TranTac: Tactile Sensing for Robotic Insertion

• TranTac is a tactile sensing framework that uses a single embedded 6-axis IMU to measure micrometer-scale deformations during contact-rich insertion tasks.
• It integrates transient tactile signal processing with transformer-based encoding and diffusion-policy control to output corrective 6-DoF pose adjustments.
• Empirical evaluations show robust performance in tasks like USB, key, and prism-slot insertion with high success rates even with minimal training data.

TranTac is a tactile sensing and control framework for contact-rich robotic manipulation that uses a single contact-sensitive 6-axis inertial measurement unit embedded within elastomeric gripper tips to detect dynamic translational and torsional deformations at the micrometer scale during fine insertion tasks such as key insertion and USB plugging. It combines transient tactile signal processing, transformer-based tactile encoders, multimodal fusion with vision, and a diffusion policy that outputs 6-DoF pose corrections of the grasped object. In the reported formulation, TranTac is explicitly positioned as a data-efficient and low-cost alternative to touch sensing approaches that are either insensitive to subtle changes or require excessive sensor data, and it achieves strong insertion performance when used alone or in combination with vision (Wu et al., 20 Sep 2025).

1. Definition and scope

TranTac denotes the framework introduced by Wu et al. for leveraging transient tactile signals in robotic manipulation, with particular emphasis on insertion tasks in which visual perception is insufficient to detect misalignment (Wu et al., 20 Sep 2025). Its central premise is that dynamic contact events at the gripper’s tip encode visually imperceptible pose changes of the grasped object, and that these cues can be exploited to imitate human insertion behaviors and to perform online correction of the object pose.

The framework is organized around four coupled elements: a customized sensing system based on an embedded 6-axis IMU, a signal model that maps inertial measurements to elastomer deformation, a transformer-based encoder for tactile time series, and a diffusion-policy controller conditioned on fused visual, tactile, and proprioceptive observations. The target regime is not generic tactile perception in isolation, but contact-rich manipulation in which the robot must react to subtle collisions, localize misalignment, and adjust a 6-DoF pose trajectory during insertion.

A common source of confusion is nomenclature. TranTac should be distinguished from "TransTac: Visuo-Tactile Modality Transition via Ultraviolet-Encoded Transparent Elastomers," which describes a transparent ultraviolet-encoded binocular vision-based tactile sensor for integrating visual observation and marker-based tactile reconstruction within a single compact device (Yang et al., 3 Jun 2026). The two systems address tactile sensing from different architectural premises: TranTac centers on transient inertial signals embedded in elastomeric fingertips, whereas TransTac centers on transparent visuo-tactile imaging.

2. Sensor hardware and deformation model

The sensing substrate of TranTac is a custom fingertip built around an ST LSM6DSR iNEMO sensor measuring linear acceleration over 3 axes at $\pm 16\,g$ and angular velocity over 3 axes at $\pm 4000\,dps$ at $3500\,Hz$ (Wu et al., 20 Sep 2025). The IMU is packaged in a two-stage molded PDMS fingertip using Dow SYLGARD 184, and data are streamed via flex cable to a Raspberry Pi 5 and then to a PC for real-time buffering. The reported tip dimensions are $11\times 11\times 8\,mm$, with a hardware cost of $$5$and a data rate of$42\,KB/s$at$3500\,Hz$$.</p> <p>The paper models the elastomeric tip in continuous time by relating measured acceleration and angular velocity to translational and torsional deformation. Let$$y_a(t)\in\mathbb{R}^3$be the measured linear acceleration and$y_\omega(t)\in\mathbb{R}^3$$the measured angular velocity. The translational deformation$$\Delta x(t)\in\mathbb{R}^3$and torsional deformation angle vector$\pm 4000\,dps$0 are given by

$\pm 4000\,dps$1

and

$\pm 4000\,dps$2

Here, $\pm 4000\,dps$3 and $\pm 4000\,dps$4 are calibration matrices mapping the IMU frame to local tip deformation coordinates, and $\pm 4000\,dps$5, $\pm 4000\,dps$6 are constant offsets estimated during zero-motion calibration.

The interpretation given in the source is specific: double integration recovers micrometer-scale tip translation, while single integration recovers tip rotation. This provides a compact route to estimating dynamic local deformation without requiring spatially dense tactile imaging. A plausible implication is that TranTac prioritizes temporal fidelity and deformation transients over explicit spatial contact maps.

3. Transient tactile signal processing

TranTac’s processing pipeline is explicitly designed for transient events rather than static load estimation (Wu et al., 20 Sep 2025). Preprocessing uses a high-pass filter with cut-off approximately $\pm 4000\,dps$7 to remove quasi-static drift, a low-pass filter with cut-off approximately $\pm 4000\,dps$8 to reject sensor electronic noise, normalization by the sensor full-scale range, and sliding-window segmentation using a $\pm 4000\,dps$9 collision window corresponding to $3500\,Hz$0 samples aligned to contact events.

Within each window $3500\,Hz$1, the filtered signals are written as

$3500\,Hz$2

after which the band-limited deformations $3500\,Hz$3 and $3500\,Hz$4 are computed using the deformation equations above. The resulting deformation estimates are embedded into a 6-DoF pose increment of the grasped object:

$3500\,Hz$5

This formulation makes the tactile signal an estimator of object pose change rather than merely a contact/no-contact indicator. In the reported setup, the tactile observation stream is substantially denser than the visual stream: $3500\,Hz$6 IMU samples per visual frame, reflecting the mismatch between $3500\,Hz$7 tactile sensing and $3500\,Hz$8 RGB capture. This suggests that TranTac treats high-rate contact transients as the primary source of corrective motion information when insertion enters regimes where visual cues become insufficiently discriminative.

4. Representation learning and policy architecture

The tactile encoder in TranTac is transformer-based (Wu et al., 20 Sep 2025). Each $3500\,Hz$9-dimensional IMU sample is projected by an MLP into a token of dimension $11\times 11\times 8\,mm$0, and a 1D sinusoidal positional encoding is added. For a token sequence $11\times 11\times 8\,mm$1, the paper uses standard multi-head attention:

$11\times 11\times 8\,mm$2

$11\times 11\times 8\,mm$3

$11\times 11\times 8\,mm$4

The tactile encoder stack uses $11\times 11\times 8\,mm$5 layers and outputs $11\times 11\times 8\,mm$6 tokens of dimension $11\times 11\times 8\,mm$7 for each fingertip.

For multimodal fusion, a ResNet-18 processes RGB images at $11\times 11\times 8\,mm$8 into a $11\times 11\times 8\,mm$9 feature map, which is flattened into $$0 visual tokens with 2D positional encoding. These visual tokens are concatenated with the tactile tokens from both fingertips and passed through $$1 fusion layers of standard Transformer blocks, after which the fused representation is pooled and projected to a $$2-dimensional embedding $$3. The diffusion policy is then conditioned on observations $$4.

The control module models a ground-truth future trajectory $$5 through a diffusion process:

$$6

with closed form

$$7

and training objective

$$8

At inference, the first 6-DoF pose delta in the recovered denoised trajectory is applied as the robot command. Action chunking and exponential temporal averaging are used for smooth control at $$9.

5. Demonstration learning and task regime

TranTac is trained by imitation learning using $and a data rate of$0 trajectories per task collected via teleoperation with a see-through VR headset at $and a data rate of$1 (Wu et al., 20 Sep 2025). The tasks listed are $and a data rate of$2 prism-slot insertion, USB insertion, key insertion, and circle-square insertion. Training uses Adam with learning rate $and a data rate of$3, batch size $and a data rate of$4, and $and a data rate of$5 steps. An optional behavior cloning baseline is also defined by

$and a data rate of$6

The source characterizes the framework as data-efficient in two explicit senses. First, it reports that $and a data rate of$7 demonstrations, corresponding to at most $and a data rate of$8 of data, yield robust policies. Second, the tactile interface produces much lower data volume than image-based visuo-tactile systems: $and a data rate of$9 for the TranTac tip, compared with competing visuo-tactile systems listed as $at$0–$at$1 for spatial-only signals and $at$2–$at$3 for image streams.

The emphasis on transient cues is central to the regime of applicability. TranTac is not presented as a general force sensor, and it does not claim static or pseudo-static force sensing. Instead, it targets manipulation episodes in which brief contact events, collision impulses, and torsional perturbations carry the relevant state information for correcting insertion trajectories.

6. Empirical performance

The reported empirical evaluation includes both visuo-tactile policies and tactile-only misaligned insertion tasks (Wu et al., 20 Sep 2025). In the visuo-tactile setting, the source reports an average success rate of $at$4 on object grasping and insertion tasks when TranTac is combined with vision, outperforming both a vision-only policy and a policy augmented with end-effector 6D force/torque sensing. The detailed success rates over $at$5 trials per task are as follows.

Sensor modality	Rect. Insertion	Circle-Square
Vision Only	75%	60%
Vision + Force/Torque	50%	70%
Vision + TranTac	80%	80%

Sensor modality	USB Insertion	Key Insertion
Vision Only	30%	80%
Vision + Force/Torque	40%	40%
Vision + TranTac	65%	90%

These figures show that the largest absolute gain over vision-only appears in USB insertion, where success rises from $at$6 to $at$7, while key insertion improves from $at$8 to $at$9. Relative to the force/torque baseline, the improvement is substantial across all four tasks listed.

For tactile-only contact localization under misaligned insertion, the paper reports an average success rate of $y_a(t)\in\mathbb{R}^3$0 on the training object, a $y_a(t)\in\mathbb{R}^3$1 prism, over $y_a(t)\in\mathbb{R}^3$2 trials per object with initial lateral offsets $y_a(t)\in\mathbb{R}^3$3. Generalization is evaluated by training on a single prism-slot pair and testing on unseen objects including a USB plug and a metal key; the source states that insertion can still be completed with an average success rate of nearly $y_a(t)\in\mathbb{R}^3$4.

The paper also reports contact localization performance as a distinct capability. In context, this means that transient tactile signals are sufficient not only for policy conditioning during visuo-tactile control but also for tactile-only correction in misaligned insertion. This suggests that the learned representation captures task-relevant structure in the local deformation dynamics rather than merely reflecting gross collision magnitude.

7. Generalization, limitations, and relation to adjacent tactile paradigms

The generalization claim in the source is narrowly specified: single-object training suffices to generalize to novel peg-in-hole geometries, and the evaluation includes unseen cylinders, prisms, a USB plug, and a metal key (Wu et al., 20 Sep 2025). Within that scope, TranTac is presented as a framework capable of transferring insertion behavior learned from one geometry to related but unseen contact conditions.

The limitations are also explicit. TranTac has no static or pseudo-static force sensing because zero-frequency content is not captured. It lacks explicit spatial contact mapping, and future work is described as increasing IMU count or using super-resolution decoding. The paper further notes an incomplete physics model between 6-DoF motion and IMU signals, that the elastomeric tip size can be further miniaturized, and that diffusion-policy inference of approximately $y_a(t)\in\mathbb{R}^3$5 per step limits real-time reactivity.

These limitations help position TranTac among tactile sensing paradigms. Compared with vision-based tactile sensors, its sensing bandwidth and compactness are advantageous in transient contact regimes, while the absence of explicit spatial contact maps constrains geometric observability. Compared with end-effector force/torque sensing, it moves the sensing locus to the deformable fingertip and targets micrometer-scale dynamic deformation rather than distal wrench estimation. A plausible implication is that TranTac is best understood as a specialized method for high-frequency local contact inference during manipulation, rather than as a replacement for all tactile or force sensing modalities.

The broader conceptual contrast with TransTac is instructive. TransTac preserves visual transparency through a clear elastomer interface and integrates RGB observation with marker-based tactile reconstruction in a single compact device (Yang et al., 3 Jun 2026). TranTac, by contrast, derives its utility from transient inertial signals embedded in soft tips and couples them to transformer and diffusion-policy architectures for online 6-DoF corrective control. The shared theme is multimodal contact perception, but the sensing physics, data representation, and intended control loop are fundamentally different.