<- Back to research

ROS 2 tactile sensor pipeline for robot skin data replay

A robotics software guide to ROS 2 tactile data messages, rosbag workflows, force-torque context, and replayable robot skin evaluation.

ROS 2 Kiltedros2_controlforce/torque broadcastersensor data pipeline
Illustration for ROS 2 tactile sensor pipeline for robot skin data replay
RoboSkin.ai explanatory illustration; not a photograph or measurement from the cited experiment.

Updated technical brief - November 3, 2026

Why this source matters

A tactile recording is useful when another engineer can tell what each value means, which transformation produced it and which robot decision consumed it. Successfully publishing a topic answers only part of that question.

Explore tactile sensing for the broader comparison.

This interpretation uses the current public ROS 2 Kilted ros2_control documentation as a concrete interface example. It is not an installation tutorial or a fresh verification of the distribution support calendar. The separate publish-and-replay walkthrough targets a pinned Lyrical starter kit; do not mix its environment with Kilted packages. Its documented execution limits remain separate from this source review.

Core idea

Define the observation before choosing its transport. A pressure field, optical image, six-axis wrench and learned slip label contain different information. Putting them on ROS topics does not make them interchangeable or establish physical units.

The documented force_torque_sensor_broadcaster publishes geometry_msgs/msg/WrenchStamped from force/torque state interfaces. It accepts either a sensor-name convention or explicitly named interfaces, with a declared frame. That is a force/torque interface, not an automatic adapter for an arbitrary tactile array. Kilted broadcaster documentation.

ObservationMeaning to retainInformation that can be lost in a shortcut
Taxel arrayLocation map, channel order, units and validitySpatial contact distribution when reduced to one value
Optical tactile imageSensor identity, acquisition and imaging configurationDeformation evidence when only a classifier label is saved
Force/torque sampleComponents, reference frame and calibrationMeaning of direction and moment when the frame is omitted
Derived eventDetector version, input interval and event definitionEvidence needed to explain a false or late decision

Tactile data is not one data type

An aggregate wrench does not preserve a full contact map. Conversely, an array labeled pressure does not become a calibrated force/torque measurement just because its values can be added. Any conversion needs a documented physical model, geometry and calibration appropriate to the measured quantity.

Preserve the original stream when adding a derived one. Name the transformation and retain its configuration. This makes it possible to distinguish a changed sensor response from a changed processing rule. The calibration guide explains the reference evidence needed before assigning physical units.

Reader value

The Kilted release notes describe per-axis multipliers applied after existing offset logic, sequential filter chains with a separate filtered output, and transformation of wrench messages into target frames. The current broadcaster documentation also describes its TF2-based transformer node. These are documented software facilities; their presence does not validate a particular sensor calibration or its timing. Release notes.

For each output, keep a processing record:

Record fieldQuestion it answers
Input topic/type and sensor identityWhich observation entered the calculation?
Offset, scale and calibration revisionWhich mapping produced the physical estimate?
Filter order and parametersWhich processing changed the values and their temporal behavior?
Source/target frames and transform recordWhere are the components expressed?
Output topic and consumerWhich version of the observation reached the decision?

This table is an editorial record template, not an official tactile message standard. Merely changing a frame name is not a coordinate transformation. Similarly, a numerically plausible filtered signal is not proof that the right reference or transformation was used.

ROS 2 evaluation checklist

  • Record the distribution and exact relevant package versions; a documentation page is not an inventory of the installed system.
  • Distinguish raw, calibrated, filtered, transformed and event topics.
  • Define each timestamp's meaning and clock domain.
  • Keep missing or invalid observations distinguishable from measured zero.
  • Save calibration, geometry and processing configuration with the session.
  • State which observations the controller actually consumes.
  • Verify that a recording includes the required streams and context before evaluating a failure.

Why rosbag and replay matter

A recording can support an investigation only if it retains the evidence the question requires. To examine a late response, keep relevant acquisition, arrival and decision timing with a documented clock relationship. Message order alone does not reconstruct an unrecorded delay.

Separate replay fidelity from physical repeatability. Recovering the recorded values does not recreate the original friction, object pose, actuator state or controller scheduling. An offline replay may help compare processing revisions; a claim about physical recovery still requires the appropriate robot experiment.

The existing ROS 2 walkthrough owns the version-scoped commands and synthetic exercise. This article adds the evidence questions needed when replacing an exercise with a real sensing path, without claiming that those commands were executed here.

What this means for robot skin

Robot skin can change after remounting, cover replacement or repair. Preserve the sensor identity and calibration revision so a later analyst can recognize that change. Otherwise a model may appear to drift when its input assembly actually changed.

The GenForce transferable force sensing interpretation explains why calibration-transfer conditions matter. The Dream-Tac world-action model interpretation distinguishes model timing and prediction evidence from the complete control loop.

What this does not prove yet

Publishing a valid message does not prove complete delivery, correct calibration, deterministic latency, hardware compatibility or safe control. The checks must follow the specific claim. A source review and a successful software build also do not establish a live sensor-to-action result.

Where this fits next

Start with one observed failure and name the evidence needed to distinguish its possible causes. If the question concerns a contact location, retain geometry. If it concerns a late response, retain timing. If it concerns a changed estimate, retain the raw input and processing revision. Use the robot hand tactile sensor route to connect those records to the hand's task contacts.

Practical questions

  • Does one force/torque topic replace distributed touch? It represents a different observation; retain the spatial data needed by the task.
  • Does filtering establish calibration? No. Signal processing and comparison with a physical reference answer different questions.
  • Can replay prove the original robot would recover? No. Replaying observations is not replaying the complete physical interaction.
  • Which guide provides commands? Use the separate Lyrical walkthrough and its explicit verification record; this page reviews interface and evidence choices.

What not to infer

RoboSkin.ai did not install or execute ROS 2, benchmark a controller or validate a hardware driver for this interpretation. Package behavior is attributed to the cited documentation. The proposed records and checks are editorial integration guidance, with no universal rate or safety guarantee.

Sources

ROS 2 Kilted distributions — retained reference; the support calendar could not be reread in this pass.

ros2_control Kilted release notes

Kilted force/torque broadcaster and transformer documentation