Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

Tactile Sensors Are an Effective Substitute for Wrist Cameras in Contact-Rich or Occluded Manipulation

Aijia Yang, Arsh Tangri, Sergey Zakharov, Krishna Murthy Jatavallabhula

November 2026

Montage: occluded retrieval, contact-rich insertion and whiteboard wiping, with the tactile maps on the right.

TL;DR A controlled study on a real arm with ACT policies and two low-cost tactile pads on the fingers. Touch without a wrist camera beats vision with a wrist camera on bag retrieval (81.7% vs 50.0%), plug insertion (46.7% vs 33.3%) and whiteboard wiping (71.7% vs 46.7%).

What they do

Manipulation policies usually get close-range feedback from a wrist camera, but the fingers often hide the contact that matters. This paper asks whether finger-mounted touch can take the wrist camera’s place. It is a controlled study: the same demonstrations, the same ACT policy and the same training settings, with only the tactile input and the wrist images switched on or off.

How it works

  • Hardware. A YAM arm with scene, egocentric and wrist RGB cameras, and two FlexiTac piezoresistive arrays on the inner faces of compliant TPU fingers. Touch is recorded at about 99 Hz, images at about 30 Hz.
  • Policy. ACT predicts 19-step chunks of 7-DoF joint positions. Each 32x16 tactile map is resized to 224x224 and passed through the same ImageNet-pretrained ResNet-18 as the camera views, so touch enters as two extra images.
  • Tasks. Tight-clearance block placement, rigid-block and sponge placement, straw insertion, marker retrieval from an opaque bag, plug insertion and whiteboard wiping. The last three are run with and without wrist images.
  • Evaluation. 1,000 real rollouts over 20 task-and-observation conditions from matched initial conditions, with Holm-corrected Barnard tests.

Results

  • Without wrist images, adding touch raises success by 60.0 points on bag retrieval (21.7% to 81.7%), 31.7 on plug insertion (15.0% to 46.7%) and 50.0 on wiping (21.7% to 71.7%). Adding a wrist camera instead gives 28.3, 18.3 and 25.0 points.
  • Touch without wrist images beats vision with wrist images on all three: 81.7% vs 50.0%, 46.7% vs 33.3% and 71.7% vs 46.7%, significant for retrieval and wiping.
  • Straw insertion improves from 43.3% to 65.0% with touch. On the placement tasks the gains (0 to 13.3 points) are not significant.
  • Adding wrist images to the visuotactile policy does not help consistently (81.7% to 71.7% on retrieval, 71.7% to 65.0% on wiping).

Limitations

One robot, one tactile sensor type, one policy-learning method and one training seed. Why the visuotactile policy scores lower with wrist images is left unresolved. The low-cost pads show baseline variation after replacement. Scene and egocentric cameras are kept in every condition, so touch replaces only the wrist view.

Videos

Retrieving a marker from an opaque bag, with the camera views and both tactile maps.
Placing a block into a box.

Figures

From Figure 1: The YAM station with scene, egocentric and wrist cameras and tactile pads on the gripper fingers.
From Figure 1: The YAM station with scene, egocentric and wrist cameras and tactile pads on the gripper fingers.
From Figure 1: Camera views next to the two tactile maps.
From Figure 1: Camera views next to the two tactile maps.
From Figure 3: Tactile gain by task and wrist-camera availability. Posterior mean and 90% credible interval for the d...
From Figure 3: Tactile gain by task and wrist-camera availability. Posterior mean and 90% credible interval for the difference in success probability between vision+tactile and vision-only.
From Figure 4: Posterior success probability for every task. Top row: the wrist-on tasks (T1-T3), two conditions each...
From Figure 4: Posterior success probability for every task. Top row: the wrist-on tasks (T1-T3), two conditions each. Bottom row: the ablated tasks (T4-T6), four conditions each. Letters come from Holm-corrected pairwise Barnard tests within each panel; conditions sharing a letter could not be distinguished.
Abstract

Learned manipulation policies often rely on wrist-mounted cameras for close-range visual feedback that complements external views of the workspace. However, contact-rich manipulation requires establishing and maintaining contacts that can remain hidden even from wrist cameras. Tactile sensors mounted on fingers can reliably and directly measure contact, without needing visual access. In this work, we show that this feedback enables policies to recover the performance advantage of wrist cameras using scene and egocentric vision together with touch. Our approach incorporates two low-cost tactile arrays into a visuotactile ACT policy, without wrist images during training or deployment. We evaluate six real-world task families spanning placement, insertion, occluded retrieval, and sustained-contact tool use. On three contact-rich tasks, we independently vary wrist-camera and tactile availability while keeping the demonstrations, policy architecture, and optimization settings fixed. Adding wrist cameras improves success by 18.3-28.3 percentage points over vision without wrist images. Adding tactile sensing instead improves success by 31.7-60.0 percentage points and yields higher observed success than vision-only policies with wrist cameras on all three tasks, with significant advantages on two under Holm-corrected tests. These results support finger-mounted tactile sensing as an effective alternative to wrist cameras in the evaluated contact-rich settings.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.