Part II: Hands — Robot and Human

Chapter 4: Robot Hand Design — Machines Built to Grasp

Written: 2026-04-01 Last updated: 2026-04-01

Overview

Tactile sensors (Chapter 2) and data (Chapter 3) are meaningless without a robot hand to house them. This chapter explores the spectrum of robot hand design — from parallel grippers to dexterous hands — the open-source revolution that brought costs from $100,000+ to $2,000, and the state of the art in sensor-integrated design.

After reading this chapter, you will be able to... - Explain the design trade-offs between parallel grippers and dexterous hands. - Understand the design philosophy and impact of open-source low-cost hands. - Distinguish key actuation methods: tendon-driven, direct-drive, and others. - Describe the Allegro Hand ecosystem's role in research.

4.1 Parallel Grippers vs. Dexterous Hands: Design Trade-offs

Seminar 3 (Inchul) systematically compared these two design extremes.

Parallel Grippers:

  • Two opposing fingers, 1 DoF
  • Simple control, high reliability, dominant in industrial settings
  • Limitations: Poor shape adaptation, difficulty maintaining continuous contact, vulnerable with thin or multiple objects

Dexterous Hands:

  • 4-5 fingers, 12-22 DoF
  • Diverse grasp types and in-hand manipulation capability
  • Limitations: Control complexity, combinatorial explosion of contact states, high cost, frequent breakage
Key Paper: Bicchi 2000. "Hands for Dexterous Manipulation and Robust Grasping: A Difficult Road Toward Simplicity." IEEE Transactions on Robotics and Automation, 16(6), 652-662. A landmark argument for simplicity in hand design, introducing the underactuation and synergy concepts that shaped modern affordable hand designs including SoftHand and LEAP Hand.

The key insight from Seminar 3 transcends this dichotomy: intelligent mechanisms (detailed in Chapter 5) can combine the simplicity of parallel grippers with the adaptability of dexterous hands.

4.2 The Open-Source Revolution: LEAP Hand, ISyHand, ORCA

Since 2023, the emergence of low-cost open-source dexterous hands has fundamentally transformed the research ecosystem.

LEAP Hand (2023)

Developed by Shaw, Agarwal, and Pathak[1] at CMU, LEAP Hand ignited the open-source hand revolution:

  • $2,000 cost (1/8 of the $16,000 Allegro Hand)
  • 3D-printable: Anyone can build one
  • 4 fingers, 16 DoF
  • Outperforms Allegro Hand on benchmark tasks
  • 200+ citations (RSS 2023)
Key Paper: Shaw et al. 2023. "LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning." RSS 2023. A $2,000, 3D-printable open-source dexterous hand that democratized manipulation research.

LEAP Hand's design philosophy is "maximum dexterity at minimum cost." A novel kinematic structure reproduces the essential degrees of freedom of the human hand using only low-cost servo motors and 3D-printed components. The key innovation is a universal abduction-adduction mechanism at the MCP (knuckle) joint: unlike Allegro (which loses DoF in the extended position) or prior designs (which lose DoF when flexed), LEAP Hand retains all degrees of freedom in all finger positions. This leads to a significantly larger thumb opposability volume and higher manipulability ellipsoid, translating to better grasping (19.5 N pull-out force, exceeding human grip) and faster in-hand cube rotation than Allegro.

Figure 4.1: LEAP Hand overview — (a) anthropomorphic dexterous hand costing ~$2,000 and assemblable in 4 hours, (b-c) scale comparison with human hand (b) and LEAP Hand-C (c), (d-h) grasping of diverse objects. Source: Shaw et al. (2023), Fig. 1.
Figure 4.1: LEAP Hand overview — (a) anthropomorphic dexterous hand costing ~$2,000 and assemblable in 4 hours, (b-c) scale comparison with human hand (b) and LEAP Hand-C (c), (d-h) grasping of diverse objects. Source: Shaw et al. (2023), Fig. 1.
Figure 4.2: Relative size of popular robot hands, left to right: human hand, Allegro, LEAP-C, LEAP, InMoov, D'Manus. LEAP is similar in size to Allegro and about 30% larger than a human hand. Source: Shaw et al. (2023), Fig. 3.
Figure 4.2: Relative size of popular robot hands, left to right: human hand, Allegro, LEAP-C, LEAP, InMoov, D'Manus. LEAP is similar in size to Allegro and about 30% larger than a human hand. Source: Shaw et al. (2023), Fig. 3.

ISyHand (2025)

ISyHand[5] pushes cost even lower to $1,300 while introducing an articulated palm:

  • 4-hour assembly time
  • Off-the-shelf + 3D-printed components
  • Palm articulation expands grasp diversity
Figure 4.3: ISyHand — an articulated palm (base / middle / distal palm link) actively wraps around objects to form stable grasps of diverse shapes. Source: Richardson et al. (2025), Fig. 1.
Figure 4.3: ISyHand — an articulated palm (base / middle / distal palm link) actively wraps around objects to form stable grasps of diverse shapes. Source: Richardson et al. (2025), Fig. 1.

ORCA (2025)

Born from ETH's Katzschmann group, ORCA is a 17-DoF tendon-driven hand:

  • Under 2,000 CHF, 8-hour build time by a single person
  • Integrated tactile sensors (FSR-based binary touch on all 5 fingertips)
  • Poppable pin joints that dislocate instead of breaking, plus auto-calibration via center-of-rotation tendon routing
  • Demonstrated 10,000+ continuous operation cycles (20+ hours) without hardware failure
  • Payload capacity up to 10.5 kg (103 N) in a four-finger grasp
  • Zero-shot sim-to-real RL for in-hand ball reorientation
  • Commercialization pursued via Mimic Robotics
Figure 4.4: ORCA Hand overview — identical form factor to the human hand, 17 DoF hand + 1 DoF wrist, silicone skin, integrated tactile sensors, poppable pin joints (A1-A3), ratchet spools (A2), cooling fans. Demonstrated 2,000+ continuous cycles without hardware failure. Source: Christoph et al. (2025), Fig. 1.
Figure 4.4: ORCA Hand overview — identical form factor to the human hand, 17 DoF hand + 1 DoF wrist, silicone skin, integrated tactile sensors, poppable pin joints (A1-A3), ratchet spools (A2), cooling fans. Demonstrated 2,000+ continuous cycles without hardware failure. Source: Christoph et al. (2025), Fig. 1.
Figure 4.5: ORCA Hand versatility — (A-D) teleoperation via ROKOKO gloves (holding a drill, gripping a cube, pouring liquid, self-wrapping), (E) imitation learning from demonstrations, (F) zero-shot sim-to-real RL for in-hand ball reorientation. Source: Christoph et al. (2025), Fig. 2.
Figure 4.5: ORCA Hand versatility — (A-D) teleoperation via ROKOKO gloves (holding a drill, gripping a cube, pouring liquid, self-wrapping), (E) imitation learning from demonstrations, (F) zero-shot sim-to-real RL for in-hand ball reorientation. Source: Christoph et al. (2025), Fig. 2.

These three hands share a common philosophy — open-source + low-cost + 3D-printable — reducing the entry barrier for dexterous manipulation research from $100K+ to under $2K.

Hand Cost DoF Actuation Tactile Open-Source Year
Shadow Hand $100K+ 24 Tendon/pneumatic Optional (BioTac) No 1990s
Allegro Hand $16K 16 Direct drive No No 2012
LEAP Hand $2K 16 Direct drive No Yes 2023
ISyHand $1.3K — Direct drive No Yes 2025
ORCA $2K 17 Tendon Yes Yes 2025
F-TAC Hand — — — 17 sensors Partial 2025

4.3 Tendon-Driven Designs: Pisa/IIT SoftHand, CRAFT, Mimic Robotics

Tendon-driven hands place actuators outside the fingers (e.g., in the forearm), transmitting force via tendons. The advantages are finger miniaturization, weight reduction, and natural compliance.

Pisa/IIT SoftHand

A direct product of Bicchi's [2000] "road toward simplicity" philosophy:

  • SoftHand 1: 1 actuator, 19 joints — adaptive synergy enables shape adaptation to diverse objects
  • SoftHand 2: 2 actuators, friction-based actuation expansion — more diverse grasp types
  • 3D-printable, modular design
Key Paper: Bicchi and Kumar 2000. "Robotic Grasping and Contact: A Review." IEEE ICRA 2000. Foundational review of robotic grasping theory covering force/form closure, grasp quality metrics, and contact models.

CRAFT Hand (2026)

Lin et al. [14] integrate hybrid hard-soft compliance with tendon-driven actuation. Rigid links provide precision position control while soft elements enable shape adaptation. Fifteen motors mounted directly on the fingers drive the hand through tendons, at under $600, and the hand covers all 33/33 grasp types in the Feix taxonomy.

Figure 4.6: CRAFT Hand — (A) full hand view showing a compact anthropomorphic form factor with forearm-mounted actuators, (B) holding a tennis ball, (C) one finger with soft PLA links + soft TPU joints, (D, E) compliant conformation to a surface during contact. Source: Lin et al. (2026), Fig. 1.
Figure 4.6: CRAFT Hand — (A) full hand view showing a compact anthropomorphic form factor with forearm-mounted actuators, (B) holding a tennis ball, (C) one finger with soft PLA links + soft TPU joints, (D, E) compliant conformation to a surface during contact. Source: Lin et al. (2026), Fig. 1.

Mimic Robotics (ORCA Hand)

A startup from ETH's Katzschmann group, Mimic Robotics pursues Physical AI for factory environments using the ORCA Hand. The lightweight tendon-driven design achieves near-human compliance with integrated tactile sensors.

CATCH-919 Hand

Zhang et al.[4] designed a hand with 9 actuators and 19 DoFs featuring fingertip hyperextension, achieving 33 stable grasp types.

4.4 Core Design Principles: DoF, Actuation, Compliance, Cost

Four axes define robot hand design:

4.4.1 Degrees of Freedom

The human hand has approximately 27 DoF, but research shows that most everyday grasps can be explained by 2-3 synergies [3]. This observation provides the theoretical basis for SoftHand's 1-2 actuator design.

DoF Range Representative Grasp Types In-Hand Manipulation
1 (gripper) Industrial parallel gripper Power grasp None
1-2 (underactuated) SoftHand, Dollar's Hand Adaptive power grasp Limited
12-16 LEAP, Allegro Power + precision Basic
17-22 ORCA, Shadow Diverse types Capable

4.4.2 Actuation Methods

  • Direct drive: Motor directly coupled to joint. Allegro, LEAP. Fast control response, simple structure.
  • Tendon-driven: Force transmitted via tendons. SoftHand, ORCA, Shadow. Miniaturization, compliance, cost reduction.
  • Pneumatic: Air-pressure driven. Deformable, safe. Lower control precision.
  • Hydraulic: Sanctuary AI Phoenix Gen 8. High power, high precision. Complex, expensive.

4.4.3 Compliance

Compliance is essential for contact-rich manipulation. As emphasized in Seminar 1, position control alone is insufficient — torque control-capable dexterous hands are necessary.

Soft Robotic Hand with Tactile Palm-Finger Coordination [2025, Nature Communications] achieved diverse-shape object grasping through coordinated palm-finger tactile sensing on soft materials.

4.4.4 Cost

The five-year price compression trend:

  • Shadow Hand: $100K+ (1990s-present)
  • Allegro Hand: $16K (2012-present)
  • LEAP Hand: $2K (2023)
  • ISyHand: $1.3K (2025)

This compression impacts not only research democratization but also the economic viability of industrial deployment.

4.5 Sensor-Integrated Design: Marrying Hands with Touch

Sensor technologies from Chapter 2 achieve maximum impact when co-optimized with hand design.

The F-TAC Hand [2025, Nature Machine Intelligence] covered 70% of the hand surface with 17 vision-based tactile sensors, achieving 100% multi-object grasp success. Sensor placement was optimized based on contact probability distributions (→ Chapter 2.4).

ORCA[6] integrated tactile sensors from the design stage, avoiding the difficulties of after-the-fact sensor attachment.

Integrated Linkage-Driven Dexterous Anthropomorphic Robotic Hand [7] proposed a novel design combining coupled motion in free space with adaptive grasping during contact through a linkage-driven mechanism.

Figure 4.7: F-TAC Hand grasp workspace — 17 high-resolution vision-based tactile sensors empower the hand to perform all 33 human grasps of the Feix taxonomy (power, precision, intermediate categories). Source: Zhao et al. (2025), Fig. 3.
Figure 4.7: F-TAC Hand grasp workspace — 17 high-resolution vision-based tactile sensors empower the hand to perform all 33 human grasps of the Feix taxonomy (power, precision, intermediate categories). Source: Zhao et al. (2025), Fig. 3.
Figure 4.8: F-TAC Hand comprehensive tactile sensing — encoder-decoder networks learn from raw tactile signals (a) and physics-based simulation data (b–d), enabling object identification and contact-pose estimation from contact geometry (e–g). Source: Zhao et al. (2025), Fig. 4.
Figure 4.8: F-TAC Hand comprehensive tactile sensing — encoder-decoder networks learn from raw tactile signals (a) and physics-based simulation data (b–d), enabling object identification and contact-pose estimation from contact geometry (e–g). Source: Zhao et al. (2025), Fig. 4.
Figure 4.9: TacPalm SoftHand — a soft gripper with a high-density tactile sensor array integrated into the palm. Palm-finger coordination senses contact across diverse grasp types (pinch, power grasp, etc.). Source: Zhang et al. (2025), Nature Communications, Fig. 1.
Figure 4.9: TacPalm SoftHand — a soft gripper with a high-density tactile sensor array integrated into the palm. Palm-finger coordination senses contact across diverse grasp types (pinch, power grasp, etc.). Source: Zhang et al. (2025), Nature Communications, Fig. 1.

4.6 The Allegro Hand Ecosystem and Research Platforms

Wonik Robotics' Allegro Hand ($16K) has served as the de facto standard for dexterous manipulation research since before open-source hands appeared. The 4-finger, 16-DoF, direct-drive Allegro has been the core platform for:

  • DeXtreme [12]: Allegro Hand + Isaac Gym for sim-to-real dexterous manipulation (→ Chapter 10.2)
  • RGMC (Robotic Grasping and Manipulation Competition): Held annually at ICRA, with Allegro as a primary platform
  • D(R,O) Grasp [13]: 89% real-world success with LEAP Hand, 87.53% average simulation success across multiple dexterous hands (ICRA 2025 Best Paper, arXiv:2410.01702)

Wonik is currently pursuing integration with Meta FAIR's Digit Plexus, an important step toward standardized sensor-hand interfaces.

Review papers provide broader context:

  • Kadalagere Sampath et al.[1]: Comprehensive review of human-like manipulation using dexterous hands
  • Anthropomorphic Five-Fingered Hand Manipulation[10]: Comparison of hybrid transmission schemes
  • Soft Robotic Dexterous Hands Advances[11]: Latest trends in soft robotic dexterous hands
Figure 4.10: D(R,O) Grasp — unified representation of robot-object interaction. Configuration-invariant pretraining enables cross-embodiment grasp prediction transferrable across multiple dexterous hands. Source: Wei et al. (2024), Fig. 1.
Figure 4.10: D(R,O) Grasp — unified representation of robot-object interaction. Configuration-invariant pretraining enables cross-embodiment grasp prediction transferrable across multiple dexterous hands. Source: Wei et al. (2024), Fig. 1.

Summary and Outlook

Robot hand design is converging along three trends: (1) open-source designs under $2K (LEAP, ISyHand, ORCA), (2) standardization of integrated tactile sensing (F-TAC Hand, ORCA), and (3) hybrid actuation combining rigid precision with soft compliance. When these trends merge, sub-$1K dexterous hands with integrated tactile will become reality (→ Chapter 14.2).

The next chapter examines intelligent mechanisms that combine the simplicity of parallel grippers with the adaptability of dexterous hands through physical design (→ Chapter 5: Intelligent Mechanisms).

Manufacturing-Cell Checkpoint

When a robot hand enters a manufacturing cell, mechanical design must be judged by more than paper benchmarks. DoF and underactuation should be evaluated by whether the target SKU family requires a particular contact transition, not by how human-like the hand appears. Fingertip and palm compliance protect fragile packages, but excessive compliance reduces force-closure margin and repeatability. Cable routing, dust sealing, cleanable pads, replacement time, spare parts, and calibration scripts often determine uptime more directly than an extra joint.

A practical review starts from the task family. Split the line into pinching, enveloping, palm support, in-hand shift, and sequential grasping. Then map each task to required contact surfaces, force limits, allowable deformation, and finger-repositioning needs. Finally check whether the tactile layout actually observes those states. This turns a vague demand for a dexterous hand into concrete requirements for hand geometry, sensor placement, and control bandwidth.

Operational Reading Note

The practical value of this chapter is not only the concept of hand mechanics and sensor placement; it is the set of engineering decisions that the concept changes. A deployable robot-hand project should start by asking what state becomes observable after this chapter is applied. The answer should be concrete: contact existence, contact patch, normal force, shear direction, slip margin, object pose, task phase, operator override, or product-damage risk. If a variable cannot be logged or consumed by a controller, it remains an explanatory idea rather than a system capability.

The second decision is the unit of evidence. Research demos often report one success metric, but tactile manipulation improves through failure records. A useful attempt record contains the object or SKU, the selected grasp candidate, the robot hand and sensor configuration, calibration version, task phase, tactile summary, policy action, safety intervention, and final outcome. This record is what connects the sensor chapters to the data chapter, the control chapters to the learning chapters, and the manufacturing chapters to QA.

The third decision is where the chapter sits in the control stack. Some ideas belong in fast reflex loops, some in contact MPC, some in policy inputs, and some only in offline diagnosis. Mixing these time scales creates brittle systems: a VLA cannot react to millisecond slip, and a low-level force controller cannot infer the next process step. The right architecture separates fast contact stabilization, mid-level grasp or rearrangement control, and slow task reasoning.

Finally, the chapter should be evaluated by the failure modes it removes. A method that improves benchmark success but leaves the team unable to distinguish perception failure, contact-acquisition failure, force-closure failure, execution-time slip, or maintenance drift is not yet production-ready. A method with slightly lower headline performance but better logs, safer force limits, and clearer recovery hooks may be the stronger basis for manufacturing Physical AI.

References

  1. Shaw, K., Agarwal, A., & Pathak, D. (2023). LEAP Hand: Low-cost, efficient, and anthropomorphic hand for robot learning. RSS 2023. arXiv:2309.06440. scholar
  2. Bicchi, A. (2000). Hands for dexterous manipulation and robust grasping: A difficult road toward simplicity. IEEE Transactions on Robotics and Automation, 16(6), 652-662. scholar
  3. Bicchi, A., & Kumar, V. (2000). Robotic grasping and contact: A review. IEEE ICRA 2000. scholar
  4. Zhao, Z., Li, W., Li, Y., et al. (2025). Embedding high-resolution touch across robotic hands enables adaptive human-like grasping. Nature Machine Intelligence. https://doi.org/10.1038/s42256-025-01053-3 #39 scholar
  5. Richardson, B. A., Grüninger, F., Mack, L., Stueckler, J., & Kuchenbecker, K. J. (2025). ISyHand: A dexterous multi-finger robot hand with an articulated palm. IEEE-RAS Humanoids 2025. arXiv:2509.26236. scholar
  6. Christoph, C. C., Eberlein, M., Katsimalis, F., Roberti, A., Sympetheros, A., Vogt, M. R., Liconti, D., Yang, C., Cangan, B. G., Hinchet, R. J., & Katzschmann, R. K. (2025). ORCA: An open-source, reliable, cost-effective, anthropomorphic robotic hand. arXiv preprint. arXiv:2504.04259. scholar
  7. Kim, U., Jung, D., Jeong, H., Park, J., Jung, H.-M., Cheong, J., Choi, H. R., Do, H., & Park, C. (2021). Integrated linkage-driven dexterous anthropomorphic robotic hand. Nature Communications, 12, 7177. https://doi.org/10.1038/s41467-021-27261-0 scholar
  8. Zhang, N., Ren, J., Dong, Y., Gu, G., & Zhu, X. (2025). Soft robotic hand with tactile palm-finger coordination. Nature Communications, 16, 2395. https://doi.org/10.1038/s41467-025-57741-6 #40 scholar
  9. Kadalagere Sampath, et al. (2023). Review on human-like robot manipulation using dexterous hands. Cognitive Computation and Systems (IET/Wiley). scholar
  10. Various. (2025). Human-like dexterous manipulation for anthropomorphic five-fingered hands: A review. Journal of Engineering Science and Technology Review. https://doi.org/10.1016/j.jestch.2025.101938 scholar
  11. Various. (2025). Soft robotic dexterous hands: Advances and challenges. International Journal of Advanced Manufacturing and Mechatronic. scholar
  12. Handa, A., et al. (2023). DeXtreme: Transfer of agile in-hand manipulation from simulation to reality. ICRA 2023. scholar
  13. Wei, Z., Xu, Z., Guo, J., Hou, Y., Gao, C., Cai, Z., Luo, J., & Shao, L. (2024). D(R,O) Grasp: A unified representation of robot and object interaction for cross-embodiment dexterous grasping. ICRA 2025 (Best Paper). arXiv:2410.01702. scholar
  14. Lin, L., Patel, S., Moon, J., Lazebnik, S., & Jain, U. (2026). CRAFT: A tendon-driven hand with hybrid hard-soft compliance. arXiv preprint. arXiv:2603.12120. scholar
  15. Zhang, Z., Han, T., Pan, J., & Wang, Z. (2025). CATCH-919 Hand: A 9-actuator 19-DOF anthropomorphic robotic hand. arXiv preprint. scholar
  16. Yu, M., Jiang, Y., Chen, C., Jia, Y., & Li, X. (2025). RGMC Champion: Kinematic trajectory optimization for in-hand manipulation. IEEE Robotics and Automation Letters. scholar