What Tactile Sensing Adds to Robot Learning
Most robot learning datasets contain three modalities: joint positions (proprioception), camera images (vision), and sometimes wrist force/torque. None of these tell the policy whether the gripper has a stable grasp on the current object.
Tactile sensing fills this gap. A fingertip-mounted sensor measures the contact pressure distribution at the gripper surface — the spatial pattern of forces that determines whether a grasp will hold or slip. This provides direct contact state supervision that transforms how imitation learning policies handle contact-rich tasks:
Hardware Checklist
Before starting Unit 1, verify you have every item below in hand.
- Paxini Gen3 Sensor Module — fingertip, finger-pad, or palm variant depending on your gripper. If you have not ordered yet, see the Paxini Gen3 hub for ordering and variant selection guidance.
- USB-C cable (included in box) — the short 0.5 m cable that ships with the sensor. For arm integration, you will also need a USB hub — any powered USB 3.0 hub works.
- Mounting adapter for your gripper — included in the box if you specified your gripper at order time. For custom grippers, use the adhesive mount and M2 screws (included). Allow adhesive 30 minutes to cure before applying load.
- Robot arm with Python SDK — any compatible arm (OpenArm, Franka, UR series, or similar). You need a working Python interface that returns joint positions with timestamps. Unit 3 covers integration specifics.
- Wrist-mounted camera (for Unit 4+) — any USB or GigE camera. 30 fps minimum. Already attached to your arm setup is fine.
Software Checklist
Confirm these items before Unit 1.
- Python 3.10 or higher — run
python3 --version. The Paxini SDK requires 3.10+. - pip and venv — run
pip3 --version. Create a fresh virtual environment for this path:python3 -m venv paxini-env && source paxini-env/bin/activate. - ~5 GB free disk space — for the SDK, dataset recordings, and model checkpoints. A 50-demo tactile dataset is approximately 1.5 GB.
- Your robot arm SDK installed and working — confirm you can read joint positions from Python before Unit 3. This path does not cover arm SDK setup.
- GPU (for Unit 5) — an NVIDIA GPU with 8GB+ VRAM reduces ACT training from ~3 hours to ~45 minutes. Training on CPU works but is slow. Cloud GPU options are covered in Unit 5.
Time Estimates
| Unit | What You Do | Time |
|---|---|---|
| 0 | This orientation | ~30 min |
| 1 | SDK install, first reading, heatmap verify | ~1 h |
| 2 | Data structure, calibration, grasp detection | ~1.5 h |
| 3 | Mount sensor, sync with arm, verify timestamps | ~2 h |
| 4 | Record 50 demo dataset with tactile | ~2 h |
| 5 | Train tactile policy, evaluate vs baseline | ~1.5 h |
| Total | ~8 h 30 min | |
Units 1–2 can be done in one session without your robot arm (sensor only). Units 3–4 require the full hardware setup and work best in a single session each. Unit 5 training can run unattended.
How to Get Help
- Check the completion check at the bottom of whatever unit you are in — it specifies exactly what you should see when everything is working.
- Read the troubleshooting section in the software guide — it covers the top 3 issues with exact fixes.
- Post in the forum questions thread with your OS, SDK version (
paxini.__version__), and the full error output.
Orientation Complete When...
You have checked every item in both checklists, you have your Paxini Gen3 sensor in hand with the correct mounting adapter, and you have set aside at least one uninterrupted hour for Unit 1.