Learning Path · Intermediate

Paxini Gen3: From Sensing to Tactile-Enhanced Robot Learning

In ~8 hours, go from your first sensor reading to a trained robot manipulation policy that uses tactile feedback. You will build a complete grasp quality detection pipeline integrated with robot arm data collection — the same workflow used in contact-rich manipulation research.

Total Time ~8 hours
Difficulty Intermediate
Units 6 (+ Orientation)
Hardware Required Paxini Gen3 + any robot arm
Prerequisites Basic Python, a robot arm
Before you start this path: You should be comfortable writing basic Python scripts (reading data from a device, parsing arrays) and have a robot arm with a Python SDK you can control. You do not need prior robotics research experience. If you are new to robot data collection in general, complete the OpenArm learning path first — it covers data collection fundamentals in more depth.

What You Will Build

By the end of this path you will have:

Live Tactile Heatmap

A streaming pressure visualization running at 500 Hz from your Gen3 sensor, confirming every taxel is functional.

Grasp Quality Detector

An online classifier that distinguishes stable grasps from slip-prone ones in real time during robot operation.

Tactile Dataset (50 demos)

A full LeRobot-format dataset with synchronized tactile + joint + camera channels, quality-checked and ready for training.

Tactile-Aware Policy

A trained ACT or Diffusion Policy model that uses tactile input — evaluated against a vision-only baseline.

Path Overview

Complete units in order. Each unit ends with a concrete completion check so you know when to move on.

0

Orientation: Before You Begin Setup

What tactile sensing adds to robot learning, hardware and software checklist, what you will build, and how to get help if you are stuck.

~30 min · No technical content
1

Install and Verify the Sensor

Install the Paxini SDK, connect the Gen3 over USB-C, write a 5-line streaming script, and run the live heatmap to confirm all taxels are working.

~1 hour
2

Understand Tactile Data

Learn the spatial structure of the pressure array, calibrate contact detection thresholds, implement grasp detection logic, and visualize contact events over time.

~1.5 hours
3

Sync Tactile with Robot Arm

Mount the sensor on your gripper, route USB along the arm, and use MultiSourceSync to record synchronized arm + tactile + camera frames with verified timestamp alignment.

~2 hours
4

Record Tactile Demonstrations

Use the extended LeRobot dataset format to record 50 grasp-and-place demonstrations with tactile channels. Apply the quality checklist and flag slip events.

~2 hours
5

Train a Tactile-Aware Policy

Add tactile as an observation modality in ACT or Diffusion Policy, train on your dataset, and evaluate against a vision-only baseline on deformable and transparent objects.

~1.5 hours (+ training time)
Start the Path — Unit 0 →

Or jump directly to Unit 1 if you have already done orientation.