Abstract

Vision-only grasping pipelines plateau when objects are occluded, deformable, or have uniform surface appearance. Tactile sensing arrays embedded in robotic fingerpads give manipulators a second sensing modality that is contact-local and inherently complementary to cameras. Recent hardware advances in piezoresistive and capacitive taxel arrays, combined with learned tactile representations, are pushing dexterous grasping into regimes that were considered intractable twelve months ago.

Hardware Landscape

The dominant taxel technologies competing in 2026 are piezoresistive elastomer sheets (popularized by the GelSight family from MIT and commercial derivatives from Tactile Robotics) and capacitive grids embedded in silicone fingertips (the approach taken by Sanctuary AI and several university spin-offs). Spatial resolution has crossed the 1mm taxel pitch threshold in lab prototypes, which is sufficient to detect edge geometry and surface slip onset separately. The practical bottleneck has shifted from sensing density to signal conditioning: high-taxel-count arrays generate bandwidth that onboard microcontrollers cannot saturate without custom ASICs or aggressive subsampling.

Learned Tactile Representations

Raw taxel readings are high-dimensional and difficult to use directly in grasp planning. Self-supervised pre-training on tactile time series, using contrastive objectives similar to those applied in visual representation learning, has produced compact embeddings that generalize across object categories. Stanford’s HATO project demonstrated that a tactile encoder pre-trained on household objects transfers to industrial parts with minimal fine-tuning. The encoder architecture matters: convolutional backbones outperform MLPs on spatially structured deformation patterns, but temporal aggregation via short LSTM or state-space layers is necessary to detect slip before it escalates to drop events.

Integration with Grasp Controllers

Tactile feedback is most useful when closed tightly into the grasp controller rather than used as a post-hoc check. Model predictive controllers that treat contact force distributions as state observations can reactively adjust finger poses within 10-20ms, well inside the timescale of incipient slip. The challenge is calibration drift: elastomer materials creep under repeated contact, shifting the resting baseline on timescales of hours. Teams at CMU’s Manipulation Lab have published drift-correction routines that use periodic no-load calibration presses, but field-deployed systems need solutions that work without interrupting operation.

Open Problems

Waterproofing and hygienic cleaning remain unsolved for food and medical applications. Most current taxel arrays are destroyed by autoclave cycles or harsh chemical washes. Until packaging advances close that gap, tactile-equipped grippers will remain in controlled industrial settings rather than expanding into the broader service robotics market that could most benefit from them.