# Teaching Robots to Feel with AI Worker Tactile Imitation Learning

**URL:** <https://discourse.openrobotics.org/t/teaching-robots-to-feel-with-ai-worker-tactile-imitation-learning/58932>\
**Category:** Projects\
**Tags:** ros2, physical-ai\
**Created:** [October 18, 2026, 11:05pm UTC](https://discourse.openrobotics.org/t/teaching-robots-to-feel-with-ai-worker-tactile-imitation-learning/58932 "2026-10-18T23:05:40Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Pyo](https://sea2.discourse-cdn.com/flex022/user_avatar/discourse.openrobotics.org/pyo/32/15617_2.png) [@Pyo](https://discourse.openrobotics.org/u/Pyo)\
**Post date:** [October 18, 2026, 11:05pm UTC](https://discourse.openrobotics.org/t/teaching-robots-to-feel-with-ai-worker-tactile-imitation-learning/58932/1 "2026-10-18T23:05:40Z")

</div>

[![](https://us1.discourse-cdn.com/flex022/uploads/ros/original/3X/c/0/c0fb020cf96ffb1f380f0cca38c98e9974ece48c.jpeg "AI WORKER #22: Teaching Robots to Feel with AI Worker Tactile Imitation Learning") ](https://www.youtube.com/watch?v=DOjpnAeFu4Y)

Anyone who has trained vision-based imitation learning policies knows where cameras hit their limit: contact-rich manipulation with deformable objects.

Recently, we tested a bimanual paper cup unstacking and stacking task on the AI Worker equipped with dual 20-DoF HX5-D20 hands. Separating two nested paper cups requires just enough grip force to slide the top cup out without crushing it or dragging the bottom cup along. Visually, grabbing one cup versus two cups looks almost identical from the camera feeds.

Since each fingertip on the HX5-D20 has a tactile pressure array, we wanted to see how much tactile feedback actually helps and how the model architecture affects the outcome across a 54-DoF action space. We trained and evaluated three models on the same task:

1. Baseline ACT (0% success): Using only RGB cameras and joint states, the policy struggled to regulate contact force. It either knocked the cups over or failed to separate them at all.

2. Tactile ACT (10% success): Simply projecting the raw fingertip pressure arrays through an MLP and adding them as input tokens to ACT made the grasp much more stable. However, the robot still tended to pick up both nested cups together rather than sliding the top one off.

3. ViTacFormer (40% success): Instead of treating touch as a passive input, this two-stage architecture first predicts future tactile states from visual, proprioceptive, and tactile history, then conditions the action decoder on those predicted touch tokens. Anticipating how fingertip pressure should change allowed the robot to peel off the top cup, stack it, and finish stacking the remaining cup.

The 40% success rate shows there is still plenty of room to push performance further, but jumping from 0% with vision-only to 40% by predicting future tactile states highlights why architecture design matters just as much as adding tactile sensors to the hardware.

Check out the full comparison video, code, and technical documentation below.

🎬 Video: [https://youtu.be/DOjpnAeFu4Y](https://youtu.be/DOjpnAeFu4Y)

📝 Docs: [HX5-D20: ACT with Tactile Sensors | ROBOTIS Docs](https://docs.robotis.com/docs/systems/aiworker/resources/technical_story/tactile_inference/)

💬 Community: [ROBOTIS | Actuator for Physical AI](https://www.robotis.com/en/community.php)