All articles
RoboticsPhysical AITactile SensingDexterous Manipulation

Tactile sensing for robots in 2026: why touch is becoming the missing layer of dexterous robot hands

Robots have learned to see and plan, but most still cannot feel what their fingers are doing. Between October and November 2026 a run of results from Stanford, NYU and Columbia, Toyota Research Institute, York University, Sharpa and Google DeepMind showed how much touch changes the picture, from a $10 full-hand tactile skin to fingertip sensors that beat wrist cameras. What a robot tactile sensor actually measures, what the new evidence shows, where its limits are and a practical playbook for technology leaders.

By Yassineh Pourebrahim · · 9 min read

Over the past two years robots have become much better at seeing a scene, understanding an instruction and planning a sequence of steps. The last few centimetres are still the hardest part. When fingers close around a cup, a cable or a test tube, the contact itself is often hidden from every camera, and a robot that relies on vision alone has to guess whether it is holding the object firmly, slipping or pressing too hard. Tactile sensing for robots is the attempt to close that gap. In October and November 2026 several independent results, from university labs, a corporate research institute, a robot-hand maker and Google DeepMind, made a stronger case than ever that touch is becoming a core layer of the robot stack rather than a research curiosity.

What a robot tactile sensor actually measures

A camera tells a robot where things are. A tactile sensor tells it what is happening at the point of contact. The devices in this article measure different combinations of four signals:

  • Contact and location: whether a finger or palm region is touching something, and where.
  • Normal force: how hard the surface is being pressed.
  • Shear force: the sideways force along the surface, which is the early sign that an object is starting to slip or that a stopper is coming free.
  • Contact geometry: the shape of the contact patch, which helps estimate how an object sits in the hand.

The technologies differ. Sharpa's Dynamic Tactile Array uses vision-based sensing: a miniature camera inside each fingertip watches a compliant surface deform. Stanford's SoTa skin is capacitive. NYU and Columbia's PanoTac combines a piezoresistive array with magnets and magnetometers to capture shear. York University's work uses tri-axial piezoresistive force sensors on each finger. The common idea is that the robot gets a direct measurement of contact instead of inferring it from pixels.

What the new tactile sensing research shows

SoTa: a $10 full-hand tactile skin for humans and robots

On 1 October 2026 researchers from Stanford University, with co-authors from Tsinghua University and the University of Cambridge, published SoTa, a soft capacitive tactile skin that covers the whole hand, fingers and palm, with a shared layout of 202 taxels (tactile pixels). The same layout is used on a human demonstrator's hand and on the robot hand, so a single tactile encoder can learn from both without a learned mapping between sensors. The authors report under $10 in materials per skin, more than 97% of the initial response span retained after 10,000 loading cycles, and electrical continuity through 1,280 tight-fist folding cycles.

The most important number is about data rather than hardware. With a fixed budget of robot demonstrations, adding cheaper human demonstrations wearing the same skin raised mean success across eight evaluation conditions from 22.8% to 45.9%, and improved success in all five out-of-distribution conditions. Tactile observations improved in-distribution success over vision-only policies on all three tasks (box reorientation, cup picking and plug insertion), although the authors note that these benefits did not consistently extend to out-of-distribution conditions. They plan to open-source the resources needed to build the skins.

PanoTac: shear feedback for fingertips under $50

On 2 November 2026 researchers from New York University, City University of New York and Columbia University published PanoTac, a compliant fingertip for a parallel-jaw gripper that gives a 12 × 32 grid of taxels, each with one normal and two shear components, at 60 Hz, for less than $50 per fingertip. On three contact-rich tasks, test-tube preparation, grain scooping and salt pouring, visuo-tactile policies reached 91.7% overall success against 45.0% for vision-only policies, an average gain of 46.7 percentage points. Ablations showed that removing either the spatial contact information or the shear feedback reduced performance.

The authors are candid about limits: the magnetometers are sensitive to ferromagnetic objects and external magnetic fields, force resolution is bounded by the elastomer and the number of magnetometers, and the current fingertip is shaped for a two-finger gripper rather than a dexterous hand.

Toyota Research Institute: touch can replace the wrist camera

On 7 November 2026 a study with authors from Toyota Research Institute asked a practical engineering question: can finger-mounted tactile sensors substitute for the wrist-mounted cameras that many manipulation systems rely on? Using two low-cost tactile arrays on a robot arm and a fixed imitation-learning pipeline (Action Chunking with Transformers), the team varied wrist-camera and tactile availability while keeping the demonstrations and training identical. On three contact-rich tasks, adding wrist cameras improved success by 18.3 to 28.3 percentage points over vision without wrist images. Adding tactile sensing instead improved success by 31.7 to 60.0 points, and vision plus touch without wrist cameras achieved higher observed success than vision-only policies with wrist cameras on all three tasks.

The study is careful about scope: one robot platform, one tactile sensor type, one policy-learning method and one training seed, and low-cost tactile pads showed baseline variation after replacement. Its summary is worth quoting in spirit: vision guides the robot toward contact, while touch resolves interaction states that cameras cannot reliably observe.

York University: slip control without calibration or training data

Not every useful result needs machine learning. In a paper published in Frontiers in Robotics and AI on 9 February 2026, researchers at York University in Toronto described a slip-recovery controller for an anthropomorphic hand with a tri-axial piezoresistive sensor on each fingertip. Slip is detected from changes in tangential force against a baseline measured online, and only the finger that is slipping increases its grip until the slip stops, with motor-current protection to avoid damaging the actuator or the object. The method needs no per-object calibration, friction model, dense tactile array, line of sight or training dataset, and was tested on an aluminium tube, a plastic water bottle and a sponge. The authors position it for warehouse picking and human-robot collaboration.

Dexterous robot hands are shipping with touch built in

Research matters most when it reaches hardware that integrators can buy. Sharpa's Wave is a human-scale, five-finger hand with 22 active degrees of freedom. According to the company's product page, its Dynamic Tactile Array detects pressure, slip, force changes and contact location at the fingertip, with 6-dimensional force and torque sensing, a 0–30 N force detection range, 1 mm spatial resolution, up to 180 frames per second and 20 ms latency in the higher-end configuration. The hand integrates with ROS 2, can be simulated in MuJoCo and Isaac Sim, and is aimed at service, manufacturing and assembly, and research use. Sharpa also states that humanoid robots are still at an early stage of development globally.

On 10 November 2026 Google DeepMind introduced Gemini Robotics 2.5, its most capable vision-language-action model to date, evaluated on more than 200 tasks such as serving food with tongs, tying a bow on a gift box and pipetting a liquid. DeepMind reports a 30% to 100% higher success rate than previous baselines, depending on the task type, with tasks completed 20% to 30% faster. In a proof-of-concept on Mercedes-Benz production lines in Kecskemét and Sindelfingen, Apptronik's Apollo 2 humanoid, which DeepMind notes features the five-finger, 22-degree-of-freedom SharpaWave hand, performed automotive kitting with a 96% success rate. DeepMind does not break down how much of that result comes from tactile input, so it should not be read as a touch benchmark. It does show that dexterous, sensorised hands are now part of the platforms on which frontier robot models are tested.

Visuo-tactile manipulation: what technology leaders should take from this

I read four lessons from these results.

  1. Touch fixes the part vision cannot see. Three independent studies, using three different sensor technologies, found large gains on contact-rich tasks. The gains are largest exactly where cameras are occluded: insertion, retrieval, scooping and holding.
  2. Cost is no longer the main barrier. A full-hand skin under $10 in materials and a three-axis fingertip under $50 change the economics of instrumenting a robot cell. The harder costs are integration, calibration drift and maintenance.
  3. Data strategy matters more than sensor specs. SoTa's biggest gain came from letting humans wearing the same skin supply demonstrations. Matching sensors across humans and robots may prove as important as raw resolution.
  4. Simple control still has a place. The York result shows that a well-designed, calibration-free controller can deliver slip recovery without any training data. Not every tactile problem needs a foundation model.

The broader context is the physical AI stack. In Physical AI in 2026 I described the vision-language-action models that give robots their task intelligence, and in World models in 2026 the simulated worlds where those models are trained and tested; touch is still one of the hardest signals to simulate well, which is another reason real tactile data is valuable. In Humanoid robots in manufacturing 2026 I looked at factory pilots, where the difference between a dropped part and a completed shift often happens at the fingertip.

A practical playbook for tactile sensing in robot programmes

  1. Start with the failure log. List where your robots fail today. If failures cluster around insertion, slippage, occlusion or delicate objects, tactile sensing is a candidate; if they are about navigation or perception at a distance, it is not.
  2. Run a controlled comparison. Follow the Toyota Research Institute design: same demonstrations, same policy, with and without touch, with and without wrist cameras. Report success over many trials and out-of-distribution conditions, not a best clip.
  3. Budget for drift and replacement. Low-cost pads vary after replacement and magnetic sensors react to nearby metal. Plan normalisation, recalibration and spare parts as part of operations.
  4. Design the data pipeline for touch. Decide early whether human demonstrations will be collected with matching sensors, how tactile streams are time-aligned with video and actions, and how they are stored and labelled.
  5. Keep a non-learned safety layer. Grip-force limits, slip reflexes and motor-current protection, as in the York controller, should sit beneath any learned policy.
  6. Check the vendor's real numbers. For hands and sensors ask for durability cycles, latency, force range, behaviour after replacement and evidence from live sites, and treat lab success rates as hypotheses until they hold in your environment.

Closing thought

For most of robotics history, touch was the sense we promised to add later. The results of this autumn suggest that later has arrived for contact-rich work. Sensors are cheap enough, policies can use them, and the evidence that they improve real tasks is coming from several labs at once. The open questions are now operational: durability, calibration, data and integration. Those are exactly the questions technology leaders are good at asking. Data is the other half of the story; in Egocentric video for robot learning in 2026 I look at how first-person human video is becoming the bulk of robot training data, and why models like T-Rex still need touch on top.

References

Related articles

RoboticsPhysical AIRobot LearningTraining Data

Egocentric video for robot learning in 2026: how human video is becoming the training data behind robot foundation models

The scarcest input in robotics is not compute or hardware but data about how hands actually do work. In November 2026 NVIDIA's GR00T 2, Panasonic's T-Rex and the new HubEgo-100K dataset showed how far first-person human video can now carry a robot, building on earlier open datasets such as EgoVerse and World In Your Hands. What egocentric video for robot learning is, what the new evidence shows, where it breaks down and a practical playbook for technology leaders planning robot training data.

· 9 min read

Physical AIRoboticsVision-Language-ActionHumanoid Robots

Physical AI in 2026: what vision-language-action models mean for robots at work

In mid-October 2026, Google, NEURA Robotics and Galaxea each showed a piece of the same picture: robot foundation models that see, reason and act. What has really changed, what is still vendor-reported, and how a technology leader should evaluate physical AI.

· 8 min read

World ModelsPhysical AIRoboticsSimulation

World models in 2026: Genie 4, Cosmos 3 and what simulated worlds mean for robots and enterprise AI

Between 27 October and 9 November 2026 world models moved from research demos to tools that developers can download, stream and test against. Google DeepMind previewed Genie 4, NVIDIA released open Cosmos 3 simulation checkpoints, World Labs streamed its Atlas model in real time, LiberAI trained a world action model on 2.2 million hours of human experience, Apple ran a humanoid trained almost entirely in simulation and MIT unified six kinds of physics in one model. What world models can and cannot do yet, where the reality gap still bites, and a practical playbook for technology leaders.

· 11 min read