Giving Robots a Sense of Touch: Physical AI Advances with Tactile Sensors

Robotic manipulation remains one of the harder unsolved problems in automation engineering. Vision-based systems have matured considerably — object localization, pose estimation, and grasp planning from RGB-D data are now reliable enough for structured industrial environments. What vision cannot provide is contact information: whether a grasp is stable, whether a surface is beginning to slip, or how force is distributed across a fingertip during a hold. These signals are what close the control loop during manipulation, and without them, systems compensate through excessive grip force, conservative motion profiles, and large training datasets designed to paper over sensing uncertainty.

Tactile sensing addresses this gap directly, but industrial adoption has been slow. The primary barrier is not technological immaturity — capable tactile hardware has existed in research settings for decades. The barrier is interpretation: unlike cameras, where resolution, frame rate, and dynamic range map predictably onto system performance, tactile sensing lacks a shared understanding. There is no industry consensus on what signals a useful tactile sensor must capture, at what bandwidth, or at what spatial resolution.

That ambiguity has a direct cost: A company planning hundreds of thousands of grasps to train a policy needs confidence the sensor is capturing the right physical phenomena. One approach is not to derive that specification analytically, but to look at the system that already performs dexterous manipulation better than any robot built to date.

Human Hands as the Benchmark

The sensory-motor physiology of the human hand is the best-characterized model of dexterous manipulation available. Johansson and Vallbo’s foundational 1979¹ study recorded from 334 low-threshold mechanoreceptive units in the glabrous skin of the hand, classifying them into four groups based on adaptation rate and receptive field size. The four classes divide into two functional modes: SA types respond to sustained indentation, encoding static pressure distribution, edge geometry, and skin stretch; FA types respond to dynamic mechanical events — flutter, vibration, and contact transients.

The practical importance of this dual-modality architecture becomes clear during object handling. Johansson and Westling² demonstrated directly that human grasp control is event-driven: FA-type afferents provide the fast transient signals that trigger corrective responses when slip is detected, while SA types maintain the sustained contact map that guides grip force regulation. The two channels are not redundant — they operate at different frequencies and serve different control functions, and both are required for reliable manipulation.

Table 1: Human Mechanoreceptor Reference vs. TSF-85 Modalities.

This physiology gives engineers a concrete specification target. A tactile sensor capable of replicating dexterous manipulation needs to capture both static pressure distribution and dynamic contact events, ideally through the same sensing element and over the same contact region, to avoid interference and added mechanical complexity that come from stacking separate sensor layers. It also needs a channel for fingertip orientation — the spatial context required to correctly interpret a tactile pressure map when the finger pose is not directly observable from the actuator. Table 1 maps each mechanoreceptor class to its functional role and to the corresponding sensing channel in the TSF-85.

The TSF-85 Design

Capacitive sensing was selected as the transduction principle for several reasons specific to gripper fingertip constraints. Unlike optical sensors, which require an imaging cavity, internal illumination, and an elastomer surface that degrades under repeated compression, capacitive sensing has no optical path and no consumable mechanical element. Unlike magnetic sensing, it requires no ferromagnetic compatibility and introduces no calibration complexity from field interference. Critically, capacitive circuit architectures are manufacturable at the scale and cost that industrial deployment demands. The key design challenge was the integration of two distinct capacitive circuits — one static, one dynamic — onto a single PCB layer without crosstalk, using a shared dielectric.

Figure 1 - TSF-85 PCB design with shared dynamic and static sensing. (Image: Robotiq)

The sensor PCB spans 22 mm × 37 mm and carries two distinct capacitive circuits on the same plane. The first is a static array of 28 taxels arranged in a 4×7 grid, read out via a capacitance-to-digital converter (CDC) that measures absolute capacitance at each taxel and maps pressure distribution across the contact surface (Figure 1). The second is a single dynamic taxel integrated around the perimeter of the static array — sharing the same dielectric layer but processed through a high-gain transimpedance amplifier (TIA) that measures capacitance change rather than absolute value. The TIA path captures signals up to 1,000 Hz, spanning the frequency ranges associated with both FA I (5–50 Hz) and FA II (40–400 Hz) mechanoreceptor classes established in Table 1.

The Shared-Dielectric Architecture

Figure 2 - Different layers of the TSF-85. (Image: Robotiq)

Both circuits operate through the same dielectric layer, whose microstructure governs the electromechanical response under load. The current design represents an evolution from the precision-molded microstructures described in the original published architecture, refined to improve manufacturing consistency at industrial scale while preserving the sensitivity characteristics required for both sensing modes (Figure 2).

A special geometry of the capacitive traces exploits fringe capacitance to improve sensitivity and spatial uniformity across the static array. Critically, this shared-layer approach eliminates the registration errors and inter-layer crosstalk that would result from stacking separate sensing layers — a common integration penalty in multimodal tactile sensor designs.

Proprioception via IMU

An integrated IMU completes the sensing picture. For an underactuated gripper, finger orientation is not directly observable from the motor encoder at the base joint. The IMU provides fingertip attitude, giving the system the spatial context needed to interpret tactile maps correctly. The accelerometer channel also provides a second, independent source of vibration data, effectively allowing two independent measurements of the same underlying physical phenomenon.

Use Cases

Tactile signals must be processed and paired with algorithms to extract manipulation-relevant information. Over a decade of research with this sensor has validated several high-value use cases.

Kwiatkowski et al.³ demonstrated grasp stability prediction by fusing tactile signals with proprioceptive data through convolutional neural networks, achieving reliable failure detection across novel objects prior to any lift attempt. That work was extended in Kwiatkowski et al.⁴, which examined the limitations of small and unbalanced tactile datasets on classification performance, a finding directly relevant to teams designing large-scale data collection programs.

Roberge et al.⁵ addressed dynamic tactile event classification using sparse coding of power spectral density signals, distinguishing object-gripper slip from object-world slip in real time. That group subsequently extended the work to in-hand object recognition through grasp-centric exploration, enabling object identification during a squeeze without prior object models.⁶

Stepputtis et al.⁷ demonstrated dynamic regrasping by training a deep predictive model to anticipate slip propagation, then exploiting gravity and acceleration as extrinsic perturbations to reposition objects within the gripper without added mechanical complexity.

Sensor Life Cycle and Signal Processing

Translating research-validated use cases into deployment requires sensors that produce consistent, repeatable output across millions of cycles. Two metrics determine whether that bar is met: durability over operational lifetime, and signal consistency across sensors and taxels.

Figure 3: Sum of the taxel readings over 1.5M grasp cycles. (Image: Robotiq)

Accelerated life-cycle testing beyond 1.5 million grasp cycles on an uneven contact surface shows stable tactile response with no meaningful degradation in contact location or relative intensity across the array — and testing is ongoing. For context, robotic foundation models typically require training data in the order of billions of examples. That volume is predominantly generated in simulation, but simulation does not accurately reproduce real-world contact mechanics, making physical data collection a necessary component of any manipulation training pipeline. Figure 3 plots the sum of taxel readings across the full test duration.

Figure 4: Single taxel response from 37 different sensors. The left figure shows the raw signals and the right figure uses a gain to align the sensors to output 500 counts at a load of 100N. (Image: Robotiq)

Signal consistency must be treated on two fronts, between taxels and between sensors. Both types of variance may be addressed through a straightforward calibration routine: applying a known load to the sensor and computing the gain factor required to align each taxel’s output to a target value. Early results show this approach substantially reduces inter-taxel variance and aligns responses across sensors, as shown in Figure 4.

Figure 5: TSF-85 mounted on a Robotiq 2f-85 Gripper. Response from three taxels from five loading and unloading cycles. (Image: Robotiq)

Figure 5 shows the inter-taxel variation. While each taxel has a unique response, the repeatability of the signal can be observed on the five loading and unloading curves in the plot.

The hysteresis of the sensor is observable in Figure 5. Because the sensor response exhibits hysteresis, it is optimized for contact detection and orientation estimation, which depend on relative signal patterns, rather than absolute force measurement.

Why Now

Tactile sensors have existed for years, but what has changed is the systems it feeds into. Physical AI systems have matured to the point where multi-modal contact data can be productively consumed. The next step is to validate the impact on performance and training efficiency.

This article was written by Jennifer Kwiatkowski, AI Specialist, Robotiq (Quebec City, Canada). For more information, visit here .

References

R. S. Johansson and A. B. Vallbo, “Tactile sensibility in the human hand: relative and absolute densities of four types of mechanoreceptive units in glabrous skin,” Journal of Physiology, vol. 286, pp. 283–300, 1979, doi: 10.1113/jphysiol.1979.sp012619.
R. S. Johansson and G. Westling, “Roles of glabrous skin receptors and sensorimotor memory in automatic control of precision grip when lifting rougher or more slippery objects,” Experimental Brain Research, vol. 56, pp. 550–564, 1984.
J. Kwiatkowski, D. Cockburn, and V. Duchaine, “Grasp stability assessment through the fusion of proprioception and tactile signals using convolutional neural networks,” in Proc. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Vancouver, BC, Canada, 2017, pp. 286–292, doi: 10.1109/IROS.2017.8202170.
J. Kwiatkowski, M. Jolaei, A. Bernier, and V. Duchaine, “The good grasp, the bad grasp, and the plateau in tactile-based grasp stability prediction,” in Proc. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Kyoto, Japan, 2022, pp. 4653–4659, doi: 10.1109/IROS47612.2022.9981360.
J.-P. Roberge, S. Rispal, T. Wong, and V. Duchaine, “Unsupervised feature learning for classifying dynamic tactile events using sparse coding,” in Proc. IEEE International Conference on Robotics and Automation (ICRA), Stockholm, Sweden, 2016, pp. 2675–2681, doi: 10.1109/ICRA.2016.7487428.
J.-P. Roberge, L. L’Écuyer-Lapierre, J. Kwiatkowski, P. Nadeau, and V. Duchaine, “Tactile-based object recognition using a grasp-centric exploration,” in Proc. IEEE International Conference on Automation Science and Engineering (CASE), 2021, pp. 494–501.
S. Stepputtis, Y. Yang, and H. Ben Amor, “Extrinsic dexterity through active slip control using deep predictive models,” in Proc. IEEE International Conference on Robotics and Automation (ICRA), Brisbane, QLD, Australia, 2018, pp. 3180–3185, doi: 10.1109/ICRA.2018.8461055.