Dexterous Robotic Hands in the Foundation-Model Era: A Full-Stack Survey of Hardware, Data, Cross-Embodiment Transfer, Learning, and Evaluation
Organizations: Xi’an Jiaotong University, Xi’an, China
Abstract
Dexterous robotic hands are increasingly coupled to large-scale human data and foundation models, yet progress in hardware, representation learning and evaluation is often assessed separately. This survey reviews work from January 2024 to early November 2026, with earlier studies included to establish technical lineages. We organize the literature into five interdependent layers: hardware; data collection and curation; cross-embodiment transfer; learning and foundation models; and evaluation and applications. Human hand-object understanding and tactile sensing provide two cross-cutting perspectives. A system-level architecture, comparative tables and five case analyses connect design choices across these layers. We distinguish within-study experimental comparisons from heterogeneous corpus statistics, vendor specifications and company-reported results. Three conclusions emerge. First, the value of human data depends on kinematic, visual and contact alignment, not recording volume alone. Second, shared action representations and foundation policies do not remove embodiment-specific constraints on force, latency and contact. Third, task completion is insufficient to establish physical correctness or deployment readiness. Evidence remains limited for zero-shot transfer of contact-rich skills between morphologically different hands, long-duration reliability, and cross-sensor tactile generalization. We identify eleven open problems and propose reporting interfaces and controlled comparisons that can make full-stack progress more measurable. The review is a structured narrative synthesis rather than a pooled performance ranking or an exhaustive systematic review.
Figures & tables
| Interface | Information needed for diagnosis and reuse |
| Capture transfer | Raw observations, hand/object state, calibration, timing and provenance |
| Transfer learning | Action semantics, masks, embodiment metadata and feasibility tests |
| Policy control | Target frames and units, chunk horizon, update rate and delay |
| Control evaluation | Outcome, force, precision, damage, interventions and latency |
| Evaluation design | Failure mode, hardware/sensor configuration, environment and recovery |
| Layer | 2024 | 2025 | 2026 (through 4 November) |
| Hardware | – | ORCA [ 41 ] ; Sharpa Wave [ 42 ] | Dex5-S [ 10 ] ; MIDAS [ 43 ] ; Aero Hand [ 12 ] |
| Data | DexCap [ 44 ] ; Open-TeleVision [ 45 ] | AgiBot World [ 46 ] ; DexWild [ 17 ] ; EgoDex [ 19 ] ; DEXOP [ 16 ] | EgoScale [ 20 ] ; RealDexUMI [ 18 ] ; TouchScale [ 47 ] |
| Transfer | CrossDex [ 48 ] | SPIDER [ 49 ] | TactAlign [ 50 ] ; UniDex [ 51 ] ; TopoRetarget [ 52 ] ; OmniHOI [ 53 ] |
| Models | [ 26 ] | Helix [ 33 ] ; GR00T N1 [ 29 ] ; [ 27 ] | DreamZero [ 31 ] ; [ 28 ] ; GEN-1.5 [ 23 ] |
| Evaluation | – | TeleOpBench [ 54 ] | DexJoCo [ 55 ] ; DexVerse [ 56 ] ; HandsOn [ 57 ] |
| Term | Meaning |
| Active / total DoF | Independently actuated inputs / all joints including coupled and passive ones |
| Backdrivability | Ability of external forces to move the joint through the transmission; eases compliant contact and sim-to-real |
| Retargeting | Mapping human hand motion (keypoints, joint angles) to robot joint commands |
| Embodiment gap | Mismatch in morphology, action semantics, observation, dynamics and timing between source and target |
| Robot-free capture | Data collection in which the human acts directly, with wearable hardware approximating the robot |
| Isomorphic device | Collection device with the same kinematics (and ideally sensors) as the deployed hand |
| Route | Principle | Examples | Strengths | Weaknesses | Typical use |
| Direct / quasi-direct | One micro-motor + reducer per joint, in finger or palm | Wuji Hand [ 68 ] , Revo 3 [ 69 ] , Sharpa Wave [ 42 ] , ADH-5-13 [ 70 ] , MIDAS [ 43 ] , Allegro [ 90 ] | Independent joints, low backlash, backdrivable, simple joint–motor map; good for RL and sim-to-real | Limited space and heat dissipation; many motors; modest finger force | Research, RL, data-collection platforms |
| Linkage | Linear actuators drive coupled linkages (e.g., 6 active / 12 total) | Linkage hands on early humanoids [ 11 ] , ByteDexter [ 71 ] | Stiff, strong, durable, cheap, easy to mass-produce | Coupled joints, weak abduction, limited in-hand dexterity | Industrial picking, logistics |
| Tendon | Forearm motors pull (sheathed) tendons | Shadow [ 5 ] , Optimus V3 [ 74 ] , ORCA [ 41 ] , RUKA [ 75 , 76 ] , MM-Hand [ 77 ] , CRAFT [ 13 ] | Light, low inertia, compliant, high DoF | Friction, creep, slack; tensioning and calibration; non-linear mapping | Anthropomorphic and humanoid hands |
| Hybrid | Tendons for flexion; direct drive/linkage for abduction and opposition | Xiaomi 21-DoF [ 84 ] , CYJ Hand-0 [ 85 ] | Combines force, low weight and precision per joint | Highest system complexity and maintenance cost | 2026 high-end products |
| Hydraulic | Miniature hydraulic valves | Sanctuary AI [ 87 ] | High power density, fine force resolution | Complexity, cost, leakage risk | Research platforms |
| Servo (open-source) | Off-the-shelf bus servos at the joints | LEAP [ 66 ] , Aero Hand Open [ 12 ] , AmazingHand [ 89 ] | Very cheap, easy to build and repair | Bulky, non-anthropomorphic proportions, limited force and life | Entry-level research, teaching |
| Hand (maker) | Year | Route | DoF | Mass | Force / payload | Sensing | Notes |
| RH56 series (Inspire) [ 91 , 92 ] | 2023+ | Link. | 6 active / 12 | 0.5–0.6 kg | – | Optional fingertip touch | Standard hand on early humanoids; 10k units shipped in 2025 (reported) |
| RH524J1 (Inspire) [ 93 ] | 2026 | Hyb. | 24 (22 hand + 2 wrist) | – | – | Proximity, touch, depth | Tendon-hybrid; modular, thermal management |
| O6 (Linkerbot) [ 94 , 11 ] | 2025 | Link. | 6 | – | 3 kg per finger | – | Volume product, about US$930 |
| O30 (Linkerbot) [ 11 , 95 ] | 2026 | DD | 20 | 740 g | 30 kg rated | – | Each joint independently driven; aimed at RL |
| OmniHand 3 Ultra-M (AGILINK) [ 96 , 95 ] | 2026 | DD | 20 | 630 g | – | 5 fingertip visuo-tactile + 300+ palm taxels | Paired with DUET contact layer |
| Revo 3 (BrainCo) [ 69 ] | 2026 | DD | 21 | – | 70 N grip | Full-palm + fingertip visuo-tactile | Backdrivable; 33 grasp types |
| Hand | Year | Route | DoF | Mass / cost | Highlights |
| LEAP Hand [ 66 ] | 2023 | Servo DD | 16 (4 fingers) | 600 g / US$2k | Universal MCP abduction; 4-h assembly; 1/8 Allegro cost |
| ORCA [ 41 , 106 ] | 2025 | Tendon | 17 | – / CHF 2k | Dislocating joints, self-calibration; 10k cycles; commercial lite/standard/touch |
| RUKA / RUKA-v2 [ 75 , 76 ] | 2025–26 | Tendon | 15 underact. (v1); + 2-DoF wrist (v2) | US$1.3k–1.5k | Fully open CAD; learned controllers |
| CRAFT [ 13 ] | 2026 | Tendon | 15 actuated | US$600 | Hybrid hard–soft compliance; 33/33 GRASP types |
| MM-Hand [ 77 ] | 2026 | Remote tendon | 21 | – | Quick-release tendons; in-palm stereo; sheath-friction analysis |
| MIDAS [ 43 ] | 2026 | DD | 13 of 16 | 700 g / US$3k | 283 three-axis taxels; sim model, retargeting, teleop pipeline |
| System | Date | Route | Hardware | Key idea / finding | Scale |
| DexCap [ 44 ] | 2024-03 | Glove | Mocap glove, chest RGB-D + SLAM | Portable robot-free capture with human-in-the-loop correction | Tens of minutes/task |
| Open-TeleVision [ 45 ] | 2024-07 | VR teleop | Vision Pro, active stereo neck | Immersive active vision | Tens of demos/task |
| Bunny-VisionPro [ 135 ] | 2024-07 | VR teleop | Vision Pro, vibrotactile | Haptic cues and safety constraints | – |
| ACE [ 136 ] | 2024-08 | Exoskeleton | Arm exo + hand tracking | Cross-platform visual exoskeleton | – |
| DOGlove [ 137 ] | 2025-02 | Glove | Open force-feedback glove | Manipulation under occlusion via haptics | – |
| AgiBot World [ 46 ] | 2025-03 | Teleop factory | 100 robots, VR teleop | Standardized pipeline; scaling of GO-1 | 1M+ traj., 2,976 h |
| Teleop | Wearable | Egocentric | Sim/gen. | |
| Cost per hour | + | ++ | ++ (marginal) | |
| Throughput | + | ++ | ++ | |
| Executability | ++ | + (iso.) / | ||
| Embodiment gap | none | small–med. | largest | sim2real |
| Contact / force | ++ | |||
| Scene diversity | + | ++ |
| Field | Content and rationale |
| Hand state | MANO 21 keypoints and/or glove joint angles at native rate; keep raw, not retargeted [ 207 , 20 ] |
| Wrist pose | 6D pose with tracker type and calibration; main source of action noise [ 15 , 16 ] |
| Vision | Raw egocentric, wrist and palm streams with intrinsics/extrinsics; mirrored camera placement on robot [ 17 , 35 ] |
| Touch | Raw taxel signals with sensor model, layout and calibration; optionally a MANO-UV projection [ 231 ] |
| Objects | Meshes and 6D poses where available; required by physics-aware retargeting [ 49 , 53 , 232 ] |
| Robot data | Commands, measured joint states, currents/torques, control rate and latency [ 21 ] |
| Dataset | Year | Content / scale |
| FPHA [ 261 ] | 2017 | 100k+ RGB-D frames, 45 actions, magnetic hand pose |
| HO-3D [ 262 ] | 2019 | 77,558 frames, 10 objects, markerless 3D |
| ContactPose [ 259 ] | 2020 | 2,306 grasps, thermal contact maps |
| GRAB [ 264 ] | 2020 | Whole-body mocap grasps, 51 objects, contact |
| DexYCB [ 263 ] | 2021 | Multi-view YCB grasps; handover benchmark |
| HOI4D [ 266 ] | 2022 | 2.4M ego frames, 800 instances, 16 categories |
| Approach | Gap | Representative work | Contact/force | Reported evidence |
| Kinematic retargeting + video | H R | EgoScale [ 20 ] , UniDex [ 51 ] , Do as I Do [ 296 ] | No | +54% from 20k h pretraining; 8 hands, 52k trajectories |
| Physics/contact-aware retargeting | H R | SPIDER [ 49 ] , OmniHOI [ 53 ] , contact-anchored [ 232 ] , TopoRetarget [ 52 ] | In sim | 9 embodiments; 5 hands 39–89% vs. 31%; 7.3 59.3% |
| Generative visual conversion | H R | EgoEngine [ 212 ] , RoboTryOn [ 225 ] , DexStitch [ 297 ] | No | Zero real-demo dexterous policy (EgoEngine) |
| Deployed-hand / isomorphic capture | H R, arms | DexUMI [ 15 ] , RealDexUMI [ 18 ] | Yes | 88.75%; one checkpoint on 3 arms |
| Sim RL grounding | H R | ManipTrans [ 222 ] , Human2Sim2Robot [ 223 ] | In sim | One human demo to real |
| Unified slot action space | R R | FAAS [ 51 ] , DexFormer [ 298 ] , Being-H0.5 [ 21 ] , AcrossDEX [ 299 ] | No | 30 robots, one checkpoint on 5 platforms |
| Situation | Suggested starting point |
| One target hand, budget for hardware | Isomorphic or deployed-hand capture with shared touch [ 18 , 15 , 149 ] ; add teleoperation only for failure-prone steps |
| Several hands, mostly grasping | Morphology-aligned or canonical-hand policies [ 303 , 302 ] ; cross-hand grasp synthesis [ 318 ] |
| Several hands, general tasks | Function-aligned slot actions with masks and embodiment adapters [ 51 , 21 ] ; ablate relative-EEF vs. joint wrist actions [ 300 , 28 ] |
| Large unlabeled video | Latent actions or world-action pretraining [ 306 , 31 , 319 ] , then paired mid-training [ 20 ] |
| Human video, contact-rich tasks | Contact-anchored or physics-aware retargeting with object meshes [ 232 , 49 , 53 ] ; residual policies on the robot |
| Visual gap dominates | Mirror camera placement [ 17 , 35 ] ; then inpainting or generative conversion with feasibility checks [ 212 , 225 ] |
| Model | Year | Pretraining scale / headline result |
| T3 [ 343 ] | 2024 | 3M+ samples, 13 sensors; +25% insertion vs. scratch |
| Sparsh [ 344 ] | 2024 | 460k+ images; +95.1% avg. on TacBench |
| Sparsh-skin [ 329 ] | 2025 | Full-hand magnetic skin; 41% over prior |
| Sparsh-X [ 345 ] | 2025 | 1M interactions, 4 modalities; +63% policy |
| OmniVTLA [ 357 ] | 2025 | 135k samples; 100% pick-place on hands |
| FTP-1 [ 346 ] | 2026 | 3,000 h, 21 sensors; +31% unseen sensors |
| Work | Year | Scale / headline result |
| DexGraspNet [ 394 ] | 2022 | 1.32M grasps, 5k+ objects (Shadow) |
| UniDexGrasp++ [ 404 ] | 2023 | 85.4% train / 78.2% test success |
| Get a Grip [ 402 ] | 2024 | 3.5M grasps, 4.3k objects; learned evaluators |
| DexGraspNet 2.0 [ 403 ] | 2024 | 427M grasps, 8,270 scenes; 90.7% real in clutter |
| BODex [ 396 ] | 2024 | 49 grasps/s/GPU; 81% on real Shadow Hand |
| DexGrasp Anything [ 401 ] | 2025 | 3.4M+ poses, 15k+ objects |
| System | Slow (semantics) | Mid (visuomotor) | Fast (contact/whole body) |
| Helix 02 [ 33 ] | S2 VLM latents | S1, 200 Hz, fingers incl. | S0, 1 kHz, 10M params |
| CraftNet [ 34 ] | language/task | S1 Motion Brain | S0 Interaction Brain (contact) |
| DUET [ 96 ] | manipulation layer | contact layer (touch) | |
| GO-2 [ 432 ] | action CoT | async. slow intent / fast tracking | |
| T-Rex [ 219 ] | – | low-rate action expert | high-rate tactile expert |
| Organization | Model (latest) | Route | Architecture notes | Data / dexterity evidence |
| Physical Intelligence | [ 26 , 27 , 28 ] | VLA, RL | Flow-matching expert; context conditioning incl. metadata | Zero-shot cross-embodiment folding (grippers) |
| NVIDIA | GR00T N1 N1.7 [ 29 , 300 ] ; DreamZero [ 31 ] | VLA, human data, WAM | Relative EEF actions; embodiment adapters | 20k h human video; +54% on 22-DoF hand [ 20 ] |
| Google DeepMind | Gemini Robotics 1.5 / 2 [ 317 , 425 ] | VLA | Thinking VLA; Motion Transfer | Multi-embodiment transfer; multi-finger control |
| Figure AI | Helix 02 [ 33 ] | Hierarchical | S2/S1 200 Hz/S0 1 kHz | Fingers in S1; palm cams, fingertip force [ 99 ] |
| Sharpa | CraftNet [ 34 ] | Hierarchical VTLA | Motion Brain + Interaction Brain | 22-DoF tactile hand [ 42 ] ; glove data |
| Generalist AI | GEN-0 GEN-1.5 [ 23 ] | Human data | 100 Hz actions, in-context prompting | Hundreds of thousands of hours (press-reported) [ 24 ] |
| Name | Date | Type | Tasks | Embodiment / platform | Scale / finding |
| Bi-DexHands [ 472 ] | 2022 | Sim | Bimanual skills staged by infant development | Two Shadow hands, Isaac Gym | RL solves tasks at 48-month level |
| DexGraspNet [ 394 ] | 2022 | Sim data | Grasp synthesis | Shadow Hand | 1.32M grasps, 5k+ objects |
| ARCTIC [ 269 ] | 2023 | Human data | Bimanual articulated objects | MANO + mocap | 2.1M frames |
| RoboPianist [ 477 ] | 2023 | Sim | Bimanual piano playing | Two Shadow hands | High-rate precision |
| DexArt [ 474 ] | 2023 | Sim | Faucets, lids, laptops, buckets | Allegro, SAPIEN | Generalization across instances |
| OakInk2 [ 268 ] | 2024 | Human data | Affordance–primitive–task hierarchy | Multi-view + mocap | Task-aware motion generation |
| Dimension | Record and report |
| Embodiment | Active/total DoF; coupling; controller; sensor type, layout and calibration |
| Adaptation | Target-hand data, tuning compute, calibration, interventions and privileged state |
| Generalization | Held-out objects, tasks, sites and hands; possible training overlap |
| Task outcome | Successes/trials; per-task results; progress rubric; uncertainty method |
| Physical outcome | Pose/dosing error, contact-force limits, slip and object damage |
| Timing | End-to-end latency, update rate, chunk horizon and deadline misses |
| Skill / domain | Res. | Bench. | Deploy | Main bottleneck |
| Grasping / sorting | 3 | 3 | 3 | Long tail, transparent objects, cycle time |
| In-hand reorientation | 3 | 2 | 1 | Pose estimation; little demand |
| Tool use | 2 | 3 | 2 | Functional contact choice, impacts |
| Bimanual | 3 | 3 | 2 | Error accumulation, synchronization |
| Deformable / fragile | 2 | 1 | 1 | Simulation gap; needs touch |
| Insertion / assembly | 2 | 1 | 2 | Sub-mm precision, insertion force |
| Step | Typical actions | Dex. need | Diff. | Phase |
| Fetch & load | Open containers, pour into machine, scrape | Med. | Med. | P1 |
| Seasoning | Scoop powder, squeeze bottles, tear sachets | High | Med.–high | P1 |
| Packing | Box, close lids, add cutlery, bag | High | Med. | P1 |
| Plating | Portion, arrange, wipe rims | High | High | P2 |
| Dish handling | Scrape, sort, rack | Med. | Med. | P2 |
| Cleaning | Wipe, return tools | Med. | Med. | P2 |
| Claim | Supporting evidence | Boundary or counterpoint | Decisive next comparison |
| Aligned human data can improve learning | EgoScale and DexWild report benefits of human pretraining/co-training [ 20 , 17 ] . | More human data is not always better; gains depend on the bridge and mixture. | Vary source hours and alignment quality at fixed compute, hand and robot data. |
| Contact should constrain transfer | Physics and contact-aware methods improve transferred trajectories [ 49 , 232 ] . | Simulated contact feasibility does not guarantee real friction or force accuracy. | Match human clips and target hands; compare kinematic, contact and real-task errors. |
| No action space is universally best | Relative EEF and joint-space choices receive different support [ 314 , 28 ] . | Embodiment sets, tasks and training pipelines differ across reports. | Compare action spaces on the same data, backbone and morphologically distant hands. |
| Touch and fast feedback are promising | Tactile mid-training and hierarchies show task-specific gains [ 47 , 219 ] . | Sensor heterogeneity persists; a separate contact layer is not always tested against end-to-end control. | Factor sensing, controller split and latency independently; evaluate unseen sensors. |
| Zero-shot dexterity remains narrow | Real cross-hand grasp transfer has been reported [ 303 ] . | Grasp transfer does not establish long-horizon tool use or in-hand skill transfer. | Report zero-shot and fixed-budget adaptation on grasping, reorientation and tool use. |
| Completion is not deployment readiness | Laboratory tasks expose divergence between completion and validity [ 475 ] . | Uptime, repair, damage and cost evidence is sparse or company-reported. | Use fixed-duration trials with intervention, failure and maintenance logs. |
| Stakeholder | What the evidence supports | What to measure and report |
| Hand designers | Backdrivable joints and low backlash for learning; human-like proportions, wrist and camera placement where human data are used; hybrid actuation for force plus precision; modular, quick-swap fingers, tendons and skins | Active DoF and coupling map; continuous and peak fingertip force; endurance under specified load; tactile resolution and coverage; repair time |
| Data teams | Store raw human and robot signals with calibration; mirror camera placement; isomorphic devices with shared touch; diverse collectors and sites; online quality checks; consent and de-identification from day one | Valid-data rate; synchronization error; collector and site diversity; cost per accepted hour; provenance |
| Cross-embodiment researchers | Contact-preserving retargeting; slot action spaces with masks; paired human–robot data; tests on morphologically distant hands | Zero-shot and -shot curves across at least two hand types; contact error, not only joint error |
| Model developers | Large-scale pretraining plus paired mid-training; touch as an input with gating or a separate fast layer; RL and interventions after imitation; fast inference for high-rate control | Success with confidence intervals; latency and control rate; behavior under perturbation; data scale per stage |
| Benchmark designers | Functional, tactile and long-horizon tasks; embodiment-decoupled task definitions; real-world counterparts | Precision, force, damage and takeovers alongside success; sim–real correlation; trial counts |
| Deployers | Use dexterous hands only where needed; deterministic safety layers; sensor-closed dosing; staged rollout with gates | Task success, takeovers per hour, availability, cycle time vs. skilled worker, MTBF, incident counts |
Robotic Manipulation
Robotic manipulation studies how robots perceive, plan, and control physical interactions to grasp, move, assemble, and use objects.
See all 1,805 papers