1. What is a soft robot? (the short version)

• A soft robot is a robot whose soft body deforms to make it move. The body is silicone rubber, plastic film, oil, or sometimes cultured muscle. Unlike a conventional robot built from a metal skeleton and joints, its “moving parts” and its “body” are one and the same [Ref. 1].
• There are several ways to power it, each with a different weakness. Pneumatic (needs a compressor), electrohydraulic (needs a supply of hundreds to thousands of volts), heat (slow to cool) and living muscle (needs a culture bath). All of them still have open problems outside the laboratory.
• “Soft” does not mean “safe”, “energy-saving” or “can grasp anything”. What softness buys is shape adaptation and tolerance of bumps. The price is hard control, low output and heavy power supplies.

First, here is how a rigid robot and a soft robot differ in four parts (Fig. 2).

2. How soft is “soft”?

To put “soft” in numbers, people usually use the Young’s modulus, a measure of stiffness when something is pulled. But a paper that sorted out definitions of soft robots (Chubb et al., 2019) points out that the Young’s modulus ranges called “soft” differ between sources [Ref. 2] Sourced. In other words, we found no common line of the form “below this modulus, it is soft” (not confirmed). So we look at real material numbers instead.

Silicone rubbers widely used for soft-robot prototypes (Ecoflex, Dragon Skin) have their numbers in manufacturer data sheets. Figure 3 plots some of them.

How to read it: the “100% modulus” is the force per unit cross-section when the rubber is stretched to twice its length. Ecoflex 00-30, for example, is 10 psi, or about 69 kPa (0.069 MPa) Our calculation. Dragon Skin 30 is 86 psi, about 593 kPa Our calculation: stretching it to the same twofold length takes about 8.6 times the stress. Elongation at break is 900% for Ecoflex 00-30 and 364% for Dragon Skin 30 [Ref. 3, 4] Sourced. The harder the grade, the less it stretches (the arrangement is Our calculation).

3. Way 1: inflate it with air

The oldest way to move a soft robot is air pressure. A handy comparison is a balloon: put air in a rubber bag and it swells; let the air out and it returns. But a balloon only swells; it does not bend. To make it bend, you attach a “hard-to-stretch layer” to one side (Fig. 4).

Original SVG, detailed explanation and sources

Original SVG reference: Concept diagram (vector drawing). How a pneumatic soft actuator bends: a cross-section of a stretchy layer with a row of chambers bonded to a hard-to-stretch layer. The shapes and proportions are for illustration, not those of a real device. The layer materials follow Shepherd et al. (2011) [Ref. 5].

Shepherd et al. used this approach to build a four-legged robot made entirely of elastomer (rubber). It had just five actuators and simple valves, no sensors, and ran below 10 psi [Ref. 5] Sourced. The stretchy layer is Ecoflex 00-30 or 00-50; the hard-to-stretch layer is PDMS (Sylgard 184). Molds were 3D-printed and the parts cast [Ref. 5] Sourced.

In numbers: 7 psi is about 48 kPa Our calculation (roughly half an atmosphere on top of ambient). A crawling speed of 24 +/- 3 m/h is about 6.7 mm per second Our calculation. The robot also crawled under a gap only 2.0 cm high [Ref. 5] Sourced. But it was tethered by tubing to an external compressed-air source [Ref. 5] Sourced. That points to one of the biggest open problems of the pneumatic approach: can the robot carry its own power?

There are other pneumatic examples. Hawkes et al. reported a vine-like robot that grows from its tip: pressurizing a thin-walled vessel stored inside-out lets the tip lengthen, and an asymmetry in the tip’s lengthening steers it [Ref. 7] Sourced. Wehner et al. reported a robot made only of soft materials, including its control: a microfluidic “logic circuit” autonomously regulates the decomposition of an on-board fuel, and the resulting gas inflates the pneumatic network [Ref. 8] Sourced. It shows that “completely soft, no cable” is one of the goals.

4. Way 2: electrically driven “oil pouches” (HASEL and HALVE)

Instead of air, there is a method that uses electric force to move oil. A handy comparison is a zipper. Apply a voltage across two electrodes in a pouch and they attract each other, closing from one end like a zipper. The oil inside is squeezed out, another part of the pouch swells, and the pouch as a whole shortens (Fig. 5, left). This is called a HASEL (pronounced “hazel”; hydraulically amplified self-healing electrostatic actuator). Combining electrostatic force with oil pressure is its trademark, and in 2018 it was reported that it can repeatedly self-heal after electrical breakdown [Ref. 9] Sourced.

The “Peano-HASEL”, reported the same year, is a pouch that can be made by heat-sealing commercial plastic film, needing no frame or prestretch. The paper reports up to about 10% contraction, a strain rate of 900% per second, operation at 50 Hz and lifting more than 200 times its own weight [Ref. 10] Sourced. These are numbers from laboratory prototypes.

Original SVG, detailed explanation and sources

Original SVG reference: Concept diagram (vector drawing). Left: how an electrohydraulic actuator contracts. Right: layer structure of the conventional type (Peano-HASEL) and the low-voltage type (HALVE). Dimensions and permittivities follow Gravert et al. (2024) [Ref. 12]; this is not the exact structure of any real product.

Cutting the voltage with materials: HALVE

The weak point of HASEL is its high voltage. Gravert et al. (ETH Zurich) note that conventional electrohydraulic actuators need several thousand volts, which makes the driving electronics bulky and inefficient [Ref. 12] Sourced. They also list exposed electrodes, which make the devices hard to use near people.

Their answer is the HALVE (hydraulically amplified low-voltage electrostatic) actuator. Conventionally, a single film (such as BoPET, biaxially oriented polyester) served as both “structure” and “insulation (the capacitor’s dielectric)”. HALVE separates the two roles. A strong polymer outer shell provides structure and outer insulation, and a thin layer of a high-permittivity polymer (P(VDF-TrFE-CTFE)) earns the electrostatic force (Fig. 5, right). The paper states that replacing 15 µm of BoPET (relative permittivity 3.3) with 5 µm of P(VDF-TrFE-CTFE) (relative permittivity about 40) could in theory cut the required voltage about six-fold [Ref. 12] Sourced.

In measurements, HALVE reached an average power density of 50.5 W/kg and a peak strain rate of 971% per second at 1100 V, which the paper says is 4.9 times lower in voltage than the previous state of the art [Ref. 12] Sourced. The paper sets this beside typical mammalian skeletal muscle (about 50 W/kg and a peak of about 500% per second) [Ref. 12] Sourced. It also showed devices that are safe to touch, waterproof and “self-clearing” (when insulation fails, the electrode is destroyed locally and the short circuit clears) [Ref. 12] Sourced. The power electronics were also made compact: a 28 cm robotic fish carried them and swam untethered. Its speed, however, was at most 3.8 cm/s (0.14 body lengths per second) at 2 Hz, the average power draw was 0.6 W, and one battery charge lasted about 54 minutes of continuous actuation [Ref. 12] Sourced.

More detail: how much does the force change if you halve the voltage?

Electrostatic force is proportional to the square of the electric field (Maxwell stress). For the same geometry the field is proportional to voltage, so halving the voltage cuts the force to about one quarter Our calculation. For HALVE to cut the voltage and keep the force, a material with high relative permittivity must draw a large force from the same voltage (the “about six-fold” estimate above rests on this idea). This explanation is ours, not a derivation from the paper’s equations.

5. A legged robot test: is the power use really small?

An example that tested whether an artificial muscle works as a robot’s power source is the electrohydraulic musculoskeletal leg (PELE) of Buchner et al. It was a joint effort of the Soft Robotics Lab at ETH Zurich and the Max Planck Institute for Intelligent Systems, among others, published in Nature Communications in 2024 [Ref. 13] Sourced. It is a carbon-fiber skeleton with 3D-printed joints and an antagonistic (pulling-in-opposite-directions) pair of electrohydraulic muscles. The muscles are stacks of Peano-HASEL connected to the joints by tendons [Ref. 13] Sourced.

The main results are as follows [Ref. 13] Sourced.

• Gait motions above 5 Hz and high jumps up to 40% of the leg height (highest 128 mm)
• Hopping over grass, sand, gravel, pebbles and large rocks using open-loop force control only (control that does not measure and correct for the ground)
• Self-sensing of obstacles from the change in the muscles’ capacitance
• A minimum net cost of transport (COT), a measure of movement efficiency, of 0.73 (at a robot mass of 225 g)

COT is the energy needed to move a given weight a given distance, divided by the work of gravity on it (E / mgl), a dimensionless number. For a 225 g robot at 0.73, moving 1 m takes about 1.6 J Our calculation (0.73 x 0.225 kg x 9.81 m/s² x 1 m). The paper states that this value excludes the loss in high-voltage conversion (typically about 75% efficient) [Ref. 13] Sourced. Assuming 75% would give about 2.1 J/m Our calculation, but that is our own estimate under our assumption, not a value in the paper.

The paper also compared PELE with a direct-drive motor leg of the same size and mass. During squatting the average power was 25.3 W for the motor leg and 306 mW for PELE (1.2% of the motor leg). Just holding a posture took 17 W for the motor leg and 9.4 mW for PELE (Fig. 6) [Ref. 13] Sourced. As a ratio, that is about 1 to 1800 Our calculation.

The leg’s voltage-force relationship also needs care. The paper notes a dead zone of 0 to 3 kV (depending on the external load) in which the muscles do not respond to voltage [Ref. 13] Sourced. The controller overlaps extensor and flexor activation to compensate. Even a body that is comparatively soft has awkward behavior in its drive, such as a range where voltage does nothing.

6. Way 3: heat, living muscle and jamming

Besides pneumatic and electrohydraulic drives, three other means of power or function often come up. Figure 7 lines up the five methods by “power in”, “example” and “limit” – our own arrangement.

Heat-driven systems (overview)

Shape memory alloys and coiled fibers contract when warmed by an electric current. They give a large force from a small volume, but they take time to cool, so motion tends to be slow. Details are in our explainers on shape memory alloys and on soft actuators and artificial muscles.

Moving with living muscle (biohybrid)

Another approach uses real cultured muscle as the power source. In 2018 a University of Tokyo group reported a robot whose joint was moved by an antagonistic pair of skeletal muscle tissues. The joint turned about 90 degrees, the lifetime was about one week, and it could pick and place objects [Ref. 14] Sourced. In 2025 the group reported a hand with fingers powered by bundles of muscle tissue (MuMuTA). One bundle produced a contractile force of about 8 mN over a contractile length of about 4 mm, and cables turned that into bending of multi-jointed fingers [Ref. 15] Sourced. 8 mN corresponds to the weight of about 0.8 g Our calculation.

Switching stiffness: jamming

The last method is not about “moving” but about “switching stiffness”. Brown et al. reported a gripper made from a thin rubber bag filled with powder (ground coffee). Press the bag onto an object and the powder flows around its shape. Pump the air out (a pressure difference of about 75 kPa) and the grains lock together and harden, so the bag can lift the object while wrapped around it [Ref. 16] Sourced. The membrane is latex rubber 0.3 mm thick. For many items (light bulbs, LEDs, bottle caps, tubing and others) the pick-up success rate over 10 trials each was 100%. The objects it could not grip were those the bag cannot wrap around the sides of: large hemispheres, thin disks lying flat, and very soft items such as cotton balls [Ref. 16] Sourced.

7. How do you control a soft body?

The other wall for soft robots is control. For a rigid arm, measuring the joint angles fixes the pose. A soft body bends continuously, and the quantities describing “where and how much it is bent” are in principle very many (Fig. 9, top). In a survey, Della Santina, Duriez and Rus define continuum soft robots as “mechanical systems entirely made of continuously deformable elements” and say that what characterizes the class is continuum (or multi-body) dynamics, a non-negligible elastic potential field, and uncertainty in sensing and actuation [Ref. 17] Sourced.

So how is it handled? There are roughly three ideas. (1) Compute with a model: predict the body’s shape and force from a continuum-mechanics model and use it. The survey covers both the systematic overview and the open problems [Ref. 17] Sourced. (2) Use classical control at the low level: in PELE, target tendon positions come from inverse kinematics (a calculation that works backward from a target tip position to the joint angles), and muscle force is adjusted by PID control [Ref. 13] Sourced. (3) Let the softness of the body do some of the work: with pairs of antagonistic muscles, adjust stiffness and let the body absorb the ground’s shape. PELE hopped over grass, sand and stones with no elaborate correction [Ref. 13] Sourced.

8. Where robots and drones could use this, and the limits

Here is a summary from the standpoint of robots and drones. Examples of soft mechanisms mounted on drones were not covered by the sources we checked (not confirmed). The readings below are our commentary.

The facts in each row are as stated in the bracketed references. The arrangement of the table is Our calculation. The right-hand column is our commentary (readings without a source).

9. Limits, what we could not confirm, and a summary

What we could not confirm

• A common Young’s-modulus definition of “soft”: we confirmed that ranges differ between sources, but found no common line [Ref. 2]
• Examples on drones: we did not confirm specific reports of drones carrying soft grippers or legs
• Lifetime and reliability for industrial or outdoor use: our examples are laboratory experiments, and we found no multi-year durability data
• Standards aimed directly at artificial muscles or soft robots: a search of ISO standards for “soft robot” and “artificial muscle” turned up none that we could confirm (related general standards are in Chapter 10)
• Full texts of some papers: the Rus and Tolley review (Nature) and the papers by Wehner et al. and Hawkes et al. were checked at the abstract level; we state no numbers from their main text
• ASTM standards: ASTM International’s site states that it prohibits AI use, so we did not open it. We state only the fact of the numbers the manufacturer’s data sheet cites

Summary

• A soft robot is one whose “moving parts” and “body” are merged. Softness is read from per-material data sheets (Fig. 3) Our calculation
• There are five ways to move one. Pneumatic is simple and has a long history but tends to be tethered to its power source. Electrohydraulic has shown muscle-like output but needs high-voltage supplies. Heat, living muscle and jamming each have other limits (Fig. 7)
• HALVE cut the voltage about 4.9-fold through a materials idea (a high-permittivity polymer). Force density and thin-film defects remain challenges Sourced
• An electrohydraulic leg showed low power use at rest. But the comparison partner and conditions are limited, so “more efficient than motors” cannot be concluded (Fig. 6)
• Controlling a soft body is hard; combinations of classical control and the body’s softness are being tried (Fig. 9)

10. How to read the yardsticks and standards

A soft robot’s performance can be sorted into three groups of yardsticks: material yardsticks (hardness, tension, dielectric breakdown), actuator yardsticks (power density, strain rate, lifetime) and robot yardsticks (cost of transport, grasp success rate, power use). For every yardstick you must check what it is divided by and under what conditions it was measured, or the numbers cannot be compared. “50 W/kg”, for example, means different things depending on whether the mass is that of the muscle or of the whole device. The table lists what we could confirm (checked 25 October 2026).

Numbers, titles and status of standards were checked on the official ISO pages (25 October 2026). The texts are paid, so we stay with the scope statements. The arrangement of the table is Our calculation.

Cautions in reading (for materials engineers)

• A demonstration, a benchmark and the field are different things: a leg hopping over grass is the result of one leg in a lab. It is not the same as long-duration operation or safety certification of a robot
• A manufacturer’s stated value and independent verification are different: data-sheet values are manufacturer-stated and conditional. It is worth re-measuring with your own samples and conditions
• Standards are mostly voluntary: ISO standards generally become binding only when cited in law or contract (commentary). When you see “complies with a standard”, check which standard and which requirement
• Plans and achievements are different: standards under revision (ISO 8373, the successor to ISO/TS 15066) may change in content Not yet confirmed

11. Glossary

12. References (primary sources)

1. Rus D., Tolley M. T. “Design, fabrication and control of soft robots”, Nature 521, 467-475 (2015), DOI 10.1038/nature14543 (checked at abstract level) – https://doi.org/10.1038/nature14543
2. Chubb O. et al. “Towards an ontology for soft robots: what is soft?”, Bioinspiration & Biomimetics 14, 063001 (2019), DOI 10.1088/1748-3190/ab483f (checked at abstract level) – https://doi.org/10.1088/1748-3190/ab483f
3. Smooth-On “Ecoflex Series Technical Bulletin” (silicone data sheet, version code 011124DH) – https://www.smooth-on.com/tb/files/ECOFLEX_SERIES_TB.pdf
4. Smooth-On “Dragon Skin Series Technical Bulletin” (silicone data sheet, version code 040825PA) – https://www.smooth-on.com/tb/files/DRAGON_SKIN_SERIES_TB.pdf
5. Shepherd R. F. et al. “Multigait soft robot”, PNAS 108, 20400-20403 (2011), DOI 10.1073/pnas.1116564108 (main text checked) – https://pmc.ncbi.nlm.nih.gov/articles/PMC3251082/
6. ISO “ISO 48-4:2018 Rubber, vulcanized or thermoplastic – Determination of hardness – Part 4: Indentation hardness by durometer method (Shore hardness)” (ISO standard page) – https://www.iso.org/standard/74969.html
7. Hawkes E. W. et al. “A soft robot that navigates its environment through growth”, Science Robotics 2, eaan3028 (2017), DOI 10.1126/scirobotics.aan3028 (checked at abstract level) – https://doi.org/10.1126/scirobotics.aan3028
8. Wehner M. et al. “An integrated design and fabrication strategy for entirely soft, autonomous robots”, Nature 536, 451-455 (2016), DOI 10.1038/nature19100 (checked at abstract level) – https://doi.org/10.1038/nature19100
9. Acome E. et al. “Hydraulically amplified self-healing electrostatic actuators with muscle-like performance”, Science 359, 61-65 (2018), DOI 10.1126/science.aao6139 (checked at abstract level) – https://doi.org/10.1126/science.aao6139
10. Kellaris N. et al. “Peano-HASEL actuators: Muscle-mimetic, electrohydraulic transducers that linearly contract on activation”, Science Robotics 3, eaar3276 (2018), DOI 10.1126/scirobotics.aar3276 (checked at abstract level) – https://doi.org/10.1126/scirobotics.aar3276
11. ISO “ISO 37:2024 Rubber, vulcanized or thermoplastic – Determination of tensile stress-strain properties” (ISO standard page) – https://www.iso.org/standard/86892.html
12. Gravert S.-D. et al. “Low-voltage electrohydraulic actuators for untethered robotics”, Science Advances 10, eadi9319 (2024), DOI 10.1126/sciadv.adi9319 (open access; main text checked) – https://pmc.ncbi.nlm.nih.gov/articles/PMC10775996/
13. Buchner T. J. K. et al. “Electrohydraulic musculoskeletal robotic leg for agile, adaptive, yet energy-efficient locomotion”, Nature Communications 15, 7634 (2024), DOI 10.1038/s41467-024-51568-3 (open access; main text checked) – https://doi.org/10.1038/s41467-024-51568-3
14. Morimoto Y. et al. “Biohybrid robot powered by an antagonistic pair of skeletal muscle tissues”, Science Robotics 3, eaat4440 (2018), DOI 10.1126/scirobotics.aat4440 (checked at abstract level) – https://doi.org/10.1126/scirobotics.aat4440
15. Ren Z. et al. “Biohybrid hand actuated by multiple human muscle tissues”, Science Robotics 10, eadr5512 (2025), DOI 10.1126/scirobotics.adr5512 (checked at abstract level) – https://doi.org/10.1126/scirobotics.adr5512
16. Brown E. et al. “Universal robotic gripper based on the jamming of granular material”, PNAS 107, 18809-18814 (2010), DOI 10.1073/pnas.1003250107 (main text checked) – https://pmc.ncbi.nlm.nih.gov/articles/PMC2973877/
17. Della Santina C., Duriez C., Rus D. “Model-Based Control of Soft Robots: A Survey of the State of the Art and Open Challenges”, IEEE Control Systems 43(3), 30-65 (2023) (arXiv:2110.01358; checked at abstract level) – https://arxiv.org/abs/2110.01358
18. ISO “ISO/TS 15066:2016 Robots and robotic devices – Collaborative robots” (ISO standard page) – https://www.iso.org/standard/62996.html
19. ISO “ISO 8373:2021 Robotics – Vocabulary” (ISO standard page) – https://www.iso.org/standard/75539.html
20. ISO “ISO 10218-1:2025 Robotics – Safety requirements – Part 1: Industrial robots” (ISO standard page) – https://www.iso.org/standard/73933.html
21. Falco J., Van Wyk K., Messina E. “Performance Metrics and Test Methods for Robotic Hands”, NIST Draft Special Publication 1227 (October 2018, draft) – https://doi.org/10.6028/NIST.SP.1227-draft
22. Calli B. et al. “Benchmarking in Manipulation Research: The YCB Object and Model Set and Benchmarking Protocols”, IEEE Robotics & Automation Magazine 22(3), 36-52 (2015) (arXiv:1502.03143) – https://arxiv.org/abs/1502.03143
23. ISO “ISO 13482:2014 Robots and robotic devices – Safety requirements for personal care robots” (ISO standard page) – https://www.iso.org/standard/53820.html

13. Claim-to-source audit

Last updated: 25 October 2026 / Sources are limited to primary sources (peer-reviewed papers, official ISO and NIST pages, manufacturer data sheets). Numbers are published values from papers and data sheets, under differing conditions. A common Young’s-modulus definition of “soft”, examples on drones, long-term industrial durability and standards aimed directly at soft robots could not be confirmed, so they are not stated. Figures 2 to 7 and 9 are concept diagrams or charts with our calculation made for explanation; Figures 1 and 8 are frames for AI-generated images (no image placed yet).