Tactile sensing

Tactile sensing is the technology that gives robots a sense of touch: the ability to measure the physical contact between a robot and the world, including contact force, pressure distribution, shear, slip, vibration, texture, hardness, temperature, and proximity. It is widely described as the "missing modality" of embodied AI, the sense that complements vision and language and that turns a dexterous hand from a mechanically capable gripper into a system that can actually feel what it is doing. Without touch a robot must infer contact indirectly from cameras and joint encoders; with touch it can grasp a fragile object, detect a slip before it happens, and manipulate in the dark or inside a cluttered drawer where vision is occluded ^[1][2].

Tactile sensing sits at the center of modern robot manipulation research because manipulation is fundamentally about contact. Vision tells a robot where an object is; touch tells it what is happening at the moment the fingers close. This article covers what tactile sensing measures, the key specifications that define a sensor, the main transduction technologies and their trade-offs, the rise of high-resolution vision-based touch, fingertip sensor arrays for dexterous hands, the supplier landscape, and how tactile data is now feeding machine-learning systems from imitation learning policies to multimodal vision-language-action models.

What tactile sensing measures

Human skin is the reference design. Glabrous (hairless) skin on the human fingertip contains four main mechanoreceptor types: Merkel discs (slow-adapting, sensitive to sustained pressure and fine edges), Meissner corpuscles (fast-adapting, sensitive to light touch and low-frequency flutter), Ruffini endings (sensitive to skin stretch and shear), and Pacinian corpuscles (fast-adapting, tuned to high-frequency vibration around 250 to 350 Hz, the cues that signal texture and incipient slip) ^[3]. A robot tactile system tries to reproduce some or all of these channels. In practice a tactile sensor may report any combination of the following ^[1][4]:

• Normal force and pressure distribution: how hard, and where, contact is happening across a sensing patch.
• Shear force: lateral (tangential) force, the cue that reveals whether a grasped object is about to slide.
• Slip and micro-vibration: the small, fast vibrations produced when a surface begins to move against the skin, and when sliding across texture.
• Contact geometry: the local 3D shape of the object pressing into the sensor (edges, corners, curvature, surface relief).
• Texture and hardness: fine surface structure and material compliance, inferred from pressure patterns and vibration spectra.
• Temperature and heat flux: thermal contact, useful for material identification and safety.
• Proximity or pre-touch: some sensors detect an object just before contact.

The distinguishing feature of touch compared with vision is that it is a contact sense with essentially no occlusion problem at the point of contact, high temporal bandwidth, and direct access to force. That combination is exactly what dexterous, contact-rich manipulation needs.

Why touch is the missing modality for embodied AI

The last decade of robot learning has been dominated by vision and, more recently, language. Large foundation models trained on internet-scale images and text gave robots strong perception and instruction-following, but their success has been concentrated in visually guided tasks. Contact-rich tasks such as inserting a plug, threading a needle, buttoning a shirt, or turning a key in a lock routinely fail when the policy has only cameras to rely on, because the decisive information (are the mating surfaces aligned, is the object slipping, how much force is on the part) is simply not visible ^[5][6]. Researchers building physical AI systems increasingly frame tactile sensing as the modality that closes this gap, giving learned controllers the same close-loop force feedback that humans use without conscious thought.

Survey work traces robotic touch through several eras: origins in the 1960s and 1970s, a period of foundational work in the 1980s and early 1990s, a relatively quiet "tactile winter" from the mid-1990s to around 2009 when progress in vision outpaced touch, and a strong resurgence from about 2010 to the present driven by soft and biomimetic materials, low-cost cameras, and machine learning ^[1]. The current phase is defined by several converging themes: large-area electronic skin, tactile-instrumented robot hands, high-resolution vision-based touch, soft and biomimetic sensors, and the integration of touch into learning systems.

Key quantities and specifications

Comparing tactile sensors requires a shared vocabulary of specifications. The most important are the following.

Scroll sideways for more →

For context, the human fingertip can detect forces on the order of 3 millinewtons and resolve two contact points a few millimeters apart ^[3][12]. Matching that sensitivity, resolution, and robustness at once, on a fingertip small enough to fit a dexterous hand and durable enough to survive real work, is the central engineering challenge of the field.

Transduction technologies

There is no single best way to build a tactile sensor; each transduction principle trades resolution, bandwidth, durability, cost, and manufacturability differently. The main families are compared below. Specifications are representative examples, not universal, and vendor claims are attributed where relevant ^[4][7][13].

Scroll sideways for more →

Beyond individual fingertips, the same principles scale up to large-area electronic skin (e-skin): flexible, stretchable sensor sheets meant to cover arms, torsos, or whole robots. Research e-skins have demonstrated large coverage areas with super-resolution readout, self-powered triboelectric arrays for collision safety, and origami-inspired capacitive skins covering tens of thousands of square millimeters, though combining high resolution, large area, and low cost on curved surfaces remains difficult ^[14][15].

Vision-based tactile sensing

The most consequential recent shift in robotic touch has been the rise of camera-based (optical) tactile sensors, which trade the low bandwidth and modest resolution of electrode arrays for extremely high spatial detail.

The lineage begins with GelSight, invented at MIT by Edward Adelson and postdoc Micah Kimo Johnson around 2009 under the name "retrographic sensing." A GelSight sensor is a slab of clear elastomer coated on the outside with an opaque reflective membrane; a camera and colored LEDs sit behind the gel. When the coated gel presses against an object, the membrane takes on the object's exact shape, and photometric stereo reconstructs the 3D surface at resolution far finer than the human fingertip, on the order of tens of micrometers ^[8][16]. Painting small markers on the gel lets the camera also recover shear and force from how the markers move. GelSight, Inc., based in Waltham, Massachusetts, commercialized the technology first for industrial inspection ^[16][17]. Later variants optimized the idea for robot fingers: GelSlim made the package thin and compact, and the GelSight Wedge achieved high-resolution 3D reconstruction in a finger-shaped form factor ^[18][19].

Two developments made high-resolution optical touch central to machine-learning manipulation. First, DIGIT, released and open-sourced by Meta (then Facebook) AI in 2020, miniaturized the GelSight concept into a low-cost, compact, mass-manufacturable fingertip sensor that could actually be mounted on a multi-fingered hand, and published the designs so any lab could build one ^[2][20]. Meta later partnered with GelSight to manufacture DIGIT commercially ^[21]. Second, ReSkin, released by Meta AI with Carnegie Mellon University in 2021, took a different route: a 2 to 3 mm stretchable elastomer with embedded magnetic microparticles, read by a magnetometer, that costs under 6 US dollars per unit at modest volume and can be peeled off and replaced when it wears out, with a self-supervised algorithm to auto-calibrate each new skin ^[22][23]. Because both DIGIT and ReSkin were cheap and open, they became de facto research standards, and their image-like or field-like outputs plug naturally into the convolutional and transformer networks that dominate deep learning.

Meta pushed the agenda further on 31 October 2024 with a trio of releases: Digit 360, a fingertip sensor that the company says packs more than 8 million taxels (effectively camera pixels) and over 18 sensing features including vibration, heat, and even odor, with sensitivity to forces as small as 1 millinewton and an on-device AI accelerator for reflex-speed processing; Sparsh, described as a general-purpose touch "foundation model" pre-trained with self-supervision on more than 460,000 tactile images that Meta reports outperforms task-specific models by an average of over 95 percent under limited labeled data; and Digit Plexus, a platform that routes many fingertip and skin sensors over a single cable ^[13][24][25]. Meta announced GelSight would manufacture Digit 360 and that Wonik Robotics would build a tactile version of the Allegro Hand on Digit Plexus ^[13][24].

Fingertip sensor arrays for dexterous hands

For a dexterous hand to manipulate like a human, touch must cover the fingertips, the finger pads, and ideally the palm, and it must report more than a single pressure number per contact. Modern hands therefore integrate dense arrays of multi-axis taxels.

Taxel counts have climbed quickly. The research humanoid iCub carried roughly 108 tactile elements per hand (12 per fingertip plus 48 in the palm) ^[26]. Commercial hands from Inspire Robotics (inspire_robotics) now advertise far denser coverage: the RH56 series is offered with resistive or capacitive fingertip arrays, and the compact RH56H1 is listed with about 260 sensing points (roughly 170 across the fingertips plus 90 in the palm), according to distributor specifications ^[27]. At the high end, PaXini's DexH13 hand is described as carrying roughly 1,140 of what the company calls intelligent tactile processing units, its multidimensional Hall-effect sensing cells, a figure reported from the company's CES 2026 presentation and best treated as a vendor specification rather than an independently verified count ^[28]. PaXini similarly cites very fine repeat positioning accuracy for its hardware; such sub-0.1 mm figures are plausible for precision assemblies but should be read as manufacturer claims. Because dense wiring is costly, a parallel research direction uses deep learning to synthesize "super-resolution" tactile maps, in one case generating 2,700 virtual taxels from just 23 physical sensing units, which suggests future hands may not need to place a physical taxel at every point they can resolve ^[40].

A complementary layer is joint torque and six-axis force/torque sensing. Rather than measuring skin pressure, these units sit at the wrist and finger joints and report the full force and torque vector (three forces and three torques). Vendor datasheets for humanoid wrist and hand sensors describe representative ranges on the order of tens of newton-meters of torque with full-scale accuracies quoted around plus or minus 0.5 percent, and force resolutions of roughly 0.1 to 1 N ^[10][11]. Tesla's Optimus, for example, uses six-axis force/torque sensors in its wrists and ankles for force-controlled motion ^[11]. Together, skin-level taxel arrays and joint-level force/torque sensing give a hand both fine contact detail and whole-hand force awareness, mirroring the way humans combine cutaneous touch with proprioceptive force sense.

Suppliers and the tactile-sensor industry

Tactile sensing has shifted from a niche research supply chain to a competitive industry, propelled by the humanoid robot boom.

The most prominent pure-play supplier is PaXini Technology, founded in Shenzhen in 2021 by researchers from Waseda University's robotics lab. The company's name is taken from the Pacinian corpuscle, the human vibration receptor. PaXini says it developed the world's first 6D Hall-array multidimensional tactile sensor, able to report six-axis force plus properties such as texture, elasticity, and temperature, and it is widely described as China's largest tactile-sensor maker ^[29][30]. Its most-cited achievement is cost: PaXini states that comparable multidimensional tactile sensors cost as much as 100,000 yuan five or six years ago, while its basic models now sell for as little as 199 yuan, which it frames as roughly 0.2 percent of overseas equivalents ^[30]. Company representatives have gone further, with a data-collection director claiming that around 80 percent of humanoid robots worldwide use PaXini sensors, a figure that is a company assertion and hard to verify independently ^[30]. PaXini also builds full robots, including the TORA-ONE humanoid, and by 2026 had raised over 1 billion yuan (about 145 million US dollars) in a Series B that pushed its valuation above 10 billion yuan, with a strategic investment of more than 100 million yuan from BYD ^[31][32]. Other Chinese suppliers named in the same industry chain include Hanwei and Keli Sensing, alongside a broad field of force/torque-sensor makers now targeting robot hands.

On the open-hardware and academic side, Meta has been the dominant force through DIGIT, ReSkin, Digit 360, and the Sparsh model, with manufacturing partners GelSight and Wonik Robotics extending that ecosystem into products ^[13][24]. Legacy research hardware includes SynTouch's BioTac, a biomimetic fluid-filled fingertip that sensed force, micro-vibration, and temperature and became a research standard, though its commercial availability has reportedly narrowed in recent years ^[9][7], and the University of Bristol's TacTip, an open, 3D-printed optical sensor whose internal white pins mimic the human skin's intermediate ridges and achieve sub-millimeter super-resolution ^[33].

Market sizing should be read with caution because published estimates vary widely by definition and methodology. One market-research house (NextMSC) puts China's tactile-sensor market at about 619 million US dollars in 2023, rising toward 1.39 billion by 2030 at roughly 11 percent compound annual growth ^[34]. A separate, widely circulated Chinese industry-chain figure cites a China tactile-sensor market on the order of 1.5 billion yuan for 2026, but that number appears in single-source commentary and is not corroborated by named research reports, so it is best treated as indicative rather than authoritative. Analysts do broadly agree that force/torque and tactile sensing is among the fastest-growing robot-sensor categories, driven by collaborative robots and humanoids.

Tactile sensing in AI and robot learning

Touch data has become a first-class input to robot learning. Several distinct uses have matured:

• Slip detection: because touch directly senses shear and micro-vibration, it can catch an object starting to slide faster and more reliably than vision. Modern approaches treat slip as a binary classification problem and use CNN-LSTM networks that combine spatial and temporal tactile features to trigger a corrective grip ^[35].
• In-hand manipulation: rotating, re-orienting, and re-grasping an object within the hand is a canonical contact-rich task where touch supplies the state that cameras cannot see. Reinforcement learning policies now use tactile feedback to rotate objects and insert connectors, and imitation learning from teleoperated demonstrations increasingly records touch alongside video ^[5][35].
• Multimodal policies and VLA models: a wave of 2025 and 2026 research fuses touch into vision-language-action models. Systems such as Tactile-VLA, OmniVTLA, and force-aware VLA variants argue that vision-and-language policies fail on contact-rich tasks precisely because they lack tactile awareness, and show that adding a semantically aligned touch stream improves insertion, deformable-object handling, and force-sensitive manipulation ^[5][6].
• Simulation and pretraining: touch simulators (for example TACTO) and touch foundation models (Sparsh) let policies be pretrained or trained partly in simulation, addressing the scarcity of labeled tactile data ^[24].

These techniques are showing up in commercial humanoids. Sanctuary AI integrated a new generation of touch sensors into its Phoenix hands in February 2025, putting a seven-cell micro-barometer array in each finger pad that detects forces as low as 5 millinewtons, close to the roughly 3 millinewton sensitivity of a human fingertip, and enabling blind picking, slip detection, and force limiting when vision is blocked ^[36][37]. Figure AI took a similar view with Figure 03 in 2025, which the company says carries in-house fingertip tactile sensors able to feel forces as small as 3 grams (about the weight of a paperclip), paired with a wide-angle palm camera, so its Helix control system can distinguish a secure grasp from an incipient slip ^[38][39]. Both companies emphasize that touch is what lets a learned policy act confidently when the camera view is occluded, which is most of the time during real manipulation.

Applications and open challenges

Beyond dexterous manipulation, tactile sensing supports grasping and grip-force control, texture and material recognition, object identification by feel, surface following and assembly, safe human-robot contact, prosthetics and wearables, and quality-inspection tools. As robots move from structured factories into homes and unstructured environments, the ability to work by feel becomes more valuable, not less.

Several hard problems remain open:

• Durability and consumables: the highest-resolution sensors (vision-based gels) wear out in hundreds to a few thousand contacts, while the most durable arrays give up spatial detail. Building a fingertip that is both high-fidelity and able to survive continuous real-world work is a leading unsolved trade-off, and one Figure and Sanctuary both cite as a reason they built sensors in-house ^[38][36].
• Wiring and routing density: putting hundreds of taxels on every fingertip and across a palm creates a wiring nightmare. Platforms like Digit Plexus, magnetic sensors like ReSkin that avoid fragile in-skin wiring, and multiplexed array architectures are all attempts to manage routing density ^[24][22].
• Sim-to-real for touch: simulating contact physics and the exact response of a soft, marked gel or a compliant electrode is far harder than rendering images, so sim-to-real transfer of tactile policies lags behind vision.
• Lack of standard benchmarks: the field has many incompatible sensors and few shared datasets, tasks, or metrics, which makes results hard to compare. Touch foundation models and cross-sensor datasets are early attempts to standardize, but there is nothing yet like ImageNet for touch ^[1][24].

ELI5

Imagine trying to tie your shoelaces or pick up a single grape while wearing thick oven mitts and looking through a foggy window. You would fumble, squeeze too hard, and drop things, because you cannot feel what your fingers are doing. That is what a robot is like when it only has cameras. Tactile sensing is the robot's sense of touch: tiny sensors in its fingertips that feel how hard it is pressing, whether something is sliding, and what the surface feels like. Some of these sensors are actually tiny cameras looking at a soft rubber pad from the inside, so when the pad squishes against an object, the camera "sees" the exact shape it is touching. With touch, a robot can hold an egg without crushing it, feel when a cup starts to slip, and work in the dark or inside a bag where its eyes cannot see. Touch is the sense that lets robots be gentle and careful, which is why researchers call it the missing piece for robots that work in the real world.

See also

• Dexterous hand
• Robot manipulation
• Embodied AI
• Robot learning
• Vision-language-action model
• PaXini Technology
• Inspire Robotics
• Sanctuary AI Phoenix
• Figure 03
• Allegro Hand
• Sim-to-real
• Physical AI

References

1. "Tactile Robotics: Past and Future." arXiv:2512.01106, 2025. https://arxiv.org/abs/2512.01106 ↩
2. Lambeta, M. et al. "DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor with Application to In-Hand Manipulation." IEEE Robotics and Automation Letters, 2020. arXiv:2005.14679. https://arxiv.org/abs/2005.14679 ↩
3. "Peripheral mechanosensory receptors: Merkel cells, Ruffini endings, Meissner and Pacinian corpuscles." Kenhub. https://www.kenhub.com/en/library/physiology/peripheral-mechanosensory-receptors ↩
4. "Novel Tactile Sensor Technology and Smart Tactile Sensing Systems: A Review." Sensors (MDPI), 2017. https://pmc.ncbi.nlm.nih.gov/articles/PMC5713637/ ↩
5. "Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization." arXiv:2507.09160, 2025. https://arxiv.org/abs/2507.09160 ↩
6. "OmniVTLA: Vision-Tactile-Language-Action Model with Semantic-Aligned Tactile Sensing." arXiv:2508.08706, 2025. https://arxiv.org/abs/2508.08706 ↩
7. "Tactile Sensors for Robots: GelSight vs DIGIT vs PaXini." Stanford Vision and Robotics Center (roboticscenter.ai). https://www.roboticscenter.ai/learn/tactile-sensor-comparison ↩
8. Yuan, W., Dong, S., Adelson, E. "GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force." Sensors, 2017. https://pmc.ncbi.nlm.nih.gov/articles/PMC5751610/ ↩
9. "BioTac and BioTac SP: Sensory Modalities." SynTouch. https://www.syntouchllc.com/Products/BioTac/Sensory-Modalities/ ↩
10. "Six-axis Force Torque Sensors: The Force Sensing Technology for Humanoid Robots." Hypersen Technologies. https://en.hypersen.com/application/six-axis-force-torque-sensors-the-force-sensing-technology-for-humanoid-robots/ ↩
11. "6-Axis Force/Torque Sensors: Accelerating the Development of Humanoid Robots." XJC Sensor. https://www.xjcsensor.com/Blog/6-axis-forcetorque-sensors-accelerating-the-development-of-humanoid-robots ↩
12. "Sanctuary AI equips its Phoenix general purpose robots with new touch sensors." Sanctuary AI, February 2025. https://www.sanctuary.ai/blog/sanctuary-ai-equips-general-purpose-robots ↩
13. "Advancing embodied AI through progress in touch perception, dexterity, and human-robot interaction." Meta AI (FAIR), 31 October 2024. https://ai.meta.com/blog/fair-robotics-open-source/ ↩
14. "Large-Area High-Resolution Skin-Inspired Flexible Tactile Sensor for Robotic Electronic Skin." ACS Applied Electronic Materials, 2025. https://pubs.acs.org/doi/10.1021/acsaelm.5c01200 ↩
15. "Large-Scale and Flexible Self-Powered Triboelectric Tactile Sensing Array for Sensitive Robot Skin." ACS Nano / PMC, 2019. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6418621/ ↩
16. "A Helping Hand: MIT's entrepreneurial ecosystem boosts sensor tech startup." GelSight / MIT Spectrum, 2019. https://www.gelsight.com/a-helping-hand-mits-entrepreneurial-ecosystem-boosts-sensor-tech-startup/ ↩
17. "Visuotactile Sensors With Emphasis on GelSight Sensor: A Review." IEEE Sensors Journal, 2020. https://www.researchgate.net/publication/339818841 ↩
18. Donlon, E. et al. "GelSlim: A High-Resolution, Compact, Robust, and Calibrated Tactile-sensing Finger." IROS 2018. arXiv:1803.00628. https://arxiv.org/abs/1803.00628 ↩
19. Wang, S. et al. "GelSight Wedge: Measuring High-Resolution 3D Contact Geometry with a Compact Robot Finger." ICRA 2021. arXiv:2106.08851. https://arxiv.org/abs/2106.08851 ↩
20. "DIGIT: A Novel Design for a Low-Cost Compact High-Resolution Tactile Sensor." Meta Research. https://ai.meta.com/research/publications/digit-a-novel-design-for-a-low-cost-compact-high-resolution-tactile-sensor-with-application-to-in-hand-manipulation/ ↩
21. "GelSight Partners with Meta AI to Democratize Digital Tactile Sensing." GelSight. https://www.gelsight.com/gelsight-partners-with-meta-ai-to-democratize-digital-tactile-sensing/ ↩
22. "ReSkin: a versatile, replaceable, low-cost skin for AI research on tactile perception." Meta AI, November 2021. https://ai.meta.com/blog/reskin-a-versatile-replaceable-low-cost-skin-for-ai-research-on-tactile-perception/ ↩
23. "ReSkin Could Help Researchers Discover a Sense of Touch." Carnegie Mellon University, November 2021. https://www.cmu.edu/news/stories/archives/2021/november/reskin.html ↩
24. "Meta unveils AI tools to give robots a human touch in physical world." VentureBeat, 31 October 2024. https://venturebeat.com/ai/meta-unveils-ai-tools-to-give-robots-a-human-touch-in-physical-world ↩
25. "Meta's new robotic fingertip mimics human-level multimodal sensing capabilities." The Decoder, 2024. https://the-decoder.com/metas-new-robotic-fingertip-mimics-human-level-multimodal-sensing-capabilities/ ↩
26. "Recent advances and challenges of tactile sensing for robotics: from fundamentals to applications." Cell Reports Physical Science / ScienceDirect, 2025. https://www.sciencedirect.com/science/article/abs/pii/S2542529325000963 ↩
27. "Inspire Robots RH56H1 Dexterous Hand: Ultra Compact 5-Finger Hand with 260 Tactile Sensors." Knox Labs. https://www.knoxlabs.com/products/inspire-robots-rh56h1-dexterous-hand ↩
28. "At CES 2026, PaXini lays out a full-stack roadmap for embodied intelligence." KrASIA, January 2026. https://kr-asia.com/at-ces-2026-paxini-lays-out-a-full-stack-roadmap-for-embodied-intelligence ↩
29. "Why Is the Taiwanese Robot Master from Waseda University Moving to China?" CommonWealth Magazine, 2024. https://english.cw.com.tw/article/article.action?id=3796 ↩
30. "Tianjin develops sensors for embodied AI." China Daily, 10 June 2026. https://www.chinadaily.com.cn/a/202606/10/WS6a28cfd0a310d6866eb4d6e9.html ↩
31. "Robot Sensor Startup PaXini Raises $145 Million, Tops $1.4 Billion Valuation." Caixin Global, 7 March 2026. https://www.caixinglobal.com/2026-03-07/robot-sensor-startup-paxini-raises-145-million-tops-14-billion-valuation-102420414.html ↩
32. "BYD Finalizes Largest Single Investment in Embodied Intelligence Field: Over 100 Million Yuan Targeting PaXini's Tactile Perception Technology." 36Kr. https://eu.36kr.com/en/p/3269173428654466 ↩
33. Ward-Cherrier, B. et al. "The TacTip Family: Soft Optical Tactile Sensors with 3D-Printed Biomimetic Morphologies." Soft Robotics, 2018. https://pmc.ncbi.nlm.nih.gov/articles/PMC5905869/ ↩
34. "China Tactile Sensors Market 2030: Growth and Robot Trends." Next Move Strategy Consulting (NextMSC). https://www.nextmsc.com/report/china-tactile-sensor-market ↩
35. "Universal slip detection of robotic hand with tactile sensing." Frontiers in Neurorobotics, 2025. https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2025.1478758/full ↩
36. "Sanctuary AI integrates tactile sensors into Phoenix general purpose robots." The Robot Report, 26 February 2025. https://www.therobotreport.com/sanctuary-ai-integrates-tactile-sensors-into-phoenix-general-purpose-robots/ ↩
37. "Sanctuary AI equips Phoenix humanoids with new touch sensors." Robotics 24/7, 2025. https://www.robotics247.com/article/sanctuary-ai-equips-phoenix-humanoids-with-new-touch-sensors ↩
38. "Introducing Figure 03." Figure AI, 2025. https://www.figure.ai/news/introducing-figure-03 ↩
39. "Figure AI designs Figure 03 humanoid for AI, home use, and scaling." The Robot Report, 2025. https://www.therobotreport.com/figure-ai-designs-figure-03-humanoid-ai-home-use-scaling/ ↩
40. "Super-resolution tactile sensor arrays with sparse units enabled by deep learning." Science Advances, 2025. https://www.science.org/doi/10.1126/sciadv.adv2124 ↩