Feel2Grasp: Tactile‑Conditioned Offline RL for Re‑grasping

We study tactile‑driven re‑grasping: given an initial imperfect grasp, a robot should repeatedly adjust its grasp until it reaches a desired contact configuration. We collect teleoperated demonstrations with a front camera and two tactile sensors, encode observations into compact latents, and train an offline RL policy (IQL). A simple tactile success signal (“circle” marker visibility on the tactile pad) lets the policy keep trying until it achieves the intended grasp.

We also found that collecting high quality data was important for Offline RL. We want the policy to learn a behavior of re-grasping. With only successful grasp without re-grasping data can't train the policy to retry. For that when collecting data, we only relyed on the camera and sensor input not watching the scene directly. So when we couldn't get the successful grasp every first try and that led to regrasping motion during data collection. This helped the policy to learn how to try until it properly grasp the right position.

Tactile success signal (circle)

Each tactile sensor provides an image of the contact patch. We use a simple marker (“circle”) on the tactile pad: when the pen is grasped in the intended contact configuration, the circle becomes visible in the tactile image. We detect this per sensor (left/right) and use it as a success indicator.

What matters

Vision helps reach and stabilize around the object.
Tactile specifies the desired contact configuration.
State stacking adds short‑horizon temporal context so the policy can “try again” reliably.

Results

Three stages (two clips each). Click play to compare behaviors.

Stage 1 — No stacking, with tactile sensing

Uses tactile observations but without temporal stacking; failed to get close to the pen.

Stage 1 — No stacking, with tactile sensing

Uses tactile observations but without temporal stacking; got close the pen but no grasping behavior.

Stage 2 — No stacking, without tactile sensing

Vision‑only policy make it stable to reach to pen. Also succeed on grasping but does not re‑attempt on failure.

Stage 2 — No stacking, without tactile sensing

Same situation as above; the policy does not re‑attempt on failure.

Stage 3 — Stacking 4 states, with tactile sensing

Stacking improves temporal understanding; the policy keeps re‑grasping until tactile success.

Stage 3 — Stacking 4 states, with tactile sensing

Stacking improves temporal understanding; the policy grasp properly right away.

Repository

Code and instructions: https://github.com/mjkim001130/Feel2Grasp

Baseline Comparison: ACT

We trained an ACT baseline on the same dataset and task setup. Qualitatively, ACT performs similarly, but our offline RL policy achieves a slightly higher success rate in our runs.

ACT — clip 1

Representative run with ACT.

ACT — clip 2

Second representative run.

Other Objects

We tested the same sensing setup with additional objects to validate that the tactile signal (contact imprint) remains informative beyond a single pen instance.

Tactile sensor and objects

Examples: screw, LEGO brick, key, and their tactile imprints.

Hardware integration

Mounting used for data collection and evaluation.

Representative tactile frames

Tactile appearance differs by object geometry and grasp configuration.

BibTeX

@misc{feel2grasp2025,
  title  = {{Feel2Grasp: Tactile‑Conditioned Offline RL for Re‑grasping}},
  author = {{Kim and Kim and Kim and Baik}},
  year   = {2025},
  howpublished = {\url{https://github.com/mjkim001130/Feel2Grasp}},
}