Compendium

Soft Robotics — Robotics Reference (expanded 2026-05-17)

22 min read

Soft Robotics — Continuum Bodies, Compliant Actuation, Embodied Safety

1. At a glance

Soft robotics replaces rigid-link / rigid-joint kinematics with continuously deformable bodies built from elastomers, gels, fluids, fabrics, and active polymers. The defining material property is compliance: Young’s modulus E ≈ 10⁴–10⁹ Pa, overlapping human soft tissue (skin ≈ 10⁵ Pa, muscle ≈ 10⁴ Pa, cartilage ≈ 10⁷ Pa) and far below structural metals (steel E ≈ 2·10¹¹ Pa, aluminium ≈ 7·10¹⁰ Pa). Because the body deforms continuously, classical Denavit-Hartenberg kinematics do not apply — continuum formulations (piecewise-constant-curvature, Cosserat rod, finite-element) take their place. See [[Robotics/kinematics-dh]] for the rigid counterpart and [[Robotics/Tier3/control-algorithms]] for control-side implications.

Advantages.

  • Embodied safety — low-modulus contact intrinsically limits peak force; ISO/TS 15066:2016 power-and-force-limited collaborative thresholds are met by geometry, not just by torque sensing. A soft pneumatic gripper at 200 kPa applies ≈ 5–15 N on grasp, two orders of magnitude below a comparable rigid jaw.
  • Adaptive grasping — passive shape conformance lets one finger envelope eggs, raspberries, irregular automotive trim, and machined steel parts without re-tooling. The “morphological computation” thesis (Pfeifer & Bongard 2007 How the Body Shapes the Way We Think) is most visible here.
  • Dexterity through continuum DoF — a single tendon-driven soft trunk approximates the maneuverability of an octopus arm: theoretically infinite DoF, practically 6–20 effective modal degrees once compliance is regularised.
  • Lightweight, low inertia — silicone density ρ ≈ 1100 kg·m⁻³ versus steel ρ ≈ 7800 kg·m⁻³; the same envelope weighs roughly 1/7 as much, helping wearables and untethered platforms.

Tradeoffs.

  • Control is hard — nonlinear hyperelastic constitutive laws, hysteresis, viscoelastic creep, slow time constants (tens to hundreds of ms for pneumatics).
  • Sensing is hard — strain gauges, IMUs, and encoders presume rigidity. Flexible analogues (capacitive skin, GelSight, fiber Bragg gratings) lag rigid sensor maturity by ~10 years.
  • Repeatability is poor — cast-and-cure elastomers vary ±5–15 % cell-to-cell; calibration drift over weeks from UV, ozone, plasticiser migration.
  • Power tethers — pneumatics need a compressor or CO₂ cartridge; HASEL needs kilovolt drivers; DEAs need 1–10 kV at low current. Untethered, sustained operation remains the field’s open problem in 2026.

2. Materials

The active design palette in 2026:

Silicone elastomers. The workhorse — addition-cure (platinum-catalysed) PDMS systems. Smooth-On Ecoflex 00-30 (shore 00-30, elongation 900 %, tensile 1.4 MPa) and 00-50 (shore 00-50, slightly stiffer) for pneumatic chambers; Dragon Skin 10/20/30 NV for higher tear strength on fingers. Dow Sylgard 184 for thin membranes and microfluidics (E ≈ 1.5 MPa at 10:1 base:crosslinker ratio). Cure: room-temp 4 h or 70 °C 30 min. Vacuum degassing essential to avoid pinholes that become rupture initiators.

Polyurethanes. Higher abrasion resistance than silicone, more chemically resistant, but stiffer and less stretchy. Smooth-On VytaFlex 30/60 for moulded tooling and fingers needing toughness; Bayer (Covestro) Desmopan thermoplastic PU for extrusion / 3D-printable filaments (NinjaFlex, Recreus Filaflex). Used heavily by Soft Robotics Inc commercial mGrip end-effectors because food-contact NSF certification is cheaper.

Hydrogels. Polyacrylamide, PEG, alginate-PAAm tough hydrogels (Suo group, Harvard). Used in transparent grippers, ionically conductive sensors, biocompatible bioelectronics. Tend to dry out — sealed/hermetic skins required.

Dielectric elastomers. Pre-strained acrylic or silicone films sandwiched between compliant electrodes (carbon grease, CNT, silver nanowire). 3M VHB 4910 / 4905 is the canonical research elastomer (E ≈ 1 MPa, 1000 % stretch). Wacker Elastosil P7670 is the production silicone with lower viscoelastic losses for actuators that must cycle. Drive voltage 1–10 kV, fields 100–300 MV·m⁻¹ — below the dielectric breakdown limit of ≈ 400 MV·m⁻¹.

Shape-memory polymers / alloys. NiTi (Nitinol) 50/50 at% is the canonical alloy — austenite-martensite transformation at 40–80 °C produces recoverable strains of 8 %, generated stresses to 700 MPa. Drives small actuators (cardiac stents, catheter steering, biomimetic fish tails). Slow (thermally limited); Joule-heating typical, water-cooling for cycling. Shape-memory polymers (epoxies, polyurethanes, polynorbornenes — DiAplex, Veriflex) give larger strains (50 %) at lower forces.

See [[Engineering/Tier3/polymers-taxonomy]] for the broader polymer chart (thermoplastics, thermosets, elastomers, gels) and [[Engineering/microfluidics]] for fabrication overlap.

3. Pneumatic soft actuators

Pneumatics dominate the field because air is cheap, safe, and gives 10–500 kPa pressures matched to elastomer yield.

McKibben pneumatic artificial muscle (PAM). Invented J.L. McKibben 1958 for paediatric orthotics. A rubber bladder inside a braided sleeve: pressurised, the bladder expands radially, the braid forces axial contraction of 25–35 % at forces of 0.1–10 kN depending on diameter (10–50 mm typical). Force-length curve is nonlinear and hysteretic. Commercial: Festo MAS / Fluidic Muscle (10/20/40 mm, up to 6 bar, 6000 N); Shadow Robot Air Muscle; Suzumo PMA. Used in [[Robotics/Tier3/end-effectors-zoo]] antagonistic pairs because a single PAM only pulls.

PneuNet bending actuators. Pleated Networks of inflatable chambers in an elastomer matrix with a strain-limiting layer on one side. Mosadegh, Whitesides et al. 2014 Adv. Funct. Mater. — the canonical “pneumatic finger” geometry. Inflation 50–200 kPa, bend angles to 320°, response 100 ms-1 s depending on chamber volume. Variants: slow-PneuNet (low-pressure, soft), fast-PneuNet (segmented, faster). Most academic soft grippers descend from this.

Fiber-reinforced soft actuators. Helical wraps of inextensible fibre (Kevlar thread) constrain radial expansion to enforce bending or twisting modes (Polygerinos, Walsh et al. 2015 IEEE T-RO). The fibre helix angle dictates mode: 0° = pure extension, 54.7° = pure twist, 90° = pure radial expansion blocked → pure bending when paired with a strain-limiting layer.

Inflatable / sliding-layer. Origami balloons, layer-jamming actuators (Wall, Wood, Rus 2015) — granular or sheet jamming under vacuum stiffens by 10–50× on demand, the basis of Empire Robotics VERSABALL and a recurring theme in variable-stiffness manipulators.

Pressure sources. From smallest to largest:

  • Hand bulb / foot pump — < 50 kPa, demonstration only.
  • CO₂ cartridge (12 g) — ≈ 6 L at 1 bar, used by Harvard’s Octobot (Wehner et al. 2016 Nature) for untethered runs.
  • Portable diaphragm pump — Cobionix CB2, KNF, Parker — 1–3 bar, ~1 L·min⁻¹, battery-powered.
  • Compressor + regulator + soft valve manifold — bench standard, 6–10 bar from a small oil-less compressor (Jun-Air, Atlas Copco), regulated to 50–500 kPa.
  • Soft valves — Mahon, Whitesides et al. — elastomer kink-valves and SMA-driven micro-valves for fully untethered, electronics-free control (the Octobot lineage).

4. Hydraulic soft actuation

Water or oil is roughly 1000× less compressible than air, giving sharper force control and better trajectory tracking at the cost of mass, sealing complexity, and slower fill/drain. Used where precision matters or where electrical drive is desirable.

Classical hydraulic soft. Water-filled silicone tubes with custom valve trees (servo proportional valves, Festo MPYE; piezo Lee Co. micro-valves). Wehner et al. 2014 swimming soft fish; Marchese, Onal, Rus 2014 hydraulic soft octopus tentacle — 12 separately addressable chambers, 6 bar, sub-mm tracking.

HASEL (Hydraulically Amplified Self-healing ELectrostatic) actuators. Keplinger group, U. Colorado Boulder, 2018 Science — a dielectric polymer pouch filled with insulating oil with compliant electrodes on opposite faces; applying 5–10 kV pulls the electrodes together via Maxwell stress, displacing oil and changing pouch shape. Combines the high force of hydraulics with the speed and direct electrical drive of DEAs. Strains 100 %+, response 50 ms, self-heals after dielectric breakdown because the liquid dielectric reflows. Commercial: Artimus Robotics (Boulder, founded 2018, scaling to industrial pick-and-place and HVAC-actuator product lines through 2025-26). Peano-HASEL and donut-HASEL geometries each have different strain/force tradeoffs.

Self-healing. Beyond HASEL, a parallel research thread: polymers (Stanford Bao group; Otto group, U. Cambridge) reform covalent or hydrogen-bonded networks after a cut, restoring 80–95 % of mechanical strength in minutes-hours at room temperature.

5. Electroactive polymers (EAPs)

Direct electrical drive — no air, no pump, in principle wearable.

Dielectric Elastomer Actuators (DEAs). Pelrine, Kornbluh, Joseph 2000 Science — the founding paper. Maxwell stress σ = ε₀·εᵣ·E² compresses the elastomer between two compliant electrodes; incompressibility pushes the film outward in-plane. Strains routinely 30–100 %, peak 380 % reported (VHB 4910 pre-strained). Drive voltage 1–10 kV at < 1 mA — kV-DC supplies are small but the high voltage is a safety design problem. Spin-out Artificial Muscle Inc (Pelrine), later acquired by Bayer Material Science then dissolved; commercial DEA remains a research-lab predominate technology in 2026.

Ionic Polymer-Metal Composites (IPMCs). Nafion or Flemion ion-exchange membrane plated with platinum or gold electrodes. Applying 1–5 V drives hydrated cations toward the cathode, swelling that side and bending the strip. Low voltage, low force (mN-range), works only hydrated → encapsulated for air operation. Bar-Cohen, JPL — IPMC fish and microgrippers; widely used in microfluidic valves.

Piezoelectric polymer PVDF / P(VDF-TrFE). Direct piezo coupling in a flexible film. Low strain (< 0.1 %) but useful for high-frequency actuation and especially sensing (acoustic, tactile, ultrasonic transducers). See sensing § below.

6. Tendon-driven and cable-driven soft systems

Cables routed through a compliant body convert linear motor rotation into bending / twisting / contraction. Less elegant than fluidic but easier to control (electric motors are well-understood) and to instrument (motor encoders measure pull).

Festo BionicSoftHand 2.0 (2019, refreshed 2021, 2024) — 12 pneumatic chambers plus tendons in the thumb, taught with RL in simulation to manipulate a Rubik’s cube. The hybrid pneumatic + tendon pattern is the industry standard for anthropomorphic hands.

Shadow Robot Dexterous Hand / Modular Grasper — 24-DoF tendon-driven (Spectra polyethylene cables), 20 antagonistic pneumatic Air Muscles in the original variant, modern variants use electric motors with cable spools. Used by OpenAI 2019 Dactyl and 2020 Solving Rubik’s Cube — pneumatic + tendon platform demonstrating in-hand manipulation purely from sim2real RL.

Pisa/IIT SoftHand (Catalano, Bicchi 2014 IJRR) — 19 joints driven by a single tendon under one motor; passive elastic elements distribute force along an “adaptive synergy” so the hand naturally conforms. Demonstrates underactuation as compliance: motor count ≪ DoF, the body itself solves the grasp.

Flexure-based compliance — replace pin joints with notch flexures (FlexPivot, polymer compliant mechanisms, see [[Robotics/compliant-mechanisms]]) so a “rigid” robot becomes locally soft at the joints. The middle ground between fully soft and fully rigid; used in surgical instruments and microassembly.

See [[Robotics/cable-driven-robots]] for the parallel cable-robot family (workspace-scale rather than end-effector-scale).

7. Magnetic soft robots

A relatively recent (post-2015) thread: embed hard-magnetic micro-particles (NdFeB powder) or soft-magnetic particles (carbonyl iron) in an elastomer, then shape and pole the body so an external magnetic field actuates it remotely. No tether, no on-board power.

Diller & Sitti (CMU) 2014 — programmable shape change in millimetre-scale swimmers and crawlers, driven by Helmholtz coils.

Capsule endoscope steering. Commercial: Ankon NaviCam, Anx Robotica NaviCam, Capsovision add external-magnet steering to passive ingestible capsules for gastric screening. Olympus EndoCapsule and Medtronic PillCam still ship as passive, but field studies through 2024-26 show steerable capsules clinically maturing.

Endovascular catheter. Kim, Parada, Zhao (MIT) 2019 Sci. Robot.; Zhao, Cooper et al. (Massachusetts General / MIT) 2020 Nature — a hydrogel-coated ferromagnetic guidewire steered by an external magnet through the cerebral vasculature in a phantom, slipperier and less traumatic than a manually-pushed guidewire. Early-stage clinical translation 2024-26 with multiple startups (Bionaut Labs for intracranial drug delivery; Magnebotix).

8. Origami and kirigami soft robots

Folded and slit sheets give programmable shape change without bulk material. The flat substrate is easy to fabricate; the folds and slits encode the kinematics.

Origami magic ball (Li, Wang, Rus, Wood 2019) — a polyhedral bellows that collapses under vacuum into a 90 %-volume-reduced gripper that envelops irregular produce. MIT CSAIL / Harvard SEAS.

Kirigami crawler (Rafsanjani, Bertoldi 2018 Sci. Robot.) — laser-cut plastic sheet wrapped around an elastomer body; the kirigami skin produces anisotropic friction so cyclic inflation produces directed locomotion. Stanford and Harvard variants.

Origami robotic worm (MIT / Wood, 2014 Science) — flat sheet self-folds via SMA hinges and walks away. Demonstrates printable / programmable assembly.

9. Modeling

Hyperelastic constitutive laws. The strain-energy density W(λ₁,λ₂,λ₃) of an incompressible rubber:

  • Mooney-Rivlin (1940/1948): W = C₁(I₁−3) + C₂(I₂−3). Good to 100–200 % strain.
  • Yeoh (1993): W = Σ Cᵢ(I₁−3)ⁱ for i=1,2,3. Better for large strains 200–700 %.
  • Ogden (1972): W = Σ (μᵢ/αᵢ)(λ₁^αᵢ + λ₂^αᵢ + λ₃^αᵢ − 3). Most flexible, used for fitting commercial silicones above 500 % strain.
  • Gent (1996), Arruda-Boyce (1993) — physically motivated alternatives capturing strain-hardening at the chain-extension limit.

Fit parameters from uniaxial / planar / biaxial tensile data (ASTM D412, D624). Most FEA packages (ANSYS, Abaqus, COMSOL) support all five.

Reduced-order kinematics.

  • Piecewise Constant Curvature (PCC) — Webster & Jones 2010 IJRR — divide a continuum body into N sub-arcs, each described by 3 parameters (curvature κ, plane angle φ, length L). 3N DoF approximates infinite-DoF behaviour reasonably for slender bodies in non-contact configurations.
  • Cosserat rod theory — Renda, Boyer, Dias, Wood 2018 IEEE T-RO and earlier — treats the rod as a continuum of oriented frames with shear, extension, bending, torsion. PDE numerically solved by shooting methods or implicit integration. Exact in the slender-rod limit.
  • Modal / functional approaches — represent strain field as a Fourier or polynomial basis; Della Santina et al. 2019. Captures geometric non-linearity at low computational cost.

Learned dynamics. Thuruthel, Falotico, Renda, Laschi, Iida 2019 IEEE T-RO — recurrent neural network predicts pneumatic actuator pose from pressure history, eliminating need for first-principles model. Trained on tens of minutes of data; works inside the training distribution. Hybrid physics-ML approaches (Della Santina + sim2real) extend reach.

Simulation packages.

  • SOFA (Inria) — open-source FEM for soft body, originally surgical simulation; the de-facto research standard, supports rigid-soft coupling and contact.
  • PyElastica (Gazzola lab, UIUC) — Cosserat-rod Python library, fast for slender continua.
  • SoMo (Soft Motion) — Truby et al. — PyBullet-based, simplified contact for soft grippers + RL training.
  • Chrono::FEA, MuJoCo soft bodies (since 2022 with mesh-based soft DoFs), NVIDIA Isaac Sim soft-body extensions (2024 onwards).

10. Sensing

Rigid encoders, IMUs, force sensors do not survive infinite-DoF deformation. The soft-sensor toolkit:

Flexible strain. Carbon-loaded silicone, eutectic gallium-indium (eGaIn) microchannels (StretchSense, Liquid Wire). Resistance changes 10–100 %·strain⁻¹. Adafruit flexible strain sensor (re-branded SoftWear); MXene films (3M / Drexel Gogotsi) — 2D titanium-carbide flakes, very high gauge factor. StretchSense commercial gloves (motion capture, VR, finger telemetry).

Capacitive skin / pressure. Two compliant electrodes separated by an elastomer — capacitance C = ε₀·εᵣ·A/d changes under deformation. Pressure Profile Systems, Tekscan FlexiForce (low-cost piezoresistive 0.1–20 N), Sensel Morph capacitive surface (32k cells, 5 g resolution), Pi-Squared / Pi.SQR lab-scale arrays. Used for soft gripper grip-mapping and prosthetic skin.

GelSight tactile. Adelson group, MIT — Adam Sipe et al. originated the concept ~2009 in IROS and a 2017 Sensors expansion; coats an elastomer pad with reflective paint, observes deformation through an embedded camera. Resolves geometry, slip, texture at micron-scale lateral resolution. Commercial 2024-26: GelSight Mini at < $1k, integrated into IROS / RoboCup grippers; DIGIT open-source variant (Meta AI 2020 + community). Pair with [[Robotics/Tier3/sensor-families]] flexible / tactile entry.

Fiber Bragg gratings (FBG). Wavelength-encoded strain sensors written into optical fibre, embedded inside silicone bodies. Multiple gratings on one fibre give multi-point shape sensing without wiring routing. Used in surgical instruments (Hansen Medical / Auris/J&J Monarch lineage) for catheter shape feedback.

Conformal electronics. Rogers (Northwestern) “epidermal electronics” 2014; flexible Li-ion + thermoelectric + ionic batteries plus serpentine-routed CMOS, mounted on a stretchable elastomer membrane. Heart-rate, sweat-chemistry, neonatal monitoring. Bridges into medical wearables.

11. Control

Soft systems are nonlinear, hysteretic, viscoelastic, and effectively infinite-DoF. Five control paradigms in use:

1. PCC + feedback linearization. Treat the body as a chain of arcs (per §9), derive Jacobian J(q) mapping arc parameters q to tip pose, invert it for a desired Cartesian trajectory. Fast, interpretable, breaks down on contact and at large deformations.

2. Model predictive control (MPC) on reduced-order models. Della Santina, Bicchi, Rus 2020 — Cosserat or modal model rolled forward in a short horizon, optimize input pressures or tendon tensions. Sub-100 ms loops achievable on commodity hardware.

3. Sliding-mode / robust control. Treat un-modelled dynamics as bounded disturbance; design a high-gain robust controller. Works when actuator bandwidth ≫ disturbance bandwidth. Common in McKibben PAM control.

4. Learning-based control. Model-free RL: PPO / SAC on the actuator pressure as action, joint pose or task reward as objective. Sim2real domain randomization (Tobin, Akkaya, OpenAI 2017+) bridges the simulator gap. Notable: OpenAI Dactyl 2019 and Rubik’s Cube 2020 on the Shadow Dexterous Hand (pneumatic + tendon hybrid), trained entirely in simulation. Soft Robotics Inc mGripAI 2024 commercial — ML-guided pick selection on food-handling lines, the first SaaS-priced soft gripper “AI plan”.

5. Vision-language-action models. 2024-26 trend: Google PaLM-E, RT-X / RT-2, OpenVLA (Stanford 2024) generate continuous actions from natural-language instructions and camera observations. Adaptation to soft platforms is recent (DeepMind Aloha and Mobile Aloha extensions, Stanford Soft Aloha 2025) but inherits the same training recipe.

→ [[Robotics/Tier3/control-algorithms]] for the full control-theory map.

12. Manufacturing

Two-part silicone casting. Mix base + crosslinker, vacuum-degas, pour into a 3D-printed (PLA, ABS, SLA-resin) mould, cure at room temperature 4 h or 70 °C 30 min, demould. Two casts joined by uncured silicone create chambers; mould-design includes draft angles and ejector reliefs. The canonical “Harvard soft robotics toolkit” workflow (free downloadable mould files since 2014).

3D-printing.

  • Voxel8 (Lewis lab spin-out, MIT) — direct-ink-write multi-material elastomer + conductor, since 2015.
  • Carbon DLS (digital light synthesis) — CLIP photopolymerisation of EPU 40 / 41 / 45 elastomers, smooth surface finish, suitable for production grippers and footwear midsoles.
  • Stratasys PolyJet J826 / J835 / J850 — multi-material UV-cure droplet jetting; mixes Agilus30 (Shore 30A elastomer) with rigid VeroWhite to give graded-stiffness parts in one print.
  • UV-cure SLA / DLP — Formlabs Flexible 80A and Elastic 50A resins; faster, cheaper than PolyJet, single-material.
  • Inkbit — Vista (MIT spin-out) — vision-corrected multi-material jetting at higher accuracy; targeting medical and aerospace soft components 2024-26.

Soft lithography. Master mould in SU-8 photoresist on silicon, cast PDMS replica; the microfluidics standard, used for soft sensors and microscale soft robots.

Cast-and-cure with embedded electronics. Place a flex-PCB sensor or eGaIn channel in the mould before the silicone pour; release after cure. The dominant route for instrumenting soft fingers in 2026.

Multi-material additive is the most active manufacturing thread 2024-26. Stratasys, 3D Systems, Inkbit, and Mosaic (textile-integrated multi-extrusion) collectively make single-print soft-rigid robots realistic for production. Carbon’s industrial-scale DLS partnership with Soft Robotics Inc — public 2025 — is the bellwether for at-scale soft-gripper production.

13. Applications and commercial landscape

End-effectors.

  • Soft Robotics Inc (Bedford MA, founded 2013) — mGrip silicone-finger gripper (4-finger pneumatic, food + agri), mGripAI with vision and pick-planning (2023+ commercial). Lines at Tyson Foods, Driscoll’s berries, OSI Group.
  • Empire Robotics VERSABALL — granular-jamming sphere gripper (2014–2017 commercial, now defunct, the canonical jamming-based reference).
  • OnRobot Soft Gripper — silicone three-finger 3-pack, commercial 2018+, common on UR cobots.
  • Festo MultiChoiceGripper / DHEF — bistable finger gripper.

Wearables and exosuits.

  • Harvard Wyss / Walsh group ReWalk Soft Exosuit (assistive walking, ankle + hip torque via Bowden cables and inflatable bladders).
  • Roam Robotics — pneumatic-augmented skiing / industrial knee assist (consumer launch 2019; pivoted to industrial 2023-24).
  • Hyundai MEX / VEX / CEX — chairless exos for assembly-line workers, partially soft.
  • Pneubotics (Otherlab spinoff, 2014) — industrial-scale soft pneumatic arms, since pivoted into legacy-licensing.

Medical.

  • Capsule endoscopy. Olympus EndoCapsule, Medtronic PillCam, Capsovision CapsoCam, Anx Robotica NaviCam (with steerable variant), Ankon NaviCam Stomach System — passive and increasingly magnetically steered ingestibles.
  • Endovascular soft catheters. Steerable hydrogel-coated guidewires (Bionaut Labs, Magnebotix); commercial in pre-clinical / first-in-human trials 2024-26.
  • MIS instruments. Vincent Systems (Karlsruhe) prosthetic; J&J Monarch / Auris bronchoscope (formerly Hansen) — flexible robotic endoscopes with FBG shape sensing; Intuitive Ion robotic bronchoscope.

Search and rescue.

  • Octobot — Wehner, Truby, Whitesides, Lewis, Wood 2016 Nature, first fully untethered, electronics-free soft robot (chemical fuel + soft microfluidic logic).
  • MIT CSAIL soft snake / soft fish (SoFi) — Marchese, Onal, Rus.
  • Disney Imagineering soft characters — interactive park animatronics with soft skins for safety around guests (Project Kiwi, Groot 2023).

Agriculture.

  • Soft Robotics Inc gentle fruit picking (apples, berries, citrus) commercial since 2020.
  • Tortuga AgTech (Denver, CO) — strawberry-picking robot with soft end-effector, fleets deployed 2022+.
  • Advanced Farm Technologies (Davis, CA) — strawberry / apple harvesters, John Deere acquisition 2023.
  • Iron Ox, Naïo, FFRobotics — adjacent soft-end-effector plays.

→ See [[Robotics/Tier3/end-effectors-zoo]] for the full gripper catalogue (rigid + soft + jamming + suction + magnetic + electroadhesive).

HASEL actuator commercial scaling. Artimus Robotics shipped HVAC and industrial-actuator product lines through 2025; HASEL’s combination of electrical drive + hydraulic force is the closest practical replacement for solenoids and small pneumatic cylinders in soft systems.

LLM / VLA-controlled soft robots. Google PaLM-E (2023), RT-2 (2023), RT-X / Open X-Embodiment (2024), Stanford OpenVLA (2024). Mostly demonstrated on rigid platforms, but the same fine-tuning recipe applies to soft hands and continuum manipulators; first soft-platform VLA papers landed late 2024 and 2025-26 saw broader adoption.

Self-healing materials. Stanford Bao group + Otto group (Cambridge) polymers that re-form hydrogen-bonded or covalent networks after a cut, recovering 80–95 % of mechanical strength in minutes-hours. Pairs naturally with HASEL (self-healing dielectric).

Living / biohybrid robots. Xenobots — Bongard (Vermont) + Levin (Tufts) 2020-21 — assemblies of frog embryonic cells programmed by evolutionary search, exhibit self-locomotion and self-replication. 2024 “anthrobots” use adult human tracheal cells; therapeutic-delivery vision. Distinct from prior cell-on-elastomer biohybrid actuators (Park, Parker 2016 stingray).

Bio-inspired locomotion. Cornell, EPFL, Bristol BMI: caterpillar, octopus, jellyfish, larval-fish soft swimmers. Research demonstrators, sub-Reynolds-number swimming and crawling. Bristol Bio-Mechatronic Lab — caterpillar SMA + silicone; EPFL — soft fish; Cornell — soft octopus arm under reservoir computing control.

Untethered and sustainable power. Flexible Li-ion (Amprius, Enovix flexible packs), printed thermoelectrics, ionic-hydrogel batteries (Suo lab Harvard, 2024). Driving electric soft systems portably remains harder than driving pneumatic untethered with a CO₂ cartridge, but 2025-26 pouches now reach ≈ 150 Wh·kg⁻¹ at 30 % bending.

Soft-rigid hybrid as the practical engineering answer. Boston Dynamics Stretch (boxbot, 2022+) — soft vacuum-cup multi-finger end-effector on a rigid arm; Festo BionicSoftHand (pneumatic fingers + rigid wrist); KUKA LBR Med with elastic-jointed wrist; collaborative arms (Universal Robots, FANUC CRX, Kassow) shipping with optional soft tactile skin. Pure-soft systems remain in research; hybrids ship.

15. Failure modes and mitigations

  • Elastomer fatigue. Cyclic loading + UV + ozone exposure → crack initiation at strain concentrations (corners, fibre-bridge ends). Mitigations: fillet all internal corners, antioxidant additives (Smooth-On Sil-Poxy or commercial UV stabilisers), conservative strain limits (< 50 % of elastic limit).
  • Pneumatic leaks. Bond-line failure at silicone-silicone or silicone-fabric interfaces. Mitigations: plasma-treat surfaces, use rubber-cement primer (Sil-Poxy, DOW PR-1200), routine pressure-decay testing.
  • HV / DEA dielectric breakdown. Electrode hotspots cause carbonised pinhole through-faults. Mitigations: compliant electrodes with self-clearing (carbon grease evaporates; silver-nanowire fuses), HASEL liquid-dielectric self-healing.
  • Predictability under varying load. Hyperelastic stress-strain hysteresis differs at different speeds and after repeated cycles (Mullins effect). Mitigations: pre-conditioning (“Mullins cycling”) before deployment, closed-loop force/position with flexible sensing.
  • Cell-to-cell reproducibility. Cast-and-cure variation ±5–15 %; air bubbles, mould fill variation. Mitigations: vacuum degas + centrifugal degas, automated dispensing (Polygerinos lab process notes; commercial: Soft Robotics Inc’s proprietary cast lines).
  • Calibration drift. Soft sensors drift with temperature and creep. Mitigations: temperature-compensated capacitive sensing, periodic re-zeroing, GelSight (geometric, drift-free).

Conservative engineering rule: dual-redundant sensing where possible, soft-force limits in firmware, hard-stop in pneumatics (relief valve), and field-replaceable elastomer parts.

16. Cross-references

  • [[Robotics/end-effectors]] — broader gripper landscape
  • [[Robotics/Tier3/end-effectors-zoo]] — soft section and full taxonomy
  • [[Robotics/compliant-mechanisms]] — flexure-based compliance, the middle ground
  • [[Robotics/cable-driven-robots]] — workspace-scale tendon analogues
  • [[Robotics/kinematics-dh]] — rigid counterpart
  • [[Robotics/Tier3/control-algorithms]] — MPC, RL, sliding-mode formalisms
  • [[Robotics/Tier3/sensor-families]] — tactile, flexible, FBG sensors
  • [[Engineering/Tier3/polymers-taxonomy]] — full material chart
  • [[Engineering/microfluidics]] — fabrication overlap (soft-lithography, PDMS)

17. Citations

  • Whitesides, G. M. (2018) “Soft Robotics” Angew. Chem. Int. Ed. 57, 4258–4273 — field overview from the originator.
  • Rus, D. & Tolley, M. (2015) “Design, fabrication and control of soft robots” Nature 521, 467–475 — canonical review.
  • Trivedi, D., Rahn, C. D., Kier, W. M. & Walker, I. D. (2008) “Soft robotics: biological inspiration, state of the art, and future research” Applied Bionics and Biomechanics 5, 99–117.
  • Acome, E., Mitchell, S. K., Morrissey, T. G., Emmett, M. B., Benjamin, C., King, M., Radakovitz, M., Keplinger, C. (2018) “Hydraulically amplified self-healing electrostatic actuators with muscle-like performance” Science 359, 61–65 — HASEL.
  • Pelrine, R., Kornbluh, R., Pei, Q. & Joseph, J. (2000) “High-speed electrically actuated elastomers with strain greater than 100 %” Science 287, 836–839 — DEA founding paper.
  • Felton, S., Tolley, M., Demaine, E., Rus, D., Wood, R. (2014) “A method for building self-folding machines” Science 345, 644–646 — origami robotics.
  • Diller, E. & Sitti, M. (2014) “Three-dimensional programmable assembly by untethered magnetic robotic micro-grippers” Adv. Funct. Mater. 24, 4397–4404.
  • Kim, Y., Parada, G. A., Liu, S. & Zhao, X. (2019) “Ferromagnetic soft continuum robots” Sci. Robot. 4, eaax7329.
  • Wehner, M. et al. (2016) “An integrated design and fabrication strategy for entirely soft, autonomous robots” Nature 536, 451–455 — Octobot.
  • Webster, R. J. III & Jones, B. A. (2010) “Design and kinematic modeling of constant curvature continuum robots: a review” IJRR 29, 1661–1683 — PCC.
  • Thuruthel, T. G., Falotico, E., Renda, F., Laschi, C. & Iida, F. (2019) “Model-based reinforcement learning for closed-loop dynamic control of soft robotic manipulators” IEEE T-RO 35, 124–134.
  • Yang, G.-Z. & Whitesides, G. M. (2024) “Field-of-view review on soft robotics” Sci. Robot. — recent state-of-the-field.
  • Polygerinos, P., Wang, Z., Galloway, K. C., Wood, R. J. & Walsh, C. J. (2015) “Soft robotic glove for combined assistance and at-home rehabilitation” Rob. Auton. Syst. 73, 135–143.
  • 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” IJRR 33, 768–782.
  • OpenAI et al. (2019) “Solving Rubik’s cube with a robot hand” arXiv:1910.07113; OpenAI et al. (2018) “Learning dexterous in-hand manipulation” — Dactyl.
  • Bongard, J. & Levin, M. (2021) “Living concrete: A biological perspective on robotics” Cell — xenobot programme; Gumuskaya, G. et al. (2024) “Motile living biobots self-construct from adult human somatic progenitor seed cells” Adv. Sci. — anthrobots.
  • Soft Robotics Inc, mGripAI product literature (2024 commercial brochure).

Graph View