Chapter 5: Intelligent Mechanisms — Physical Intelligence

Overview

While Chapter 4 surveyed the full spectrum of robot hand design, this chapter focuses on three approaches where the mechanism itself exhibits intelligence: underactuation, variable stiffness actuators (VSAs), and active surfaces. Seminar 3 (Inchul) provided the richest source material, demonstrating how integrating these three mechanisms creates synergies for factory automation.

After reading this chapter, you will be able to... - Explain the principles of underactuation and its shape adaptation capabilities. - Compare four implementation approaches for variable stiffness actuators. - Understand how active surfaces contribute to continuous-contact manipulation. - Describe the impact of integrated mechanism design on factory automation.

5.1 Underactuation: Shape Adaptation with Fewer Actuators

Underactuation refers to designs where the number of actuators is less than the number of joints. Remaining joints are passively determined by springs or mechanical stops. The core value is automatic adaptation to object shape — performed by physical mechanism rather than control algorithms.

Key Paper: Bicchi 2000. "Hands for Dexterous Manipulation and Robust Grasping: A Difficult Road Toward Simplicity." IEEE T-RA. The foundational paper for underactuation theory, proving the paradoxical proposition that "fewer actuators can yield more robust grasping."

Pisa/IIT SoftHand

Among the most referenced hands in Seminar 3:

SoftHand 1 [Li et al.; Catalano et al.]:

• 1 actuator drives 19 joints
• Adaptive synergy: all joints coordinated along the first principal synergy
• Tendons slip on contact, automatically expanding contact area
• 3D-printable, modular design

SoftHand 2 [Della Santina et al.]:

• 2 actuators: primary + auxiliary synergy
• Friction-based actuation expansion enables more diverse grasp types

Palm Concavity and Adaptability

Capsi-Morales et al. proposed sectioning the palm and connecting sections via underactuation, allowing the palm itself to adapt to objects — extending the adaptive region beyond fingers.

Pull-in Mechanism

Kakogawa et al. combined underactuated fingers with a pull-in mechanism to solve the challenging task of picking up flat objects from a table surface.

Dollar's Hand

Dollar et al. pursued extreme simplification of underactuation for manufacturing environments, achieving robust grasping with minimal complexity.

ILDA: Linkage-Driven Adaptive Grasping

Kim et al. [2021, Nature Communications] proposed ILDA (Integrated Linkage-Driven Dexterous Anthropomorphic Robotic Hand), which achieves the same intelligent adaptation through a linkage-driven mechanism rather than tendons. In free space the links maintain coupled motion; on contact, the linkage passively reconfigures to fit the grasped object. This design shares the shape-adaptation behavior of underactuation while avoiding tendon wear and slack issues (→ Chapter 4.5).

5.2 Variable Stiffness Actuators: Switching Between Soft and Rigid

Variable stiffness actuators (VSAs) enable a single mechanism to transition between compliant and stiff states. Seminar 3 systematically organized four implementation approaches:

5.2.1 Parallel Beam

Fu et al. use a slider to adjust the effective beam length, continuously varying finger stiffness. Combined with ArUco marker-based vision force sensing for low-cost force feedback.

5.2.2 Pneumatic

Al Abeach et al. arrange McKibben artificial muscles antagonistically, enabling independent stiffness control while maintaining pose. This means adjustable resistance to disturbances while holding an object.

5.2.3 SMA + SMP

Wang & Ahn combine shape memory alloy (SMA) actuation with shape memory polymer (SMP) stiffness switching, achieving 55x stiffness change per hinge.

5.2.4 Particle Jamming

Wei et al. combine fiber-reinforced pneumatic actuators with particle jamming for 10x+ stiffness increase. Soft state adapts to objects; jamming the particles transitions to rigid grasping.

BISA (Bidirectional Stiffening Soft Actuator)

Lin et al.'s BISA independently controls bending stiffness and lateral stiffness, enabling more precise stiffness profile management.

Reconfigurable Passive Joints

Kopicki et al. [2025, IEEE RA-L, arXiv:2501.16006] proposed RP-joints (reconfigurable passive joints) that reconfigure under external force, then lock via tendon tension — achieving variable configuration without dedicated VSA hardware. The two-stage "shape-select → lock" approach reports 80% success on IKEA objects and 87% on YCB datasets using single-example learning.

The RP-joint-based three-fingered gripper uses 5-joint fingers (Links 1–5), and each finger adapts geometrically to flat surfaces and round objects differently — a paradigmatic example of physical intelligence where contact configurations adjust passively to object shape under an identical actuation signal.

5.3 Active Surfaces: Belt- and Roller-Based Continuous Contact Manipulation

Active surfaces drive the gripper's contact surface itself, enabling sliding or repositioning objects while maintaining contact.

Belt-Based Active Surfaces

In work by Kim et al. and Wang et al., belts are mounted on gripper finger surfaces and driven:

• Thin object approach: Belts pull cards/paper up from a table surface
• Repositioning/alignment: Belt drives adjust object position within the grasp
• Continuous contact maintenance: Manipulation occurs without releasing and re-grasping

The key insight from Seminar 3: continuous contact is easier to control and more stable than discrete contact. It eliminates the uncertainty that arises from releasing and re-grasping.

5.4 Integrating the Three Mechanisms: Adapt, Fix, Manipulate

Seminar 3's most original contribution is the temporal integration of the three mechanisms:

1. Underactuation: Initial approach → passive adaptation to object shape
2. Variable Stiffness (VSA): After adaptation → stiffness transition locks the grasp
3. Active Surface (Belt): While locked → belt drives manipulate/reposition/align

These three stages implement the natural flow of human grasping — "approach → grasp → adjust" — through mechanism.

Key Insight: Even with identical sensing and learning, stability improves dramatically when the mechanism induces continuous contact. This is the "role of physical design" — embedding intelligence in physical structure rather than relying solely on software.

5.5 Factory Automation Applications: Thin Objects, Multiple Objects, Repositioning

Seminar 3 presented concrete factory automation scenarios for this integrated mechanism:

5.5.1 Thin Object Grasping

Extremely thin objects (cards, paper, gaskets) are difficult for parallel grippers. Active belt surfaces lift edges from the table, followed by underactuated finger grasping. PP-Tac^[17] [#12] (R-Tac + Diffusion Policy, 87.5% success) attacks this problem from the tactile+learning direction; Seminar 3's mechanism approach attacks it from physical design (→ Chapter 8.3).

5.5.2 Multiple Object Handling

Grasping multiple objects simultaneously, or adding objects to an existing grasp. Underactuation's shape adaptation plays the key role; the F-TAC Hand's 100% multi-object success shows how tactile feedback reinforces this further (→ Chapter 4.5).

5.5.3 Repositioning/Alignment/Rotation

Adjusting object position or orientation within the grasp. Active belt surfaces play the central role — more stable than discrete contact (release-and-regrasp).

5.6 Continuous vs. Discrete Contact: The Role of Physical Design

Seminar 3's deepest insight is the fundamental difference between continuous and discrete contact:

Discrete Contact (most current approaches):

• Grasp → release → re-grasp
• Each transition requires new contact planning
• Uncertainty increases at state transitions
• Learning-based approaches must absorb this uncertainty

Continuous Contact (mechanism-based approach):

• Manipulation while maintaining contact
• Stable contact state → simpler control
• Mechanism "physically guarantees" continuous contact
• Reduced burden on learning and sensing

Key Perspective: When mechanisms induce continuous contact, the burden on sensing and learning is reduced. This forms one axis of the "mechanism + tactile + learning triangle" proposed in Chapter 14 (→ Chapter 14.4).

The Process-Centric Manipulation Taxonomy [2025, Nature Machine Intelligence] formalizes the connection from process specifications to tactile manipulation skills, validating 28 skills from industrial domains at ~100% success rates. This taxonomy encompasses continuous-contact manipulation skills aligned with Seminar 3's mechanism approach.

5.7 The 2026 Commercial Robot-Hand Landscape

Commercial robot hands in 2026 are moving from expensive research platforms such as Shadow and Allegro toward humanoid-scale integration parts. The shift is not just higher DoF. Product value now depends on tactile sensing, teleoperation, simulator assets, SDKs, ROS2/EtherCAT/USB integration, and compatibility with VLA/GR00T-style learning stacks.

As the Cosmax meetings show through Robotis, Tesollo, and Wuji comparisons, real purchasing decisions are more pragmatic than paper specs ^{[26] [27]}. Lead time, price, tactile add-ons, torque-control support, simulator availability, repairability, and finger-pad replacement can dominate the research schedule. In manufacturing, DoF and tactile density matter less than cycle time, lifetime, cleaning, dust protection, calibration stability, operator override, and QA traceability.

5.7.1 Separating Research Benchmarks from Manufacturing Purchase Candidates

It is not defensible to rank 2026 hands by claiming which is "most widely used" without adoption data. Allegro and Shadow remain common research benchmarks, but they are not the same category as hands being considered for a manufacturing cell. Wuji, Tesollo, and Robotis-style candidates are judged more by lead time, price, SDK quality, simulator support, torque-control APIs, tactile options, and repairability than by paper citations.

The point is not to recommend one product. It is to avoid comparing research platforms, humanoid components, and manufacturing PoC candidates under one metric. For a Cosmax-style cell, the first PoC should ask whether the hand logs tactile/force evidence, can free one finger while retaining the first object, reproduces failures in simulation, and allows pad/sensor replacement within operator workflow.

Summary and Outlook

Intelligent mechanisms are not alternatives to software intelligence (sensing, learning) but complements. Underactuation provides shape adaptation, VSAs provide state locking, and active surfaces provide continuous-contact manipulation. Their integration creates synergy beyond the sum of parts. Seminar 3's factory scenarios (thin objects, multiple objects, repositioning) represent challenges one level above current industrial deployments (logistics), and mechanism-based approaches are a key strategy for bridging this gap.

The next chapter examines human hand data collection — how we teach robots to manipulate by demonstration (→ Chapter 6: Human Hand Data Collection).

Manufacturing-Cell Checkpoint

Commercial hand selection is not a catalog ranking. In production, payload, repeatability, closing speed, and tactile options matter, but so do serviceability and integration cost. Before purchase, four questions should be answered: how quickly pads can be cleaned or replaced, how calibration drift behaves over shift-length operation, whether the SDK logs tactile packets, joint state, and fault codes under one schema, and whether the vendor can supply spare parts and firmware updates.

For cosmetics manufacturing, general dexterity is less important than fit to the process. Bottles, caps, tubes, pouches, and small cartons require different contact surfaces and force limits. Sharpa, AgiBot, XHAND, Inspire, Wuji, Unitree, Allegro, and Shadow should therefore be compared by public availability, tactile options, control interface, palm usefulness, cleanability, price, and delivery risk rather than by an unsupported claim that one hand is the most widely used.

Operational Reading Note

The practical value of this chapter is not only the concept of commercial hand purchasing and mechanism choice; 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. 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
2. Li, H., Ford, C. J., Bianchi, M., Catalano, M. G., Psomopoulou, E., & Lepora, N. F. (2022). BRL/Pisa/IIT SoftHand: A low-cost, 3D-printed, underactuated, tendon-driven hand with soft and adaptive synergies. IEEE Robotics and Automation Letters, 7(4), 8745-8751. scholar
3. Catalano, M. G., Grioli, G., Farnioli, E., Serio, A., Piazza, C., & Bicchi, A. (2014). Adaptive synergies for the design and control of the Pisa/IIT SoftHand. International Journal of Robotics Research, 33(5), 768-782. https://doi.org/10.1177/0278364913518998 scholar
4. Della Santina, C., Piazza, C., Grioli, G., Catalano, M. G., & Bicchi, A. (2018). Toward dexterous manipulation with augmented adaptive synergies: The Pisa/IIT SoftHand 2. IEEE Transactions on Robotics, 34(5), 1141-1156. scholar
5. Capsi-Morales, P., Grioli, G., Piazza, C., Bicchi, A., & Catalano, M. G. (2020). Exploring the role of palm concavity and adaptability in soft synergistic robotic hands. IEEE Robotics and Automation Letters, 5(3), 4703-4710. scholar
6. Kakogawa, A., Nishimura, H., & Ma, S. (2016). Underactuated modular finger with pull-in mechanism for a robotic gripper. IEEE International Conference on Robotics and Biomimetics (ROBIO), 556-561. scholar
7. Odhner, L. U., Jentoft, L. P., Claffee, M. R., Corson, N., Tenzer, Y., Ma, R. R., Buehler, M., Kohout, R., Howe, R. D., & Dollar, A. M. (2014). A compliant, underactuated hand for robust manipulation. International Journal of Robotics Research, 33(5), 736-752. scholar
8. Fu, J., Yu, Z., Guo, Q., Zheng, L., & Gan, D. (2023). A variable stiffness robotic gripper based on parallel beam with vision-based force sensing for flexible grasping. Robotica. https://doi.org/10.1017/S026357472300156X scholar
9. Al Abeach, L. A. T., Nefti-Meziani, S., & Davis, S. (2017). Design of a variable stiffness soft dexterous gripper. Soft Robotics, 4(3), 274-284. scholar
10. Wang, W., & Ahn, S.-H. (2017). Shape memory alloy-based soft gripper with variable stiffness for compliant and effective grasping. Soft Robotics, 4(4), 379-389. scholar
11. Wei, Y., Chen, Y., Ren, T., Chen, Q., Yan, C., Yang, Y., & Li, Y. (2016). A novel, variable stiffness robotic gripper based on integrated soft actuating and particle jamming. Soft Robotics, 3(3), 134-143. scholar
12. Kim, Y., Shin, J., Won, J., Lee, W., & Seo, T. (2024). LBH gripper: Linkage-belt based hybrid adaptive gripper design for dish collecting robots. Robotics and Autonomous Systems, 185, 104886. scholar
13. Wang, H., Gao, B., Zhao, D., & Shen, H. (2025). A reconfigurable gripper inspired by elastic belt for versatile in-hand manipulations. IEEE Robotics and Automation Letters. scholar
14. Lin, J., et al. (2023). A bioinspired bidirectional stiffening soft actuator for multimodal, compliant, and robust grasping. Soft Robotics. https://doi.org/10.1089/soro.2022.0212 scholar
15. Kopicki, M., Ansary, S. I., Tolomei, S., Angelini, F., Garabini, M., & Skrzypczyński, P. (2025). Underactuated dexterous robotic grasping with reconfigurable passive joints. IEEE Robotics and Automation Letters. arXiv:2501.16006. scholar
16. Johannsmeier, L., et al. (2025). A process-centric manipulation taxonomy for the organization, classification and synthesis of tactile robot skills. Nature Machine Intelligence. https://doi.org/10.1038/s42256-025-01045-3 scholar
17. Lin, P., Huang, Y., Li, W., Ma, J., Xiao, C., & Jiao, Z. (2025). PP-Tac: Paper picking using omnidirectional tactile feedback in dexterous robotic hands. RSS 2025. #12 scholar
18. Hogan, N. (1985). Impedance control: An approach to manipulation. JDSMC, 107(1), 1-24. scholar
19. 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
20. Sharpa. (2026). Sharpa Wave and CraftNet product information. https://www.sharpa.com/ [Sharpa, 2026] source
21. AgiBot. (2026). AGIBOT OmniHand 2025 product page and specifications. https://store.agibot.com/products/omnihand-2025-w-tactile [AgiBot, 2026] source
22. RobotEra. (2026). XHAND1 dexterous hand brochure. https://www.robotera.com/ [RobotEra, 2026] source
23. Inspire Robots. (2026). RH56DFX series dexterous hand product page. https://www.inspire-robots.com/ [Inspire Robots, 2026] source
24. Wuji Technology. (2026). Wuji Hand product documentation. https://docs.wuji.tech/docs/en/wuji-hand/latest/overview/ [Wuji Technology, 2026] source
25. Unitree Robotics. (2026). Dex5-1 dexterous hand product page. https://www.unitree.com/mobile/Dex5-1/ [Unitree Robotics, 2026] source
26. Cosmax Robotics Meeting. (2026a). Sequential multi-object grasping and active in-hand rearrangement problem statement. Internal meeting PDF, 2026-05-12. [Cosmax, 2026a] private source
27. Cosmax Robotics Meeting. (2026b). Model-based approach vs RL-based approach for in-hand manipulation. Internal meeting PDF, 2026-06-05. [Cosmax, 2026b] private source