chain

Terms that must stay distinct

Normal force acts perpendicular to the contact surface. Shear force acts along the surface. Slip is relative motion at the contact interface and may be inferred from temporal or force evidence. Pose estimation concerns the position or orientation of an object relative to a sensor or reference frame. Texture recognition is a classification problem, not force measurement.

A tactile image is not automatically a calibrated force map. A representation is not automatically a task output. A task output is not automatically safe control evidence.

Time and state

Some contact properties require a sequence rather than one frame. Sparsh concatenates frames separated by five samples for several image-based variants and uses four-frame clips for V-JEPA. At 60 frames per second, the paper describes windows of roughly 80 to 100 milliseconds.

The full system latency also includes camera capture, transport, preprocessing, encoder inference, task decoding, controller scheduling, actuator response, and queueing. A reported backbone frame rate is not end-to-end reaction time.

Practical boundary

A useful tactile system needs a defined physical property, an appropriate sensor, calibration evidence, a task-specific estimator, representative testing, and a controlled response to uncertainty or failure.

The same sensor can be strong for one contact geometry and weak for another. The same model can transfer across a reported benchmark and still fail after a gel replacement, lighting change, new object material, or new robot motion.