Dexterity

Rodney Brooks opens his dexterity essay with Heinrich Ernst's PhD work: by 1961, he writes, Ernst had connected a computer-controlled arm and hand to MIT's TX-0 and had it picking up and stacking blocks Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. The reason is contact. A walking or driving robot moves through free space and senses the world at a distance. A hand has to close the loop through the object: grip force, slip, and contact geometry decide success, and most of that state becomes invisible to cameras the moment the fingers close around it. This module covers contact-rich manipulationcontact-rich manipulationManipulation governed by making, holding, and breaking contact rather than by free-space motion: insertion, screwing, wiping, folding. Contact dynamics are discontinuous, since stick flips to slip across a friction boundary, so small sensing errors produce large outcome errors. That discontinuity is why these tasks are the stress test for touch-driven policies, and why TouchWorld's six-task tactile benchmark is built entirely from them. and the tactile gap it creates, the industry's split over whether touch is needed at all, and the two capability frontiers where that argument gets settled: in-hand manipulationin-hand manipulationRepositioning or reorienting a grasped object within the hand without setting it down: rolling a pen between the fingers, or walking a key around until it faces the lock. It is the skill that separates a hand from a gripper, and the one vision handles worst, because the fingers occlude exactly the contact that matters. Sanctuary demonstrated zero-shot in-hand reorientation with its 21-DoF hydraulic hands in 2024, and Holson's Robot Olympics makes it the gold-medal bar for tool use with the key-in-lock task. and deformable objects.

Five 2026 hands are compared below on the specs their makers disclose. Switch the sort from tactile threshold to cost and the disagreement is immediate: the two columns do not order the same way, and several cells are blank because the maker never published the number.

The tactile gap

The human hand is a sensor first and a mechanism second. A review of Roland Johansson's work reports about 17,000 low-threshold mechanoreceptorsmechanoreceptorA sensory nerve ending that fires under mechanical stimuli such as pressure, vibration, or skin stretch. The glabrous skin of the human hand carries about 17,000 low-threshold mechanoreceptors, roughly 1,000 of them packed at each fingertip, in four varieties that trade spatial localization against temporal response. Robot tactile sensors are closing on raw force threshold but remain far sparser: the best documented arrays put seven sensing cells on a fingerpad where the human finger packs about a thousand receptors. in the glabrous skin of the hand, roughly 1,000 of them at each fingertip Macefield 2022↓Why is our sense of touch so good at our fingertips?Vaughan G. Macefield, The Journal of Physiology, 2022. Brooks uses two videos from Roland Johansson's lab to make his case for touch. He describes the first match-lighting attempt as taking seven seconds and the second, after the same person's fingertips were anesthetized, as taking four times as long, with difficulty picking up and orienting the match. He says she could still sense other things in the rest of her fingers and hand and all the forces she ordinarily felt with her skeletal muscle system Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. These are the essay's descriptions of the demonstration, not an independent timing measurement or a clinical finding that all proprioception remained intact.

In his September 8, 2025 post, Benjie Holson described limitations of the learning-from-demonstration setups he was seeing, explicitly calling them a general trend with exceptions. He pointed to the lack of good standard ways to relay wrist-force information to the human teleoperator, difficulty controlling fingers beyond open and close, and the difficulty of making human-like touch sensing available to the puppeteer. His estimate of about 1 to 3 cm of task precision was a guess from videos; he said it was likely more a teleoperation limitation than a model limitation and pointed to a video he described as showing sub-centimeter tasks Holson 2025↓Benjie's Humanoid Olympic GamesBenjie Holson, 2025. Luo and colleagues define tactile robotics as developing and integrating tactile-sensing technologies into robotic systems. Their 2025 outlook discusses challenges across sensor materials, networks, simulation, benchmarking, data interpretation, multimodal learning, and active touch Luo 2025↓Tactile Robotics: An OutlookShan Luo, Nathan F. Lepora, Wenzhen Yuan et al., IEEE Transactions on Robotics (accepted), 2025. Brooks's conclusion is qualified and time-bound: "It looks like humanoid robots will need a sense of touch, and a level of touch sensing that no one has yet built in the lab" for tasks such as the match demonstration Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. This is his assessment at the time of the essay, not a current survey of every tactile sensor. The classical control stack is not silent on this question: the impedance-control half of the control module exists precisely because commanding position or force alone cannot govern a contact, and a tactile channel is what lets such a controller observe the contact it is shaping.

The bet against touch

Brooks frames the scaling counterargument as an imagined inner dialogue: end-to-end learning succeeded in speech-to-text, image labeling, and language models, so collecting examples of human hand use might yield dexterous control. His reply is that those earlier successes depended on engineered inputs, and that success at dexterity learning will likely require the right sensory data and the right thing to learn Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. Brooks reproduces an eWeek report that describes Tesla as moving Optimus training toward a "vision-only approach" instead of motion capture suits and teleoperation. The quoted report says workers wear helmet-and-backpack rigs with five in-house cameras, record tasks such as folding a t-shirt or picking up an object, and use the videos to train Optimus to mimic those actions Brooks 2025↓Why Today's Humanoids Won't Learn DexterityRodney Brooks, 2025. This is the report as reproduced in Brooks's essay, not a verified account of Tesla's complete training pipeline or evidence of tested dexterity. Figure's September 2025 Project Go-Big announcement describes a pretraining data-collection initiative rather than a general-dexterity result: its initial human-video result is navigation, with Helix trained on 100% egocentric human video, no robot demonstrations for that approach, mapping images and language to low-level SE(2) velocity commands, a transfer Figure calls zero-shot and, "to our knowledge", a first. Brookfield collection in its residential environments had begun and would expand, and the announcement's more than 100,000 residential units describe Brookfield's portfolio, not homes or trajectories collected Figure AI 2025↓Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot TransferFigure AI, 2025.

The case for the bet is real. Video of humans manipulating objects is the only manipulation data that exists at internet scale, teleoperationteleoperationDirect remote control of a robot by a human operator, used in robot learning mostly to record expert demonstrations. The ALOHA rig made the setup cheap and precise: the operator back-drives a leader arm that is a joint-for-joint twin of the follower the policy will run on, so the recorded motion is already in the robot's own action space. Mobile ALOHA extended the same leader-follower scheme to a mobile base for whole-body tasks. data is expensive and carries no force channel either, and every sensor added to a hand is cost, weight, wiring, and a new failure mode in a product that needs to be manufactured. If the bet pays off, the winners skip the tactile hardware problem entirely.

The bet on touch

The other side argues the bet cannot pay off, because the information is not in the video. Jeremy Fishel, principal researcher at Sanctuary AI, puts it in one sentence: with video alone you do not know you have touched something until well after the collision has physically moved the object Sanctuary AI 2025↓Sanctuary AI Equips General Purpose Robots with New Touch Sensors for Performing Highly Dexterous TasksSanctuary AI, 2025. Sanctuary's CEO James Wells calls touch a key enabler for creating human-level dexterity in robots, and the company built its Phoenix hands around that position: hydraulic actuation, and fingerpad arrays of micro-barometer cells sensitive to about five millinewtons, against roughly three for a human fingertip Sanctuary AI 2025↓Sanctuary AI Equips General Purpose Robots with New Touch Sensors for Performing Highly Dexterous TasksSanctuary AI, 2025RoboZaps 2026↓Sanctuary AI Phoenix 2026: Price, Is It For Sale & What It IsRoboZaps, 2026.

The hardware is arriving regardless of which thesis wins. Figure's October 2025 announcement describes a palm camera in each Figure 03 hand and says each fingertip sensor can detect "three grams of pressure", comparing the sensitivity to a paperclip's weight Figure AI 2025↓Introducing Figure 03Figure AI, 2025. Its January 2026 Helix 02 announcement lists head cameras, palm cameras, fingertip tactile sensors, and full-body proprioception as System 1 inputs. Figure describes this as "the first time we've demonstrated neural network policies that depend on these modalities" Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026. Shadow Robot's Dexterous Hand, the two-decade research benchmark, runs more than 100 sensors at 1 kHz Robot 2026↓Shadow Dexterous Hand SeriesShadow Robot, 2026. On the research flank, tactile foundation models now exist: Sparsh-X pretrained on about a million contact-rich interactions from Meta's Digit 360 sensor and lifted policy success rates by 63% over an end-to-end model using tactile images Higuera 2025↓Tactile Beyond Pixels: Multisensory Touch Representations for Robot ManipulationCarolina Higuera, Akash Sharma, Taosha Fan et al., 2025, and TouchWorld couples a tactile world model with a fast tactile refinement policy to reach 65% success on six long-horizon contact-rich tasks, 15.7 points over the strongest baseline Zhou 2026↓TouchWorld: A Predictive and Reactive Tactile Foundation Model for Dexterous ManipulationJianyi Zhou, Feiyang Hong, Yunhao Li et al., 2026.

The honest 2026 position is intermediate, and both camps occupy it. Tactile sensors ship on flagship hands, and even Figure, the loudest vision-data bet, now runs its newest policy on touch. What does not exist yet is a tactile training pipeline at anything like vision scale: the largest public tactile pretraining effort is built on about a million interactions, where vision pipelines consume billions of images. The results show the distance left. Google's Gemini Robotics 2 drives the 22-DoF SharpaWave hand, itself a tactile design Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026RoboZaps 2026↓Unitree H2 Humanoid Robot: $29,900 Price, Specs, vs H1 (2026)RoboZaps, 2026, and reports 92% success unscrewing a light bulb but 32 to 44% on its other multi-finger tasks, a spread DeepMind itself flags as the hard remainder Parada 2026↓Gemini Robotics 2 brings whole body intelligence to robotsCarolina Parada, 2026. The question of whether dexterity needs touch will be settled less by argument than by which training pipeline reaches deployment-grade reliability first.

Five hands, five bets

The hands on offer in 2026 span two orders of magnitude in price and no agreement on how dexterity will be learned. The table at the top of this module compares five of them on the specs their makers disclose, with the blanks left blank: a spec the maker has not published renders as "not disclosed" rather than as a guessed number.

The spread is the story. Tesla's 22-degree-of-freedomdegrees of freedomThe number of independent coordinates needed to specify a mechanism's configuration. An arm's degree-of-freedom count is its number of independent joints, so a 7-DoF arm places its end effector with one coordinate to spare beyond the six a rigid pose needs, and that redundancy is what lets the elbow reconfigure while the hand stays put. More degrees of freedom buy dexterity and obstacle avoidance at the price of a larger control problem. tendon design has the most articulation here and an admitted redesign behind it: days after the V3 hand patents surfaced, Musk said of the design, "this one didn't actually work" Klender 2026↓Elon Musk reveals shocking Tesla Optimus patent detailJoey Klender, Teslarati, 2026 Torres 2026↓The Forearm Is the New Hand: Inside Tesla's Optimus V3 PatentsDiana Wolf Torres, Alexander W. Torres, 2026. Sanctuary's hydraulic hand pairs the most sensitive disclosed touch with a working in-hand demo, and its maker has since pivoted to selling software RoboZaps 2026↓Sanctuary AI Phoenix 2026: Price, Is It For Sale & What It IsRoboZaps, 2026. Shadow's hand remains the research benchmark at a lab price of €110,000 including support Robot 2022↓How Much does a Robot Hand Cost?Shadow Robot, 2022. Unitree lists the entire H2 robot at $29,900, but the base model ships with non-functional placeholder hands, and its tactile option, the H2 Plus with Sharpa Wave hands, lists at $100,000 RoboZaps 2026↓Unitree H2 Humanoid Robot: $29,900 Price, Specs, vs H1 (2026)RoboZaps, 2026. Figure holds the middle: its August 6, 2024 release for Figure 02, Figure's own announcement hosted by PRNewswire, describes the second-generation robot's "4th generation hands" as "equipped with 16 degrees of freedom" without saying whether that count is per hand or combined, or giving an actuator count Figure AI 2024↓Figure unveils Figure 02, its second-generation humanoid, setting new standards in AI and roboticsFigure AI, 2024, and on Figure 03 a 3-gram fingertip threshold, the only maker in this table that puts the number on its own product page.

In-hand manipulation

In-hand manipulationin-hand manipulationRepositioning or reorienting a grasped object within the hand without setting it down: rolling a pen between the fingers, or walking a key around until it faces the lock. It is the skill that separates a hand from a gripper, and the one vision handles worst, because the fingers occlude exactly the contact that matters. Sanctuary demonstrated zero-shot in-hand reorientation with its 21-DoF hydraulic hands in 2024, and Holson's Robot Olympics makes it the gold-medal bar for tool use with the key-in-lock task. is the skill that separates a hand from a gripper: reorienting a held object without setting it down, the way you walk a key around between your fingers until it faces the lock. In the gold-medal "Use a key" event of Holson's Humanoid Olympics, a keyring with at least two keys and a keychain is dropped into the robot's waiting palm or gripper. Without putting the keys down, the robot must align, insert, and turn the correct key in a lock Holson 2025↓Benjie's Humanoid Olympic GamesBenjie Holson, 2025. These are challenge rules, not a report that a robot completed the task.

Progress here is real but narrow. Sanctuary announced in-hand manipulation with its 21-DoF hydraulic hands in December 2024, reorienting held objects without setting them down Sanctuary AI 2024↓Sanctuary AI Demonstrates In-Hand Manipulation Capabilities for Improved General Purpose Robot DexteritySanctuary AI, 2024 RoboZaps 2026↓Sanctuary AI Phoenix 2026: Price, Is It For Sale & What It IsRoboZaps, 2026, then showed a sim-trained reinforcement-learning policy reorienting objects against gravity with a 500 g weight added, a vendor-run result the company published itself Sanctuary AI 2025↓Sanctuary AI Leads the Industry in Controlling Advanced Hydraulic Hands Using Reinforcement LearningSanctuary AI, 2025. Figure says Helix 02, using Figure 03's fingertip sensors and palm cameras, performs four tasks it describes as manipulation previously out of reach for its stack, in videos it calls fully autonomous rather than teleoperated: unscrewing a bottle cap, extracting a single pill from an organizer, a task titled "Push exactly 5 ml from a syringe", and picking small metal parts from clutter. The announcement publishes no task-level success rates, volume calibration, or sensor-ablation results Figure AI 2026↓Introducing Helix 02: Full-Body AutonomyFigure AI, 2026. Shadow's hand was designed for exactly this: each finger has independent side-to-side motion for in-hand work Robot 2026↓Shadow Dexterous Hand SeriesShadow Robot, 2026. None of these is the general skill. Each is a task or a task family with its own training run.

Deformables and the long tail

Rigid objects are the easy case. A wooden block rests on one of six sides; a towel can be bunched up in more ways than any state estimator will ever enumerate, and that is before the task involves water, grease, or peanut butter Holson 2025↓Benjie's Humanoid Olympic GamesBenjie Holson, 2025. Cloth, liquids, and compliant packaging dominate actual housework, which is why Holson's task list is full of them: hang a dress shirt that starts with one sleeve inside out, clean peanut butter off your own manipulator.

The measured record is thin but no longer empty. RL-100 folds cloth and ran a juicing robot in a shopping mall for about seven hours without failure, both per-task results earned through on-robot reinforcement learning Lei 2025↓RL-100: Performant Robotic Manipulation with Real-World Reinforcement LearningKun Lei, Huanyu Li, Dongjie Yu et al., 2025. Physical Intelligence's π0.7 does laundry and espresso tasks with language steering Ai 2026↓π0.7: a Steerable Generalist Robotic Foundation Model with Emergent CapabilitiesBo Ai, Ali Amin, Raichelle Aniceto et al., 2026. The most informative datapoint is the Olympics scoreboard itself: Physical Intelligence fine-tuned its π0.6-based model on Holson's tasks and took gold in three of five categories with under nine hours of data for most tasks, but averaged 52% success and 72% task progress, while a baseline VLM without robot pretraining managed 9% progress Physical Intelligence 2025↓Moravec's Paradox and the Robot OlympicsPhysical Intelligence, 2025. Two gold tasks were physically impossible for their gripper; peeling an orange needed a tool and did not count Physical Intelligence 2025↓Moravec's Paradox and the Robot OlympicsPhysical Intelligence, 2025. Read the number both ways: half-success on tasks chosen by a critic is real progress, and half-success is not a product.

What solved would look like

Dexterity is solved when a single policy does Holson's fifteen tasks in unseen homes at human-competent speed, with no per-task data collection, and with the success rates published. Nobody is close: the best public scoreboard stands at 52% with per-task training, and the strongest multi-finger results from the largest labs sit between 32% and 92% per task. The open question is which data closes the gap. If Tesla or Figure reach household dexterity from video at scale, the tactile gap was a transitional hardware shortage. The hands are shipping either way; watch the training pipelines.