TACTX: Learning Shared Tactile Representations Across Diverse Sensors

Junsung Park1,2,*, Sachin Bhadang1,*, Carmelo Sferrazza3, Sha Yi1, Xiaolong Wang1 1UC San Diego, 2Seoul National University, 3Amazon FAR *Equal contribution; author order determined by coin flip.

A shared tactile latent aligns vision-based, magnetic, and resistive sensors, enabling tactile-conditioned manipulation policies to transfer zero-shot across physically different hardware.

Paper OpenReview Code Dataset

Sensor-Agnostic Touch

Tactile sensors provide critical information for contact-rich manipulation, yet tactile representations and policies remain tightly coupled to each specific sensor, limiting transferability across robots and hardware platforms. We propose TACTX, a framework for learning a transferable tactile representation across sensors spanning three fundamentally different transduction modalities: resistive, magnetic, and vision-based. TACTX maps heterogeneous tactile observations into a shared latent space through modality-specific encoders trained on paired contact data. Such paired interactions provide a natural alignment signal across modalities, and the encoders are jointly trained across all sensor pairs, inducing a consistent latent space for all sensor types. Across multiple manipulation tasks, TACTX improves average zero-shot cross-sensor policy success from 27.5% for vision-only transfer to 45.9%.

TACTX intro animation: a shared latent representation across three sensing modalities.
TACTX latent space animation: paired contacts aligning in the shared tactile latent space.

Paired Contacts Across Heterogeneous Sensors

TACTX trains from simultaneous paired contact observations. Two tactile sensors are mounted on opposing fingers of a parallel-jaw gripper and record the same physical contact event through different sensing mechanisms.

Pairwise datasets connect Daimon, eFlesh, and FlexiTac. Left/right sensor swaps reduce mounting bias, and 10 3D-printed objects provide point, edge, and area contact geometries for alignment.

Ten 3D-printed objects with point, edge, and area contact geometries used for paired data collection.

Pairwise Training, One Shared Latent

Each sensor has its own encoder and decoder. Encoders map native tactile observations into a shared posterior over a 16-D latent space, while decoders reconstruct each modality from that latent.

TACTX combines contrastive alignment with self- and cross-reconstruction. Minibatches draw balanced samples from every sensor-pair dataset, inducing a globally consistent tactile space from pairwise supervision alone.

TACTX architecture: modality-specific encoders and variational latent heads map each sensor into a shared latent space trained with InfoNCE, KL, and reconstruction losses.
TACTX architecture animation: forward pass for one paired contact through the shared latent space.

Zero-Shot Cross-Sensor Policy Transfer

TACTX is evaluated on four contact-rich manipulation tasks, with an additional out-of-distribution pick-and-place setting that changes object color while preserving geometry. Below, we show zero-shot transfer, training on one sensor and deploying on another, chosen randomly for each task.

Shared Tactile Representations Improves Transfer

45.9% average TACTX cross-sensor success rate
27.5% average vision-only transfer success rate
0 new policy demonstrations required at deployment

TACTX achieves the strongest transfer on most task and sensor-pair combinations. The largest gains appear in contact-rich settings such as board wiping and object reorientation, where contact geometry matters beyond binary touch.

Cross-sensor policy transfer across all tasks. Each entry reports the number of successes out of 10 trials, shown as mean ± std over 3 runs. Bold indicates the best performance for each source–deploy combination and task.
Method Source Deploy P&P P&P (OOD) Insertion Wiping Reorient
Vision
Transfer		 Daimon eFlesh 5.3 ± 0.91.7 ± 0.92.7 ± 0.53.0 ± 0.00.3 ± 0.5
FlexiTac 5.3 ± 0.51.3 ± 0.51.3 ± 0.91.3 ± 0.90.7 ± 0.9
eFlesh Daimon 7.7 ± 0.50.0 ± 0.03.7 ± 0.50.0 ± 0.07.0 ± 0.0
FlexiTac 1.3 ± 0.50.7 ± 0.50.3 ± 0.50.3 ± 0.51.3 ± 0.5
FlexiTac Daimon 7.0 ± 0.86.0 ± 0.03.3 ± 0.50.7 ± 0.57.3 ± 0.9
eFlesh 6.7 ± 0.51.0 ± 0.05.0 ± 0.00.3 ± 0.50.0 ± 0.0
Binary
Contact
Transfer		 Daimon eFlesh 4.0 ± 0.01.3 ± 0.92.7 ± 2.12.0 ± 2.81.7 ± 1.7
FlexiTac 3.3 ± 1.22.7 ± 0.52.0 ± 0.82.0 ± 1.44.0 ± 1.6
eFlesh Daimon 3.3 ± 0.90.0 ± 0.00.3 ± 0.51.0 ± 1.42.3 ± 2.1
FlexiTac 4.0 ± 1.61.3 ± 1.20.7 ± 0.90.3 ± 0.56.0 ± 2.2
FlexiTac Daimon 3.3 ± 1.72.7 ± 1.76.7 ± 0.51.3 ± 1.97.3 ± 2.4
eFlesh 1.0 ± 0.82.7 ± 1.90.3 ± 0.51.0 ± 0.82.7 ± 0.9
TactX
Transfer
(Ours)		 Daimon eFlesh 8.3 ± 0.51.0 ± 0.04.0 ± 0.84.0 ± 0.03.7 ± 0.9
FlexiTac 5.3 ± 1.22.0 ± 0.81.3 ± 1.90.3 ± 0.56.7 ± 0.9
eFlesh Daimon 9.0 ± 1.40.7 ± 0.56.0 ± 1.66.0 ± 0.85.0 ± 1.4
FlexiTac 1.3 ± 0.90.0 ± 0.03.7 ± 1.21.3 ± 0.55.0 ± 0.8
FlexiTac Daimon 8.0 ± 1.48.3 ± 1.28.3 ± 0.96.3 ± 0.97.7 ± 1.2
eFlesh 6.7 ± 0.93.0 ± 0.84.7 ± 0.55.7 ± 0.54.3 ± 0.5

Open Resources

Dataset

Paired contact data across Daimon, eFlesh, and FlexiTac with left/right-swapped sensor-pair configurations.

Code

Training and evaluation code for shared tactile representation learning, reconstruction, alignment, and ACT policy transfer.

Checkpoints

TACTX encoder and decoder checkpoints, plus downstream tactile-conditioned policy checkpoints.

BibTeX

If you find our work useful, please consider citing the paper as follows:

@misc{Anonymous2026TACTX,
  title={TACTX: Learning Shared Tactile Representations Across Diverse Sensors},
  author={Anonymous Authors},
  year={2026}
}