Which Taxels, Not How Many: Tactile Sensor Placement for Dexterous Policies

Taeyun Kim, Jinsoo Shin, Taehwa Kim, Taeyoon Kim, Jinyoung Choi

November 2026

TL;DR A controlled simulation study of where to put tactile sensors on a robot hand. For in-palm rotation, 10 taxels chosen from the teacher policy's contact statistics reach 41.4% success, while uniform placement scores 0% with 10 and needs 100 taxels to reach 51.4%.

What they do

Simulation studies of tactile dexterous policies usually spread taxels uniformly over the hand. This paper tests that default. It fixes the teacher policy, the training procedure and a pool of 2000 candidate taxels, and varies only which taxels the student sees and how their readings are arranged for the encoder. Tasks are in-palm rotation (XHand1) and grasp-and-lift under disturbances (Sharpa, Allegro), all in Genesis.

How it works

Teacher-student. A PPO teacher with privileged state and no touch is distilled with DAgger into students that see joint history, the goal and tactile input, so the teacher’s contacts do not depend on any placement.
Placement. With a budget of 10 to 300 taxels, compare uniform placement against three rules built from the teacher’s contact statistics (top-contact, contact-cover, link-proportional), plus a least-contact control and a per-link net force reference.
Representation. With all 2000 taxels, compare random, band and grid tensor layouts under CNN, LSTM and link-wise MLP encoders.

Results

In-palm rotation: top-contact and contact-cover reach 41.4% and 38.7% success with 10 taxels and about 50% with 25. Uniform gets 0% at 10 and 15, 12.4% at 25, and catches up only at 100 to 300 (56.2% at 300). All 2000 taxels give 61.5%; the teacher gets 93.8%.
No touch, least-contact taxels and per-link force give 0%, 0% and 4.4%.
Layout barely matters: the CNN scores 61.4%, 64.3% and 61.3% on random, band and grid. The link-wise MLP, which averages away taxel identity, drops to 18.6-33.2%.
Grasp-and-lift follows a different trend: at 300 taxels contact-cover beats uniform on both hands (57.0% vs. 42.6% held fraction on Sharpa, 72.2% vs. 57.0% on Allegro), while top-contact falls to 27.0% on Sharpa.
Simulating only 100 selected taxels trains 5.3x faster than the full observation.

Limitations

Everything is in simulation; sensor noise and drift are untested. The main lessons come from one hand and one task where contact concentrates on a few links, and placement behaved differently on grasp-and-lift. Students with all taxels still trail the teacher.

Figures

From Figure 2: Taxel placement on the XHand1 for in-palm rotation. Contact frequency of the teacher policy at each ca... — From Figure 2: Taxel placement on the XHand1 for in-palm rotation. Contact frequency of the teacher policy at each candidate taxel, and the taxels selected by each placement method (shown for a budget of 100). Faded fingers are not used.

From Figure 3: Tensor layouts of the selected taxels. An example contact caused by an external force on the palm is m... — From Figure 3: Tensor layouts of the selected taxels. An example contact caused by an external force on the palm is mapped into the tensor under each layout.

From Figure 5: Teacher rollouts on the three hands. Each row shows frames sampled at 2 s intervals from a single teac... — From Figure 5: Teacher rollouts on the three hands. Each row shows frames sampled at 2 s intervals from a single teacher episode. The Allegro and the Sharpa are shown on grasp-and-lift, and the XHand1 on in-palm rotation.

Abstract

Policies for dexterous manipulation depend on design choices made before training: which sensors enter the observation, where they are placed, and how their readings are arranged for the encoder. These choices are rarely examined. For tactile sensing, simulation studies place taxels uniformly over the hand, and this placement has become a default without being tested. We test this default under a single controlled protocol. The protocol holds the teacher, the training procedure, and the candidate taxels fixed and varies only the sensor configuration and the representation. We evaluate it on in-palm rotation and grasping with three robot hands. A small set of taxels selected from contact statistics matches a far larger uniform set. Taxels at rarely contacted sites are equivalent to no tactile input. With all candidate taxels, their layout in the input tensor does not affect performance, even for a convolutional encoder. Uniform placement therefore overestimates the number of taxels a task requires. We distill these results into lessons for designing tactile policies and into a minimal set of design choices that a tactile policy study should report.

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.

Matt Strong