Which Taxels, Not How Many: Tactile Sensor Placement for Dexterous Policies
Organizations: NAVER LABS, Seongnam, Republic of Korea
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.
Figures & tables
| Task-agnostic | Task-informed | |||
| Budget ( ) | Uniform | Top-contact | Contact-cover | Link-prop. |
| No tactile | 0 0.0 0.0 | |||
| Least-contact | 0 0.0 0.0 | |||
| Per-link force | 0 4.4 1.0 | |||
| 10 | 0 0.0 0.0 | 41.4 0.6 | 38.7 2.4 | 0 0.0 0.0 |
| 15 | 0 0.0 0.0 | 42.5 7.6 | 42.3 1.0 | 0 0.0 0.0 |
| Task-agnostic | Task-informed | |||
| Budget ( ) | Uniform | Top-contact | Contact-cover | Link-prop. |
| Sharpa | ||||
| No tactile | 0 9.4 0.7 | |||
| 25 | 15.9 0.7 | 22.3 6.7 | 27.9 8.5 | 21.2 3.9 |
| 50 | 24.0 0.8 | 29.1 12.5 | 38.9 15.7 | 26.6 5.5 |
| 100 | 25.6 3.4 | 24.3 1.3 | 32.0 5.4 | 26.9 2.8 |
| Tensor layout | |||
| Encoder | Random | Band | Grid |
| CNN | 61.4 0.6 | 64.3 1.3 | 61.3 1.1 |
| LSTM | 60.8 3.9 | 58.0 3.2 | 61.5 0.3 |
| Link-wise MLP | 33.2 8.6 | 25.2 3.3 | 18.6 1.1 |
| Tactile observation | Env-steps/s | Speed-up | VRAM |
| None | 42,810 | 7.6 | 0 1.5 GB |
| 32,233 | 5.7 | 0 1.6 GB | |
| 29,503 | 5.3 | 0 1.9 GB | |
| 23,817 | 4.2 | 0 2.5 GB | |
| 0 8,411 | 1.5 | 0 8.1 GB | |
| Full (ours) | 0 5,610 | 1.0 | 12.4 GB |
| Design choice | Setting |
| Sensing | |
| Candidate set | , Poisson-disk sampling on the used links |
| Taxel signal | 6-D force and torque in the link frame |
| Selection (RQ1) | |
| Budget † | |
| Placement rule † | Uniform, Top-contact, Contact-cover, Link-proportional , and Least-contact |
| Reward Term | Expression |
| Task Rewards : | |
| Goal Reached ( ) | |
| Rotation Progress ( ) | |
| Fingertip Proximity ( ) | |
| Drop ( ) | |
| Regularization : | |
| Reward Term | Expression |
| Task Rewards : | |
| Sustained Lift ( ) | |
| Goal Tracking ( ) | w_{2}\;\exp(-\lVert\mbox{\boldmathp}_{t}-\mbox{\boldmathp}^{\text{goal}}\lVert_{2}^{2}/0.25^{2}) |
| Drop ( ) | |
| Regularization : | |
| Fingertip Proximity ( ) | |
| Setting | Value |
| Simulation | |
| Physics step | 5 ms |
| Control step | 25 ms |
| Episode length | 10 s / 20 s |
| Teacher (PPO) | |
| Parallel environments | 8192 / 4096 |
| Tactile encoders | ||
| Encoder | Spatial module | Temporal |
| LSTM | – | LSTM(64) |
| CNN | Conv3 3(16)–LN–Conv3 3(32)–LN | GRU(64) |
| Link-wise MLP | FC(64)–FC(64), shared, ELU | GRU(64) |
| Shared | ||
| Proprioceptive encoder | LSTM(128)–FC(64) | |
| Task-agnostic | Task-informed | |||
| Budget ( ) | Uniform | Top-contact | Contact-cover | Link-prop. |
| Contact Events (%) | ||||
| 0 10 | 0.1 | 83.9 | 0 88.8 | 0 0.1 |
| 0 15 | 0 0.2 | 84.7 | 0 94.8 | 0 0.1 |
| 0 25 | 24.6 | 85.6 | 0 98.1 | 25.1 |
| 0 50 | 35.9 | 87.5 | 0 99.7 | 54.4 |
| Placement | Success | Drop | Under-rotation | ||||
| 0 | No tactile | 0.0 | 0.0 | 15.0 | 12.7 | 85.0 | 12.7 |
| 25 | Uniform | 12.4 | 1.3 | 30.8 | 5.1 | 56.8 | 4.5 |
| Top-contact | 50.5 | 1.3 | 16.2 | 1.1 | 33.3 | 3.2 | |
| Contact-cover | 50.0 | 1.8 | 14.6 | 3.6 | 35.4 | 4.5 | |
| Link-prop. | 8.6 | 3.1 | 30.2 | 5.3 | 61.2 | 4.7 | |
| 100 | Uniform | 51.4 | 2.3 | 16.5 | 3.2 | 32.1 | 1.1 |
| Hand | Top-100 Share (%) | Entropy | Links Spanned |
| In-Palm Rotation | |||
| XHand1 | 73.5 | 0.70 | 0 3 / 0 6 |
| Grasp-and-Lift | |||
| Allegro | 66.0 | 0.72 | 16 / 39 |
| Sharpa | 38.3 | 0.84 | 0 5 / 17 |
Tactile Sensing for Robotic Manipulation
Tactile sensing for robotic manipulation uses contact signals to help robots perceive forces, prevent slips, and handle objects reliably.
See all 189 papers