Learning paths

Learn tactile sensing: from robot skin to working data

Submit source ->

Choose a starting point based on what you need to do: understand contact sensing, shortlist hardware, or process sensor data. Each path leads to practical guides and their primary sources.

Technical workbench showing RoboSkin.ai guide routes connected to robot skin, tactile AI, and research context.
Guide-route visual for RoboSkin.ai research and category routes.

Learning path

Understand what the robot needs to sense

Start with the difference between a sensing element and an integrated robot surface. Identify the contact location, measured signal, and action the robot should take.

Start here if: Readers choosing between contact detection, a tactile fingertip, and a distributed skin

Inputs and outputs: A task description becomes a list of required contact signals and coverage areas.

Decision note

A reported sensing capability must be checked for the specific sensor and mounting arrangement.

  • Contact events
  • Pressure and force measurements
  • Surface coverage
  • Robot skin and e-skin terminology

Evaluation criteria

  • Where can contact happen?
  • Is a binary event enough?
  • Does the task need contact location or a spatial map?
  • What response will use the signal?

Learning path

Compare sensor hardware and evidence

Compare optical, magnetic, resistive, and other sensing systems using their documented outputs, geometry, calibration requirements, and task evidence.

Start here if: Engineers shortlisting tactile sensors for hands, grippers, or soft surfaces

Inputs and outputs: A sensing requirement becomes a shortlist with documented constraints and primary sources.

Decision note

The sensor directory links to research and manufacturer documentation; it is not a stock or purchasing catalog.

  • DIGIT and GelSight Mini
  • ReSkin magnetic sensing
  • Raw data versus derived estimates
  • Mounting and calibration

Evaluation criteria

  • What does the hardware measure directly?
  • Which values require a learned model?
  • Are code and calibration procedures available?
  • Was the claimed result measured on a comparable task?

Learning path

Build a tactile data and evaluation workflow

Follow the Python and ROS 2 exercises to inspect samples, preserve timestamps and units, replay observations, and define an evaluation before connecting a controller.

Start here if: Developers moving from a sensor stream to recorded data and robot feedback

Inputs and outputs: Recorded or synthetic observations become plots, replayable messages, and an explicit validation plan.

Decision note

Tutorial exercises use documented examples and synthetic data; running them does not establish hardware performance.

  • Python CSV inspection
  • ROS 2 messages and replay
  • Calibration records
  • Task-level evaluation

Evaluation criteria

  • Are time, units, and coordinate frames explicit?
  • Can invalid samples be identified?
  • Can the sequence be replayed?
  • Does evaluation measure object outcomes as well as prediction accuracy?

What should you have at the end of each path?

Use these checkpoints to decide when to move on. They are learning outputs, not a certificate of hardware readiness.

  1. Contact requirements

    A sketch of likely contact areas and a statement of the signal needed: contact event, location, normal load or shear.

    Identify the action that will use the signal, including what happens when it is missing.

    Connect the requirement to a task

  2. Hardware shortlist

    A source-linked record of the exact model, mounting surface, raw output, derived estimates and unverified specifications.

    Compare the same measurement target and configuration; a model-family claim is not a specification for every module.

    Build a shortlist

  3. Data workflow

    A small trace with timestamps, units, coordinate frames, calibration version and invalid-sample handling.

    Replay the trace and inspect gaps before proposing a controller experiment. Synthetic replay validates processing assumptions, not physical performance.

    Inspect and replay data

If one checkpoint exposes a gap, return to that requirement instead of treating all three paths as a mandatory course. For example, readable sensor samples do not by themselves establish force accuracy.

Reading guidance updated .