Soft Robotics — Continuum Bodies, Soft Actuators, Compliant Manipulation

Robots built primarily from compliant, deformable materials (elastomers, gels, fabrics, fiber-reinforced composites) rather than rigid links and discrete joints. The promise is intrinsic safety through low mechanical impedance, adaptability via passive shape conformance, biomimetic morphology that exploits material intelligence, and access to confined or unstructured spaces (lungs, intestines, rubble piles) that rigid manipulators cannot reach. The cost is everything classical robotics took for granted — there is no closed-form forward kinematics, no clean inverse-dynamics, no rigid-body Jacobian — so the field has built its own modeling (Cosserat rods, piecewise-constant curvature, FEM, generalized coordinates), actuation taxonomy (pneumatic, dielectric-elastomer, HASEL, SMA/SMP, tendon, jamming, LCE, magneto-active), and sensing stack (capacitive skins, optical-fiber, embedded resistive). The 2010-2026 wave saw soft robotics graduate from Whitesides-lab octopi to commercial food-grade grippers (Soft Robotics Inc. mGrip, RightHand RightPick3) and FDA-cleared surgical instruments (Auris/J&J Monarch flex robot).

See also

• soft-robotics
• compliant-mechanisms
• cable-driven-robots
• end-effectors
• impedance-control
• surgical-robotics
• prosthetics-rehabilitation
• sensors-force-tactile

1. At a glance

A soft robot has a body whose elastic modulus E is comparable to that of biological tissue (10 kPa to 10 MPa), as opposed to the 10-200 GPa of aluminum / steel / engineering plastics in rigid robots. That four-to-six order-of-magnitude reduction in stiffness is the design primitive. From it follow:

• Safe interaction by construction. Collision energy scales with effective stiffness; a 1 MPa elastomer finger striking a human at 1 m/s transfers ~$10^{-3}$ of the kinetic energy of an aluminum finger at the same speed. Risk-assessment under ISO/TS 15066 power-and-force-limiting collapses to a trivial calculation.
• Passive shape adaptation. A soft gripper does not need a 3D model of the object — it wraps. The same mGrip closes around a strawberry, a pear, or a steak with no replanning, no calibration, no force sensor.
• Confined-space access. A continuum surgical robot threads a 3 mm diameter through a 4 mm bronchus; a 6-DoF rigid manipulator cannot.
• Biomimetic morphology. Octopus arms, elephant trunks, snake bodies, caterpillar locomotion — millions of years of evolution use compliance, not rigidity. Soft robots reverse-engineer those solutions.

Three terms are routinely confused. Continuum robots have a structurally compliant backbone but use discrete actuators (typically tendon-driven, e.g., Webster’s concentric-tube robots). Hyper-redundant robots have many rigid links and many DoF (snake robots like CMU’s modular snake — Howie Choset’s group). Soft robots are continuum systems built from intrinsically compliant materials — the body and the actuator are one. The taxonomies blur at the edges (Festo’s BionicSoftArm has rigid couplings between pneumatic chambers, RBO Hand has rigid bones inside silicone fingers) — what matters is which mechanical primitive dominates behavior.

Where this sits. End-effectors is where most commercial soft robotics lives in 2026 — grippers for food + e-commerce + agriculture; compliant mechanisms is the rigid-flexure cousin where lumped flexures replace pin joints; cable-driven systems are the bridge from tendons-on-rigid to tendons-in-soft; impedance control is the natural control formalism for a body that is a spring; surgical is the highest-revenue clinical application; prosthetics is where soft hands compete with rigid multi-articulating hands.

First ask. What is the task envelope? Grasping unknown objects → soft pneumatic gripper. Navigating a lumen → tendon-driven continuum. Locomotion over rubble → fluidic-elastomer body. Wearable assistance → SMA / pneumatic muscle. What’s the duty cycle? High-cycle pick-and-place → silicone fatigue is real, 10 million cycles is a target. What’s the bandwidth? Pneumatic ≤ 5 Hz, DEA ≥ 100 Hz, HASEL up to 50 Hz. What’s the energy density? Pneumatic at 1 bar excels in actuation strain but needs an off-board compressor; DEAs need 4-10 kV but can be untethered with charge-pump electronics. Single-use or repeated? Soft sensors degrade — silicone-embedded carbon-black grids drift after ~10⁴ cycles.

2. First principles

2.1 Stiffness, compliance, and effective impedance

A robot’s interaction stiffness $K$ at the end-effector decomposes into:

$\frac{1}{K_{\text{eff}}}=\frac{1}{K_{\text{ctrl}}}+\frac{1}{K_{\text{mech}}}$

Rigid robots: $K_{\text{mech}}\gg K_{\text{ctrl}}$, so behavior is set by control gain. Soft robots: $K_{\text{mech}}\ll K_{\text{ctrl}}$, so behavior is dominated by material — and a velocity-controlled or position-controlled soft robot is automatically force-limited by its own stiffness. This is the formal reason a Soft Robotics Inc. mGrip cannot crush a strawberry: even at full pressure (~70 kPa), the silicone finger Young’s modulus (~1 MPa) sets a tip force ceiling around 5 N regardless of command.

2.2 Continuum modeling — three families

Piecewise Constant Curvature (PCC, Webster + Jones 2010). Approximate a continuum section as an arc with curvature $\kappa$ and length $\ell$. Two parameters per section in 2D, three in 3D (curvature, plane-of-bending angle, length). Forward kinematics reduces to a chain of arc-to-arc transforms. Closed-form, fast, but assumes negligible torsion and uniform bending — fails for sections under combined load.

Cosserat rod theory. A continuous beam parameterized by arc length $s\in [0,L]$ with position $p(s)\in {\mathbb{R}}^3$ and orientation $R(s)\in SO(3)$. Governing ODE:

$p^{\prime }(s)=R(s)v(s),R^{\prime }(s)=R(s)[\omega (s){]}_{\times }$

with internal force $n(s)$ and moment $m(s)$ satisfying

$n^{\prime }(s)+f(s)=0,m^{\prime }(s)+p^{\prime }(s)\times n(s)+l(s)=0$

where $v,\omega$ are linear/angular strain rates and $f,l$ are distributed external loads. Solves for shape under load including torsion and shear; basis for OCTArm (Walker, Clemson) and most concentric-tube robot kinematics (Dupont, BCH).

Finite Element Method (FEM). Discretize the elastomer body into tetrahedral elements with hyperelastic constitutive model (Neo-Hookean, Ogden, Mooney-Rivlin). Solve nonlinear quasi-static or dynamic equations. The dominant tool is SOFA (Simulation Open Framework Architecture, INRIA — Cotin, Duriez), now embedded in the Soft Robotics Plugin used by most academic FEM-controlled soft robots. ChainQueen (Hu et al., MIT 2019) introduced differentiable MPM (Material Point Method) for soft bodies, enabling gradient-based design + control.

Generalized coordinates / reduced models (Della Santina, Calogero — TU Delft / Pisa, 2023). Approximate the infinite-DoF body with a finite set of mode shapes $\{{\varphi }_i(s)\}$. Plug into Lagrangian; recover a finite-dim ODE that runs at 1 kHz on CPU. Standard for model-based control of pneumatic continuum arms.

2.3 Actuation taxonomy

Pneumatic Networks (PneuNets) — Whitesides Lab, Harvard, 2011 (Ilievski, Mazzeo, Shepherd, Chen, Whitesides — “Soft Robotics for Chemists”). A monolithic silicone (Ecoflex 00-30 or Dragon Skin 10) molded with internal channel networks. Pressurization differentially inflates chambers; one face is constrained (paper, fabric, stiffer silicone backing) producing a programmed curvature. Bending angle as a function of pressure for a PneuNet finger of length $L$, channel height $h$, wall thickness $t$:

$\theta (P)\approx \frac{PL}{Et}\cdot f(\text{geometry})$

PneuNets are the most-replicated soft-robot primitive in academia (every soft-robotics course makes one). Commercial descendants: Soft Robotics Inc. (Whitesides spin-out 2013).

Fluidic Elastomer Actuators (FEA). Generalization of PneuNets to arbitrary geometry — multi-axis bending, twisting, expansion. Includes McKibben pneumatic muscles (1950s, originally for prosthetics by Joseph McKibben) — a braided sleeve over an inflatable bladder that contracts axially when pressurized, producing 25-30% strain and very high force-to-weight (~1 kN/kg). Used in Festo BionicSoftHand, Bristol Robotics Lab arms, and Suit X exoskeletons.

Soft Pneumatic Linear (SPL) — bellows-style pneumatic actuators that elongate or contract linearly. Festo’s DSMP modular actuators are the commercial reference.

Dielectric Elastomer Actuators (DEA) — Pelrine, SRI International, 2000 (“High-speed electrically actuated elastomers with strain greater than 100%”). A thin elastomer film (acrylic VHB 4910 from 3M, silicone, or natural rubber) sandwiched between compliant carbon-grease or carbon-black electrodes. Apply voltage $V$ across thickness $z$; Maxwell stress compresses the film:

${\sigma }_{\text{Maxwell}}={\epsilon }_0{\epsilon }_r{\left(\frac{V}{z}\right)}^2$

For VHB 4910: ${\epsilon }_r\approx 4.7$, breakdown ~$200$ V/μm, achievable area strain ~100-200%. Operating voltages 3-10 kV. DEAs are the highest-bandwidth soft actuator (>100 Hz routine, kHz demonstrated) but the high-voltage requirement and pre-stretch fixturing keep them in research labs.

HASEL — Hydraulically Amplified Self-Healing Electrostatic Actuators (Keplinger, U. Colorado Boulder, 2018) (“Hydraulically amplified self-healing electrostatic transducers harnessing the Hertzian dipole” — Science). A liquid dielectric (vegetable oil, silicone oil) inside a flexible polymer pouch with electrodes on opposing faces. Apply voltage; electrodes zip together via electrostatic attraction, displacing the oil and producing actuator stroke. Advantages over DEA: self-healing (a dielectric breakdown just re-fills with oil), lower mechanical pre-stress requirements, scalable from grams to tens of kg. Spin-out: Artimus Robotics (Boulder, 2018). Used in Disney Imagineering animatronics.

Shape Memory Alloys (SMA — Nitinol). Ni-Ti alloy with two crystallographic phases (austenite, martensite) separated by ~70 °C transformation temperature. Cold-deformed martensite returns to austenite shape on heating (resistive Joule heating, typically 1-5 A through a 250 μm wire). Strain ~4-8%, force/weight excellent (~600 MPa stress), but bandwidth is thermal — ~1 Hz limited by cooling. Festo BionicCobot uses Nitinol bias springs.

Shape Memory Polymers (SMP). Glass-transition-based shape memory — heat above $T_g$, deform, cool below $T_g$ to lock, reheat to recover. Slower than SMA, lower force, but cheaper and 3D-printable (NinjaTek SemiFlex variants, Mitsubishi Diaplex). Used in deployable structures and one-shot grippers.

Magneto-Active Soft Materials (Lum, Boyvat, Zhang — MIT, 2021). Iron / NdFeB microparticles embedded in silicone, programmed with a spatial magnetization pattern. An external field $B$ produces a torque $\tau =m\times B$ per unit volume that distributes through the body. Magnetic actuation is wireless and remote — ideal for in-vivo applications. Spin-outs: Multi-Scale Robotics Lab (Nelson, ETH) magnetic capsule endoscopes; Bionaut Labs intracranial drug delivery.

Liquid Crystal Elastomers (LCE — Ware, U. Dallas; White, AFRL; Broer, TU Eindhoven). Polymer networks with embedded mesogens that align along a director. Heating disrupts the alignment, contracting the LCE up to 40%. Programmable via photo-alignment patterning. Bandwidth ~Hz (thermal); strain very high; force moderate. Heliograph robotics demos (Ware 2022) showed sequential-folding origami.

Tendon-Driven Soft. Rigid actuators (DC motors, McKibbens) drive Bowden cables routed through a soft body — Festo BionicSoftHand, RBO Hand 2/3 (Brock — TU Berlin), Pisa/IIT SoftHand (Bicchi — IIT Genova / Università di Pisa). The compliance is structural; the actuator is conventional. Easiest path to commercial deployment.

Granular / Layer / Fiber Jamming. Brown et al., Cornell + iRobot, 2010 (“Universal robotic gripper based on jamming of granular material” — PNAS). A balloon filled with coffee grounds (or glass beads, or coarse sand) drapes over an object; vacuum applied; the granular medium locks into a rigid shape conformed to the object. Universal grasp. Empire Robotics Versaball was the commercial product (2012-2016; company folded). Pisa/IIT “Pisa/IIT Hand” series uses layer-jamming for tunable finger stiffness; ETH ASL fiber-jamming surgical instruments achieve programmable rigidity.

Pneumatic Stiffening. Pre-pressurized elastomer chambers behave as stiff beams; depressurized chambers behave compliant. STIFF-FLOP (Cianchetti, Menciassi — Scuola Sant’Anna, Pisa, 2014) used this for surgical-arm variable-stiffness operation.

2.4 Sensing

Soft bodies cannot use rigid-robot sensors (rotary encoders presume rigid links). The soft-sensing taxonomy:

• Embedded capacitive skin — two compliant conductive layers separated by dielectric elastomer. Deformation changes capacitance. StretchSense (NZ), Bioservo SEM gloves, Pressure Profile Systems.
• Resistive grids — carbon-black or CNT-doped silicone changes resistance with strain. Cheap, drifty, hysteresis-prone. Common in DIY soft sensors.
• Optical fiber sensors — fiber Bragg gratings (FBG) embedded in silicone read curvature via wavelength shift; multi-core fibers (FBGS DTG-LBL) reconstruct full 3D shape. Surgical-robotics standard (Intuitive Da Vinci, Hansen Medical Magellan).
• EGaIn liquid-metal channels — Whitesides + Dickey eutectic-gallium-indium microchannels in PDMS produce stretchable, hysteresis-free strain gauges.
• Inductive shape reconstruction — Felton, Whitesides MIT 2014 — coils embedded in elastomer; mutual inductance encodes shape.
• IMU arrays — distribute MEMS IMUs along a continuum body; fuse via Kalman filter to estimate shape. Cheap, available off-the-shelf, but inaccurate near gravity-aligned poses.
• Vision-based proprioception — external or wrist-mounted camera observes fiducials painted on the body; deep network regresses configuration. ALOHA-style approach extended to soft arms by Della Santina’s group.

2.5 Control formalisms

Soft robotics is dominated by impedance / admittance control because the body is a spring; commanding a position is the same as commanding a force given the body’s compliance. Adding to that:

• Model Predictive Control with learned dynamics. Train a neural-net dynamics model from real or sim rollouts; run MPC on the learned model. Standard in 2024+ — Yip Lab UCSD soft-robot state estimation, Berkeley FlexLab.
• Hybrid MPC + RL. Use MPC for tracking; RL for terminal cost / value function. Closes the gap between model fidelity and task performance.
• Adaptive control. Lyapunov-based parameter estimation for unknown stiffness and damping. Pratheek Bagivalu et al., U. Edinburgh 2023.
• Direct policy learning (RL). End-to-end policy from observation to actuator pressure. Compute-expensive but no model needed — Marchese, Komorowski, Rus, MIT 2016 (octopus arm); used in most modern soft-robotic-arm reach-and-grasp demos.

3. Practical math — actuator design

3.1 PneuNet curvature design

For a PneuNet with $N$ chambers of width $w$, height $h$, wall thickness $t$, separated by walls of length $\ell$, the per-chamber expansion strain at pressure $P$ for an Ecoflex 00-30 ($E\approx 0.07$ MPa) wall:

$ϵ\approx \frac{P}{E}\cdot \frac{h}{2t}$

Each chamber’s expansion bends the inextensible bottom layer. For $P=50$ kPa, $h=4$ mm, $t=1$ mm: $ϵ\approx 1.4$. The finger curves to roughly $\theta =N\cdot ϵ\cdot w/\ell$ at full inflation. Standard 4-chamber design: $N=4$, $w=10$ mm, $\ell =4$ mm $\Rightarrow \theta \approx 14$ rad $\approx 800°$ — practically limited by self-contact and material elasticity to ~360°.

3.2 McKibben muscle force vs contraction

McKibben muscle of nominal length $L_0$, initial diameter $D_0$, braid angle ${\alpha }_0$ (typically 25-30°), contraction ratio $\eta =(L_0-L)/L_0$:

$F(P,\eta )=\frac{P{D}_0^2\pi }{4{\sin }^2{\alpha }_0}\left[3(1-\eta {)}^2{\cos }^2{\alpha }_0-{\sin }^2{\alpha }_0\right]$

At $P=5$ bar (500 kPa), $D_0=20$ mm, ${\alpha }_0=30°$: peak force at $\eta =0$ is $\approx 1.2$ kN, drops to zero at $\eta \approx 0.3$. Bandwidth limited by pneumatic dynamics — 2-5 Hz for direct-drive, higher with proportional valves close-coupled.

3.3 DEA strain vs voltage

Pre-stretched VHB 4910 (${\lambda }_x={\lambda }_y=4$ biaxial pre-stretch, thickness $z=50$ μm post-stretch), applied voltage $V$:

${ϵ}_{\text{thickness}}\approx -\frac{{\epsilon }_0{\epsilon }_r}{Y}{\left(\frac{V}{z}\right)}^2$

For VHB 4910 ($Y\approx 0.3$ MPa effective at 4× pre-stretch, ${\epsilon }_r=4.7$): $V=5$ kV gives field $E=100$ V/μm, Maxwell stress 0.42 MPa, thickness strain ~30%, lateral strain ~15% (incompressible). Breakdown around $E=200$ V/μm; operate at 50% margin.

3.4 HASEL — Peano-HASEL force

Peano-HASEL (chain of rectangular pouches): the force at hold is

$F_{\text{hold}}=\frac{{\epsilon }_0{\epsilon }_r{V}^2}{2z^2}\cdot A_{\text{electrode}}$

For a 5 cm × 1 cm electrode at $V=8$ kV through $z=20$ μm dielectric (${\epsilon }_r=3$ for vegetable oil): $F\approx 8$ N. Stroke ~15% of pouch length. Achievable lifting weight per gram of HASEL is ~200× muscle.

3.5 SMA wire current vs response time

Nitinol wire of diameter $d$, length $L$, resistance per length ${\rho }_e$:

$P_{\text{Joule}}=I^2{\rho }_{e}L$

Heat capacity per length $C_L=\rho m{c}_p(\pi d^2/4)$. Time to reach transformation temperature $\Delta T=T_A-T_{\text{ambient}}$ (Nitinol $T_A\approx 70$ °C, $\Delta T=50$ K):

${\tau }_{\text{heat}}\approx \frac{C_L\Delta T}{I^2{\rho }_e}$

For 250 μm Flexinol wire (${\rho }_e\approx 20$ Ω/m, $C_L\approx 1.7\cdot 10^{-3}$ J/(K·m)): $I=1$ A gives heating time ~85 ms. Cooling time (natural convection) ~800 ms — sets the 1 Hz bandwidth.

4. Design heuristics

• Pick the elastomer for fatigue, not just hardness. Ecoflex 00-30 (Shore 30A, 90% elongation at break) is the academic default but fatigues in <10⁵ cycles. Dragon Skin 10 (Shore 10A, 1000% elongation) lasts longer. Smooth-On Sorta-Clear for transparent demos. Wacker Elastosil M-4601 for industrial molds.
• Mold release matters more than you think. Inhibition of platinum-cure silicone by sulfur, amines, or 3D-printed PLA is the #1 cause of “my robot didn’t cure.” Use a barrier coating (Inhibit-X) or print molds in tin-cure resin.
• Bond reinforcement layers properly. Silicone-to-fabric requires Smooth-On Sil-Poxy or plasma treatment. A poorly bonded paper layer delaminates after 1000 cycles.
• Route pneumatic tubing with strain relief. A direct tube into a 1 mm port shears off after a few inflations. Use barbed fittings with epoxy potting.
• Pressure cap your controller in firmware. Soft robots burst — once. A regulator failure at 100 psi will blow your finger across the lab. Hard-code a software limit at the design pressure × 1.5 and an over-pressure interrupt.
• Account for elastomer creep. A silicone finger held at constant pressure relaxes 5-15% over the first hour. Calibrate after warmup.
• DEA + HASEL safety: every conductor is high-voltage. Use kV-rated wire, encapsulate exposed electrodes in dielectric oil, isolate operator with grounded enclosure.
• Embed sensors before casting, not after. Drilling into cured silicone destroys it; mold around the sensor.
• Model + experiment in parallel. FEM gives intuition; the silicone batch you ordered has a different modulus. Always measure tensile + cyclic samples from the actual batch.
• Plan for replacement. Soft fingers wear out. Design quick-change mounting (Festo SubD, Soft Robotics Inc. mGrip quick-release). A 1-minute finger swap is worth its weight.

5. Components & sourcing — platforms and materials

5.1 Commercial soft-gripper systems

Vendor	Product	Tech	Application
Soft Robotics Inc. (Bedford MA)	mGrip, SuperPick	Pneumatic FEA fingers	Food primary, e-commerce secondary
RightHand Robotics (Somerville MA)	RightPick3	Hybrid pneumatic + vacuum	E-commerce piece-picking
Festo (Esslingen DE)	BionicSoftHand, BionicCobot, BionicSoftArm	Pneumatic + tendon	Demonstrator + research
Schunk (Lauffen DE)	EGS / EGL with soft fingertips	Hybrid rigid + soft	Industrial pick-and-place
OnRobot (Odense DK)	Soft Gripper	Pneumatic FEA	Cobot end-of-arm
Piab (Täby SE)	piSOFTGRIP	Vacuum + soft cups	Food + delicate parts
Yaskawa Motoman	SoftFinger end-effectors	Pneumatic	Automotive trim
Bota Systems	SoftRobot grippers	Pneumatic	Lab automation

5.2 Research platforms

Platform	Lab	Notes
Octobot (Wehner, Truby, Shepherd, Wood — Harvard 2016)	Whitesides + Lewis Labs	First fully autonomous untethered soft robot; chemical fuel + microfluidic logic
RBO Hand 2 / 3 (Deimel, Brock — TU Berlin)	Robotics + Bio Lab	Pneumatic continuum fingers, opposable thumb
Pisa/IIT SoftHand (Catalano, Grioli, Bicchi — IIT Genova / U. Pisa)	Bicchi lab	Underactuated tendon-driven 5-finger, 19 joints / 1 motor
Yale OpenHand (Ma, Dollar — Yale GRAB Lab)	Dollar lab	Open-source 3D-printable underactuated hands
ROBEL D’Claw / D’Hand (Ahn, Yu, Levine et al. — Berkeley 2019)	Levine lab	Low-cost manipulation testbed for RL
OCTArm (Walker — Clemson)	Walker	9-section continuum manipulator
STIFF-FLOP (Cianchetti, Menciassi — Scuola Sant’Anna, Pisa)	Menciassi lab	Variable-stiffness surgical arm
MIT Octopus arm (Marchese, Komorowski, Rus 2016)	Rus CSAIL	RL-controlled FEA arm
Festo BionicSoftArm	Festo Bionic Learning Network	13-DoF pneumatic continuum, commercial-grade
ARMM (Vasilescu, Rus — MIT)	Rus	Cable-driven elephant-trunk

5.3 Materials and consumables

Item	Vendor	Notes
Ecoflex 00-30	Smooth-On	Soft, easy, 1:1 platinum-cure, ~$30/kg
Dragon Skin 10 / 30	Smooth-On	Tougher, higher tear strength
Sorta-Clear 18 / 37	Smooth-On	Transparent for sensor visualization
Wacker Elastosil M-4601	Wacker Chemie	Industrial silicone, thermal cure
VHB 4910 / 4905	3M	Acrylic dielectric elastomer, DEA reference
3M ScotchPlate	3M	Conductive transfer adhesive
Carbon-black grease (MG Chemicals 846)	MG	Compliant DEA electrode
Liquid metal eGaIn	Sigma-Aldrich	Stretchable conductor for sensing
Nitinol wire 250 / 375 μm	Dynalloy Flexinol, Memry	SMA actuation
Spectra / Dyneema 100 lb test	Various	Cable-drive tendons, low creep
Bowden cable Teflon liner	Capricorn	Low-friction tendon routing
Pneumatic mini solenoid valves	SMC SY3000, Festo VUVG	200 Hz switching
Proportional pressure regulators	SMC ITV2050, Festo VPPM	1-10 bar continuous
Carbon nanotube ink	C3Nano, Cabot ENERMAX	Conductive elastomer doping

5.4 High-voltage electronics (DEA / HASEL)

Component	Vendor	Notes
EMCO Q-series	XP EMCO	5-10 kV DC-DC, charge-pump
Trek 5/80 amplifier	Trek (now Apex)	Lab-grade DEA driver, 5 kV / 80 mA
Matsusada AP series	Matsusada	Custom HV power supplies
Pico Electronics HC	Pico	Miniature HV converters for untethered
Artimus Robotics drivers	Artimus	HASEL-tuned drivers (commercial)

5.5 Modeling and simulation

Tool	License	Use
SOFA + Soft Robotics Plugin	LGPL	FEM dynamics, INRIA
Abaqus / Ansys	Commercial	Hyperelastic FEM, batch design
ChainQueen (MIT 2019)	MIT	Differentiable MPM for soft
ARMM-Sim (Della Santina)	Academic	Reduced-coordinate continuum
MuJoCo (DeepMind 2021)	Apache 2.0	Soft-body via composite tendons + FEM (limited)
Genesis (CMU 2024)	Apache 2.0	Unified rigid + soft + fluid
PyElastica	MIT	Cosserat rod simulation (Gazzola lab — UIUC)

6. Reference data

6.1 Actuator comparison

Actuator	Strain	Stress	Bandwidth	Energy density	Voltage
Skeletal muscle	20-40%	0.3 MPa	10 Hz	40 J/kg	—
PneuNet	100-300%	0.05-0.3 MPa	1-5 Hz	30 J/kg	—
McKibben	25-30%	0.5-1.0 MPa	2-5 Hz	50 J/kg	—
DEA (VHB 4910)	100-200%	0.1-0.5 MPa	10-1000 Hz	10 J/kg	3-10 kV
HASEL	10-25%	0.2 MPa	10-50 Hz	20 J/kg	5-10 kV
SMA (Nitinol)	4-8%	200-700 MPa	1 Hz	5000 J/kg	low (resistive heat)
SMP	10-100%	1-10 MPa	<0.1 Hz	100 J/kg	low
LCE	30-50%	0.5 MPa	0.1-1 Hz	50 J/kg	low (thermal) or photonic
Magneto-active	10-40%	0.1 MPa	1-100 Hz	<1 J/kg (external field)	—
Granular jamming	binary	50-200 kPa hold	<1 Hz	n/a	—

6.2 Notable commercial soft-robotics deployments

Year	Company	Deployment	Significance
2015	Soft Robotics Inc.	First commercial mGrip in food packaging	Whitesides spin-out productizes PneuNets
2017	Empire Robotics	Versaball at Boeing assembly	Granular jamming gripper; company closed 2017
2018	Festo BionicCobot	Industrial trade-show demo	Pneumatic 7-DoF compliant cobot
2018	Auris Health Monarch flex robot	FDA-cleared bronchoscopy	Continuum surgical robot; acquired by J&J 2019 ($3.4B)
2019	RightHand Robotics RightPick2	E-commerce piece-picking deployments	Hybrid suction + soft fingertips
2020	Soft Robotics Inc. SuperPick	Food + meat packaging at scale	Validates soft + AI vision for food
2021	Pisa/IIT SoftHand for prosthetics	CE-marked prosthetic hand	Underactuated tendon hand for amputees
2022	Artimus Robotics HASEL grippers	Lab automation pilots	First commercial HASEL product
2023	Endiatx PillBot	Phase 1 ingestible swimmer trial	Magnetically actuated stomach-imaging robot
2024	Festo BionicSoftArm 2.0	Cobot product line	Pneumatic continuum at industrial reliability
2024	Helix-Robotics surgical flex	Phase 2 trials	Continuum colonoscopy steering
2025	mGrip + JBT Foodtech integration	Major poultry processor	Soft robotics in primary meat processing

6.3 Decision matrix — actuator choice

                  Pneumatic  DEA  HASEL  SMA  Tendon  Jamming
High bandwidth      —        ✓    ✓      —     —       —
High strain         ✓        ✓    —      —     —       —
High force          ✓        —    ✓      ✓     ✓       ✓ (hold)
Untethered          —        △    △      ✓     ✓       —
Low voltage         ✓        —    —      ✓     ✓       ✓
Surgical (MRI)      △        —    —      ✓     ✓       △
Industrial cycles   ✓        △    △      △     ✓       △
Cheap parts         ✓        ✓    ✓      ✓     ✓       ✓

6.4 Continuum-robot kinematic taxonomy

                          DoF        Actuation         Application
Concentric tube           3N         Tube rotation     Neurosurgery, eye
Tendon-driven backbone    2-3/sect   Cable             Bronchoscopy, ENT
PneuNet finger            1/section  Pressure          Gripping
FEA arm                   2/section  Pressure          Octopus mimicry
Pneumatic muscle          1          Pressure          Limb + exo
HASEL stack               1          Voltage           Lightweight lifting
SMA spring                1          Current           Compact / wearable
Magneto-active            many       External B-field  In-vivo navigation

7. Failure modes & debugging

• Bursting / blow-out. First failure of any inflatable. Cause: thin wall, stress concentration, fatigue, over-pressure. Detect: hairline cracks, audible hiss. Fix: increase wall thickness, fillet sharp corners, reduce duty cycle pressure, fabric-reinforce.
• Creep / set. Silicone deforms permanently under sustained load. Detect: rest length changes. Fix: lower nominal pressure, choose harder shore, design for periodic depressurization.
• Delamination. Bonded layers separate. Detect: bulge between layers. Fix: better surface prep (plasma, primer), redesign for shear-only bonds.
• Dielectric breakdown (DEA). Local field exceeds breakdown; carbonized failure. Detect: pinholes, smoke. Fix: lower voltage, increase thickness, mode-bias toward HASEL (self-healing).
• Sensor drift. Resistive carbon-black sensor drifts after cycling. Detect: zero-load resistance changes over time. Fix: switch to capacitive or optical, frequent re-calibration, hysteresis compensation.
• Pneumatic latency. Long tubing introduces transport delay. Detect: phase lag in pressure response. Fix: shorter tubing, larger ID near actuator, on-board valves.
• SMA overheating. Continuous current exceeds wire dissipation. Detect: wire glows, breakdown. Fix: duty-cycle limit, current sensing, cooling fins.
• Jamming reset failure. Granular bag does not return to soft. Cause: trapped vacuum, packed grains, debris. Fix: positive-pressure release pulse, vibration during release.
• Magnetic-actuation drift. Field gradient changes due to ambient ferromagnets (rebar, MRI residual). Fix: shielding, environmental calibration, closed-loop visual servoing.
• Mold-cure inhibition. Silicone fails to cure adjacent to certain substrates. Cause: sulfur (latex gloves), amines (some 3D-print resins), copper, tin. Fix: barrier coatings, inert mold materials, tin-cure silicone if compatible.
• High-voltage arc-over. Surface tracking around HV electrodes. Fix: encapsulation in dielectric oil or silicone, increase creepage distance, conformal coating.
• Tendon abrasion. Spectra / Dyneema cables fray in Bowden tubes. Fix: PTFE liner, larger bend radius, periodic replacement.

8. Case studies

8.1 Octobot — first fully soft autonomous robot (Wehner et al., Harvard 2016)

Science Robotics cover paper. Researchers in Whitesides + Lewis (3D-printing) labs at Harvard built an octopus-shaped soft robot powered by an onboard chemical reaction: H₂O₂ decomposes catalytically over platinum, producing O₂ gas that pressurizes microfluidic channels. A microfluidic oscillator logic gate alternates pressure between the eight arms — no batteries, no wires, no rigid components. Demonstrated 8-minute autonomous “swimming” motion. Significance: proved the entire stack (power + control + actuation) could be soft.

8.2 Soft Robotics Inc. mGrip in food processing (2015-2026)

George Whitesides’ Harvard work spun out in 2013 as Soft Robotics Inc. (Bedford MA). Their mGrip product is a 2-, 3-, or 4-finger pneumatic gripper with FEA-style silicone fingers (proprietary tubing geometry). At Tyson, JBT, and major poultry processors, mGrip arrays handle raw chicken, fish fillets, dough, and produce — historically impossible for rigid grippers due to crush-damage risk. Throughput: 60-100 pieces/min per arm. 2026 deployments: tens of thousands of fingers in service. Acquired by Soft Robotics (PE-backed) and re-organized 2023; product lines remain.

8.3 Auris Health Monarch + J&J ($3.4B acquisition 2019)

Frederic Moll (Intuitive Surgical co-founder) founded Auris Health in 2007 to build steerable continuum endoscopes. The Monarch platform uses two concentric flexible scopes (outer sheath + inner leader) driven by external roller wheels; the leader is tendon-actuated. FDA-cleared 2018 for bronchoscopy (lung cancer biopsy of peripheral nodules previously unreachable). J&J acquired Auris in 2019 for $3.4B, the largest soft-/continuum-robotics exit in history. Subsequent Auris Ion robotic-assisted bronchoscopy system (2019, FDA cleared) competes directly with Intuitive’s da Vinci-Ion. Validates continuum-robot business model.

8.4 RBO Hand 3 (Brock — TU Berlin, 2020)

Oliver Brock’s lab at TU Berlin developed three generations of “Robotics + Biology” (RBO) Hands — pneumatic anthropomorphic hands using PneuFlex actuators (PneuNet-derived fingers with embedded fabric). RBO Hand 3 has an opposable thumb, abduction-adduction at the MCP joints, and 8 chambers driven by 8 proportional valves. Demonstrated power and precision grasps on > 100 objects with no online sensing — purely passive shape conformance + open-loop pressure trajectories. Cited as the strongest demonstration that morphology can replace planning.

8.5 Pisa/IIT SoftHand prosthetic (Bicchi — IIT Genova / U. Pisa, 2014-2021)

Antonio Bicchi’s “synergy-based” approach — a 19-joint, 5-fingered hand driven by a single motor through an underactuated tendon network. The tendon routing implements the “first postural synergy” (Santello, Flanders, Soechting 1998 — humans grasp using a low-dimensional manifold of finger postures). Result: a single signal commands a biomechanically plausible grasp; the hand passively conforms to object shape. CE-marked prosthetic version (Prensilia / IIT spin-out) sold for ~€$25k vs €$50-80k for multi-articulating hands like Ottobock bebionic. Validated in 30+ amputee subjects through 2023.

8.6 STIFF-FLOP (Cianchetti, Menciassi — Scuola Sant’Anna, EU Horizon 2020)

A surgical continuum manipulator with variable stiffness — the same arm soft for navigation through anatomy, stiff for tissue manipulation. Stiffness via granular-jamming chambers along the body. EU collaboration culminating 2014-2018. Clinical translation slower than commercial flex-robots (Auris) but established the methodology.

8.7 Endiatx PillBot ingestible swimmer (2023-2026)

Endiatx (Hayward CA) developed PillBot — a 13×30 mm magnetically actuated capsule with two micro-impellers, a 4K camera, and wireless video, ingested for upper-GI imaging. The capsule navigates the stomach under external magnetic field control by a physician. FDA submission 2024; first commercial deployment expected 2026. Combines elements of magnetic-soft actuation with mainstream capsule endoscopy — and competes with passive devices from Medtronic + Olympus by adding controllability.

8.8 Artimus Robotics (Keplinger spin-out, Boulder, 2018)

Tim Morrissey, Christoph Keplinger, and team commercialized HASEL technology. Artimus offers gripper systems and customizable HASEL actuator stacks (10-1000 N range). Disney Imagineering licensed the technology for animatronics (smooth, biological motion); Toyota Research Institute partnered for haptics. Demonstrates that high-voltage soft actuation can be productized despite the engineering challenges.

8.9 RightHand Robotics + e-commerce (2014-2026)

MIT / Yale spin-out (Leif Jentoft, Lael Odhner, Yaroslav Tenzer — Dollar lab Yale). RightPick3 (2022 generation) uses a hybrid soft + vacuum end-effector — a central suction cup surrounded by compliant finger pads. Computer vision selects a grasp; the soft fingers tolerate small alignment errors. Deployed at scale at major US retailers + Symbotic warehouse systems. The hybrid soft-rigid approach has proved more reliable than pure-soft for high-throughput sorting.

8.10 Festo BionicSoftArm + BionicSoftHand (2019-2025)

Festo’s Bionic Learning Network produces an annual conceptual demonstrator. BionicSoftArm (2019, 2022, 2024 versions) is a 13-DoF pneumatic continuum manipulator with modular bellows-style sections; payload 3 kg, reach 1 m. BionicSoftHand pairs with it — pneumatic finger actuators with optional tendon override. Productized 2024 as part of Festo’s Motion Apps cobot line — pneumatic continuum at industrial reliability targets (10⁶ cycle life on key components). Demonstrates the path from research demonstrator to industrial product line.

8.11 RoboBee + soft-perch landing (Wood Lab Harvard 2013-2023)

Robert Wood’s Harvard Microrobotics Lab developed the RoboBee — a 90 mg insect-scale flapping-wing robot. While the airframe is rigid carbon fiber, the wings are flexible polyimide with piezoelectric actuators producing soft-equivalent flap dynamics. RoboBee X-Wing (2019) added solar-cell power. Subsequent versions used electrostatically-tethered “perching” landing pads — soft material patches that conform to surfaces. Demonstrates that soft and small are co-enabling for flying robotics.

8.12 Yip Lab UCSD — soft-robot state estimation + control (2018-2025)

Michael Yip’s lab at UCSD focuses on closed-loop control of soft continuum robots using learned dynamics models. Demonstrated MPC over a learned recurrent-neural-network surrogate of a tendon-driven manipulator running at 100 Hz. Their open-source SoftRoboticsPipeline integrates data collection, system identification, MPC, and reinforcement-learning fine-tuning. Used by several university labs through 2025 as standard infrastructure.

8.13 ETH Zürich + IIT — fiber jamming for variable-stiffness surgery (2022-2025)

Fiber-jamming (parallel high-strength fibers in a vacuum-jacketed sleeve) demonstrated by Marco Hutter (ETH RSL) and Cesare Stefanini (IIT) for surgical instruments. The instrument is soft when fibers slide; vacuum applied → fibers lock parallel → instrument becomes a stiff cantilever. Lower friction than layer or granular jamming, better cycle life. Targeting flexible-endoscope tip articulation; commercial product Endiatx + Auris-class device 2026+.

8b. Additional case studies and deployments

8.14 RightHand Robotics RightPick3 + Symbotic (2023-2026)

RightHand integrated their soft + suction hybrid end-effector into Symbotic’s warehouse system (which itself acquired the legacy SoftBank Robotics Symbotic deployments). The RightPick3 system pairs a soft compliant collar around a primary suction cup; ML-based grasp planner picks from clutter. Throughput: ~1000 picks/hour per cell, ~95% success on diverse-SKU bins. Deployed at Walmart, Target, FedEx. Cited as the most successful hybrid soft-rigid manipulation product in commercial use.

8.15 Carbon Robotics LaserWeeder (2024)

Not soft-robot per se but combines field-grade rigid frame with compliant tine-style soil engagement. The LaserWeeder identifies + lasers weeds at ~100k weeds/hour via 30 lasers; the chassis floats on compliant suspension that conforms to soil surface. Sold to row-crop operators in 2024-2025 at $1.2M+ per unit. Demonstrates that compliance + precision are complementary in agricultural robotics.

8.16 Bionaut Labs intracranial magnetic robots (2024)

Bionaut (Tel Aviv + Los Angeles) developed sub-mm magnetically-actuated soft / continuum micro-robots for intracranial drug delivery. The body is a soft polymer with embedded NdFeB; an external rotating MRI-compatible magnet field steers + propels the device through brain ventricles. Phase 1/2 clinical trials underway for ependymal cyst drainage and targeted brainstem drug delivery. First clinical-trial soft-magnetic-robot indication.

8.17 Yale OpenHand Project (Dollar — 2014-2025)

Aaron Dollar’s GRAB Lab released the OpenHand series — open-source 3D-printable underactuated hands. Tendon-driven, designed for tendon laundering through soft pulleys, fingers compliant via embedded TPU joints. Spawned hundreds of academic + maker replicas. Yale Model-T (2014) → Model-S (2015) → Model-O (2017) → mostly superseded by commercial alternatives by 2024 but seminal for democratizing dexterous manipulation.

8.18 Festo BionicFinWave + BionicSwimmer (2018-2024)

Annual Festo Bionic Learning Network demonstrators. Underwater soft robots inspired by cuttlefish (BionicFinWave 2018, undulating soft fin), shark (BionicSwimmer 2024). Not commercial products but inform Festo’s pneumatic-actuation product line and serve as recruiting + branding tools. Validates the technology-demonstrator pipeline.

8.19 Wandercraft Atalante (2018-2026)

Wandercraft (Paris) — exoskeleton for paraplegics, not strictly soft but with compliant joints throughout. CE-marked 2018. ~500 units in clinical use across European rehabilitation centers by 2024. Validates compliant-joint exoskeletons in regulated medical markets.

8.20 Vine robot — Stanford EVERSION (Hawkes-Cutkosky 2017)

Elliot Hawkes, Laura Blumenschein, Allison Okamura at Stanford. Pneumatic everting tube — the body grows from its tip by inversion of stored material. Demonstrated growing through a confined space (vine plant analog) to access otherwise unreachable locations. Spin-out: Boston-area startup pursuing search-and-rescue + inspection applications. Genuinely novel locomotion mode (no equivalent in rigid robotics).

9. Fabrication, materials, and processes

9.1 Molding workflows

Soft-robotics fabrication is dominated by silicone molding. The standard workflow:

1. CAD the body in SolidWorks / Fusion 360 / OnShape; design two-part (or three-part) mold splitting on a flat seam.
2. 3D-print the mold in tin-cure photopolymer (Formlabs Tough 2000), PLA (for non-platinum-cure silicones only), or machined aluminum (production tooling).
3. Surface-prep the mold with mold release (Smooth-On Universal Mold Release, Mann Ease Release 200) or Inhibit-X barrier coating if substrate is platinum-cure-inhibiting.
4. Mix silicone Part A + Part B (1:1 typically for Ecoflex / Dragon Skin) in a planetary mixer or vacuum chamber (THINKY ARE-310) for 1-2 minutes.
5. Degas under vacuum at ~25 inHg for 5 minutes to remove entrained air bubbles.
6. Pour into mold in a slow stream from low height.
7. Cure at room temperature (4-24 hr) or oven (60-80 °C, 30 min — Dragon Skin Fast / Smooth-Sil).
8. Demold + post-bond the constrained inextensible layer (paper, kevlar, glass-fiber fabric) with Sil-Poxy or plasma + adhesion promoter.

Common variations: lost-wax for internal cavities too complex to demold; rotational molding (Tolley group, Cornell) for hollow spherical bodies; multi-material 3D printing (Lewis Lab Harvard, Stratasys Objet) for graded-stiffness embedded structures; inkjet of conductive traces (Optomec Aerosol Jet) for embedded electrodes.

9.2 Direct-write 3D printing of elastomers

Carbon DLS, Formlabs Elastic 50A, Stratasys Agilus, and the suite of “embedded 3D printing” technologies from Lewis Lab (Harvard) at the academic frontier. The 2014 Lewis-group paper “3D Printing of Interdigitated Dielectric Elastomer Actuators” demonstrated co-printed dielectric and electrode. Current consumer-accessible: Formlabs Form 4B + Elastic 50A produces silicone-equivalent parts at Shore 50A, durable for short-cycle prototypes. Production-grade direct silicone printing: Spectroplast (ETH spin-out 2018) and German RepRap LiquidAM offer industrial silicone printers.

9.3 Constitutive models — what to put in FEM

For silicone elastomers under typical soft-robotics strain levels (10-200%):

• Neo-Hookean — single parameter ($\mu$); valid to ~50% strain; default for first-pass FEM.
• Mooney-Rivlin (2-param) — adds $C_{10},C_{01}$; valid to ~200% for many silicones.
• Ogden N=3 — three pairs $({\mu }_i,{\alpha }_i)$; valid to 400+% strain; required for Ecoflex 00-30 at large deformation.
• Yeoh — alternative 3-parameter form; popular for natural rubber.
• Gent — captures strain-hardening near limiting stretch.

Calibration from uniaxial tensile + planar shear tests via DIC (Digital Image Correlation, GOM Aramis / Correlated Solutions Vic-3D). FEM with the wrong constitutive can be off by 2-5× on tip displacement.

9.4 Bonding silicone to itself and others

A non-trivial engineering problem. Tested combinations:

Interface	Method	Strength
Silicone to silicone (same family)	Fresh-on-cured with primer	Excellent
Silicone to silicone (different family)	Sil-Poxy	Good
Silicone to fabric	Sil-Poxy + plasma	Good
Silicone to TPU	Plasma + cyanoacrylate	Moderate
Silicone to metal	Dow Corning Q3-7068 primer	Excellent
Silicone to PLA / ABS	Generally poor; encapsulate instead	Poor

Plasma treatment (handheld plasma pen or atmospheric-pressure plasma jet) increases surface energy for 30 minutes post-treatment, enabling adhesion that would otherwise fail. Standard step in industrial soft-gripper assembly.

9.5 Test rigs for soft characterization

• Quasi-static tensile — Instron, MTS, custom screw-drive with load cell. ASTM D412 / D624 protocols.
• Cyclic fatigue — same hardware, programmed cycling to failure or 10⁶ cycles.
• Burst test — water-pressure burst of pneumatic actuators. ANSI/SAE J343 adapted.
• Force-displacement at end-effector — six-axis F/T sensor (ATI Mini40, Bota SensONE) integrated into a fixture.
• Shape reconstruction ground truth — Vicon / OptiTrack with markers along the body, or laser scanner (FARO Edge) for static configurations.

10. Safety, regulatory, and clinical considerations

10.1 Industrial safety (ISO/TS 15066)

Power-and-force-limiting collaborative robots must stay below biomechanical injury thresholds (Annex A) for skull, sternum, eye, etc. The soft-robot inherent advantage: low effective stiffness multiplies into low transferred energy. Even at full-payload collision, mGrip-class pneumatic fingers typically transfer well under threshold values for forehead and forearm. However, the mounting arm (Franka, UR5e, ABB cobot) is rigid; an end-effector mounted on a rigid arm at 1 m/s strikes with arm-stiffness-dominated force regardless of soft gripper compliance. Real safety analysis must include the full kinematic chain.

10.2 Food contact (FDA 21 CFR 177)

Soft Robotics Inc. mGrip variants for direct food contact use silicones certified under 21 CFR 177.2600 (rubber articles intended for repeat use). Smooth-On Ecoflex GEL and Smooth-Sil 940 are food-contact certified; not all silicones are. Testing: extractables analysis, durability under USP / NSF conditions.

10.3 Medical (FDA 510(k), CE Mark MDR)

Continuum surgical robots (Auris Monarch, Medrobotics Flex System, Hansen Magellan) require Class II/III medical-device certification. Soft components must use medically-cleared materials (Silastic Q7 series, NuSil R-2188). Sterilization compatibility (autoclave 121 °C, gamma, EtO) is constraining — many “soft” prototype materials cannot survive autoclave cycling. Single-use disposable soft instruments are the standard workaround.

10.4 High-voltage soft (IEC 61010)

DEA / HASEL systems operate at 3-10 kV. Insulation testing (IEC 60664 creepage and clearance distances), dielectric breakdown verification, and grounded enclosure for operator safety are mandatory. Artimus Robotics ships HASEL drivers with on-board safety interlocks.

10.5 Cleanroom and aerospace

Outgassing-tested silicones (NASA SP-R-0022 and ECSS-Q-ST-70-02C) are required for spaceflight. Nusil CV-1142 and similar low-outgassing variants exist but at significantly higher cost than commodity Smooth-On products.

10a. Modeling soft robots — practical comparison

10a.1 When to use each model

Method	Speed	Accuracy	Best for
PCC (Webster-Jones)	Fastest	Lowest	Design exploration, real-time control
Cosserat rod	Fast	Medium	Slender continuum; tendon-driven
FEM (SOFA)	Slow	Highest	Validation, novel geometries
Reduced-coordinate	Medium	Medium-high	Closed-loop control
Neural surrogate	Fast (inference)	Variable	Black-box surrogate after data

For control applications: PCC at 1 kHz; FEM offline; learned-neural surrogate at 100 Hz as a middle ground.

10a.2 Data-driven dynamics

Modern practice (2023+): collect 100k+ (input pressure, joint output) pairs from a physical soft robot; train a recurrent neural network or transformer dynamics model; deploy in MPC. Yip Lab UCSD, Berkeley FlexLab, Della Santina group TU Delft all converged on this. Replaces hand-derived Cosserat models for control while still respecting their structural insight (configuration space dimensionality).

10a.3 Simulators for soft robotics — head to head

• SOFA + Soft Robotics Plugin (INRIA, free). Mature; thousands of papers; ODE-based time-stepping; FEM-based. Hard to install; Python bindings improving 2024+.
• Abaqus + Ansys (commercial). Industry standard for hyperelastic FEM; offline; not real-time.
• PyElastica (Gazzola UIUC, free). Cosserat-rod focused; Python; differentiable. Good for slender bodies.
• Genesis (CMU 2024). Unified rigid + soft + fluid; differentiable; GPU. Still maturing.
• MuJoCo Composite (DeepMind 2022). Soft via composite tendons + flexcomp; limited but fast.
• Warp + Newton (NVIDIA 2024). Differentiable Python kernels for soft / fluid.
• DiSECt (NVIDIA 2021). Differentiable cutting — surgical sim.

10a.4 Verification of soft-robot controllers

Standard practice: test in sim → test on physical prototype → test on fielded unit. For each, define metrics: tip-position accuracy, grasp success rate, cycle life. The sim-to-real gap for soft robots is larger than for rigid: silicone batch variability + temperature dependence + drift add ~10-20% noise on FEM predictions that’s hard to systematically randomize. Standard mitigation: hand-tune FEM to match measured per-unit, then accept that fielded units behave within a band.

10b. Application domains (deep dive)

10b.1 Food handling

The dominant commercial soft-robotics market in 2026. Drivers: human worker shortages in meat processing + produce packing, the inability of rigid grippers to handle delicate or variable-shape food without damage. Soft Robotics Inc., RightHand, OnRobot, Piab dominate. Specific applications:

• Bakery. Lifting dough, croissants, loaves; mGrip 3-finger arrays at automated bakeries (Bimbo, JBT FoodTech).
• Produce. Tomatoes, strawberries, peaches; gentle grasp with no bruising.
• Meat. Chicken parts (deboned + bone-in), fish fillets; FDA-grade silicone.
• Confectionery. Chocolates, gummies — temperature sensitive + sticky.
• Frozen. Pre-formed patties, fish blocks; soft fingers tolerate sub-zero temps better than rubber.

Typical line: 6-axis cobot (UR10e or Yaskawa MotoMini) + Soft Robotics mGrip + Cognex / SICK vision. Throughput 30-60 picks/min per cell.

10b.2 E-commerce and warehouse picking

Distinct from food due to higher SKU diversity + harder shapes. RightHand Robotics RightPick3 + Symbotic + Soft Robotics + several Chinese OEMs deployed at Walmart, Target, FedEx, JD.com. The hybrid suction + soft-collar pattern dominates — pure soft pneumatic struggles with rigid boxes, pure suction fails on porous or contoured items.

10b.3 Medical and surgical

Highest revenue per unit but smallest unit count. Auris Health Monarch (J&J), Medrobotics Flex System (defunct 2019, IP at Medical Microinstruments), Intuitive Surgical da Vinci SP (single-port). Disposable continuum instruments are the growth segment: Endiatx PillBot, Beat Surgical (Israel) colonoscopy assist, several startups for ENT and urology. Regulatory burden (FDA Class II/III, MDR Class IIb/III) keeps barriers high.

10b.4 Assistive and prosthetics

Pisa/IIT SoftHand (Prensilia, IIT spin-out), Ottobock bebionic + Michelangelo (multi-articulating, mostly rigid), Naked Prosthetics PIP-Drivers (3D-printed soft-rigid hybrid), Open Bionics Hero Arm (Bristol; affordable multi-grip). Soft prosthetics excel at price-per-grasp-quality; rigid multi-articulating excel at distinct independent finger control. Both coexist in the market.

10b.5 Wearable assistance and exos

Soft exos (Harvard Wyss + ReWalk subsidiary Eksoworks): assist walking via cable-driven garments rather than rigid frames. Lower-impedance, more comfortable, but lower assistance forces. CYBERDYNE HAL is rigid; Wandercraft Atalante mostly rigid. Soft exo niche: ankle-foot assist for stroke / Parkinson’s where small forces suffice.

10b.6 Inspection and exploration

Vine robots (Hawkes EVERSION), soft snake robots (Howie Choset CMU), Festo BionicSoftArm-class continuum arms for confined-space inspection. Pipe inspection, ventilation ducts, nuclear-decommissioning. Vibe Robotics, Aris Industries are small commercial players. Open frontier.

10b.7 Underwater

MIT SoFi (soft robotic fish, Rus 2018), BIKI consumer underwater drone, Pliant Energy Systems Velox + Bilbo (soft-fin propulsion). Octopus-inspired manipulators for marine ROV gripping. Mostly research stage; commercial value unproven.

10b.8 Space

NASA JPL Astrobee — not soft per se but used soft-handle interfaces. Soft-robotic propulsion (ChonyRobotics asteroid manipulator concept) is research-only. Space-grade silicone (NuSil CV-1142) is the limiting commercial input.

10c. Cross-cutting comparisons

10c.1 Soft vs rigid — when to choose

Choose soft when:
  - Object shape is variable / unknown
  - Object is fragile (crush risk)
  - Safety against humans is paramount
  - Confined-space access required
  - Biomimetic morphology is the design driver
  - Energy storage in compliance is useful (running, hopping)

Choose rigid when:
  - Precise position control is required (< 1 mm)
  - High force is needed (> 100 N continuous)
  - Long cycle life is critical (> 10M cycles)
  - Industrial repeatability standard
  - Regulatory environment requires deterministic behavior
  - Cost per cycle must be minimal

10c.2 Soft vs cable-driven vs compliant-mechanism

These three are often confused but solve different problems:

• Compliant mechanism (Howell, BYU). Rigid links connected by lumped flexures — a single elastic deformation replaces a pin joint. Pseudo-rigid analysis applies. Standard for MEMS, surgical instruments, watch movements.
• Cable-driven robot. Rigid links + flexible tendons routed through pulleys. Tendons transmit force; structure is rigid. Examples: SkyCam, Pisa/IIT SoftHand (in tendon-driven mode).
• Soft robot. Body itself deforms; no rigid skeleton. Examples: PneuNet, FEA arms, octopus arms.

The boundaries blur — RBO Hand has rigid bones in soft silicone fingers; Festo BionicSoftHand has rigid couplings between pneumatic chambers. The classification matters for which modeling and design tools apply.

10c.3 The compliance hierarchy

From rigid to soft, with examples:

1. Aluminum / steel frame robot — KUKA KR16, ABB IRB. E ≈ 70-200 GPa.
2. Composite + harmonic-drive cobot — UR10, Franka. Visible compliance under load.
3. Series-elastic actuator (SEA) robot — ANYmal, Atlas. Designed-in compliance.
4. Cable-driven with flexible base — SkyCam. Bulk compliance from cables.
5. Compliant mechanism with lumped flexures — surgical instruments. Localized soft elements.
6. Tendon-driven anthropomorphic hand — Pisa/IIT SoftHand, Shadow Hand. Hybrid.
7. Hyper-redundant snake — CMU snake. Many rigid links + ball joints.
8. Continuum robot, tendon-driven — STIFF-FLOP, Auris Monarch. Compliant backbone.
9. Continuum robot, pneumatic — Festo BionicSoftArm. Pressure-driven.
10. Soft body, PneuNet / FEA — mGrip, RBO Hand. Intrinsically compliant.
11. Soft body, DEA / HASEL — Artimus grippers. Electrostatic.
12. Soft body, magnetic / liquid — Bionaut, capsule endoscopes.

Most real robots blend levels 4-10. Pure soft (10+) is currently confined to research + niche commercial.

11. Open research directions (2024-2026)

• Self-healing. Polymers that autonomously repair after cuts. White Lab (UIUC), Sottos, Moore, Cordon Bleu chemistry. Targeted for cycle-life multiplication.
• Growing soft robots. Stanford “vine robot” (Hawkes et al. 2017) inverts from an inverted state, pneumatically growing through confined spaces. Application: pipe inspection, search-and-rescue.
• Embodied intelligence. Pfeifer “How the body shapes the way we think” — the compliance + morphology is the controller. RL trained on body design + controller jointly (Spielberg, Bhatia, Cheney 2017+).
• Bio-hybrid actuators. Cardiac-muscle-actuated swimmers (Park, Parker — Harvard 2016 “Phototactic guidance of a tissue-engineered soft-robotic ray”). Living tissue + scaffold.
• Magnetic micro-swimmers. Sitti (MPI Stuttgart), Pané (ETH), Nelson — sub-mm scale magneto-active robots for medical applications.
• Cosserat-rod differentiable simulation. PyElastica (Gazzola, UIUC) and DiffTaichi enable gradient-based design optimization.
• Soft optical / acoustic computing. Use soft material as the computational substrate — Hardesty + Reis + Pluchino — fluid-channel logic.
• Soft sensor scaling. Reproducible large-area capacitive skins remain unsolved at production scale; Bota, StretchSense, Tangible Research compete.

12. Reference workflow — building a pneumatic soft gripper

12.1 Specification phase

1. Define payload envelope: object shapes, weights, surface friction, fragility, max acceptable force.
2. Define throughput: cycles/min, expected operating life.
3. Define environment: cleanroom, washdown, food-grade, vacuum, temperature.
4. Pick gripper topology: 2-finger, 3-finger, 4-finger; opposing or parallel; with or without thumb.

12.2 Design phase

1. Model the finger geometry in CAD; parametrize chamber count, wall thickness, length.
2. FEM (SOFA or Abaqus) sweep over geometry to predict tip force vs pressure curve.
3. Select silicone family for the duty cycle (Ecoflex for low force / high strain; Dragon Skin for high cycles).
4. Design mold (split-mold for 2-part fingers; lost-wax for closed chambers).
5. Specify pneumatic system: regulator pressure range, valves (on/off vs proportional), tubing diameter, on-board manifold vs off-board.
6. Specify control architecture: PLC, microcontroller (e.g., Teensy 4.1 + valve drivers), or industrial cobot integration (UR cap, Yaskawa MotoPlus).

12.3 Prototype phase

1. 3D-print or machine the mold.
2. Mix + degas + pour silicone; cure per data sheet.
3. Demold + bond fabric / paper backing.
4. Pressure-test individually to 2× operating pressure for burst margin.
5. Cycle-test 1000 cycles at operating pressure; inspect for fatigue cracks.
6. Force-test on a load cell: tip force vs pressure curve.

12.4 Integration phase

1. Mount on cobot flange via quick-change adapter.
2. Run vision pipeline (Cognex VisionPro, Halcon, or custom YOLO + 6-DoF pose).
3. Set up pick-and-place poses; tune approach trajectory for soft-contact engage.
4. Validate on 100+ samples of payload variation.
5. Document cycle life, replacement procedure, operator training.

12.5 Deployment + maintenance

1. Spare finger inventory: 2-3 sets per shift.
2. Replace fingers at end-of-life (typically 100k-1M cycles depending on payload).
3. Monitor pressure drop / leak rate as health indicator.
4. Annual revalidation: pressure-test, force-test, vision recalibration.

12b. Quick-reference design constants

For rapid sizing in soft-robotics work, the following constants are useful to memorize:

• Atmospheric pressure: 101 kPa = 1 bar = 14.7 psi.
• Typical PneuNet operating pressure: 30-70 kPa above atmospheric.
• Typical McKibben operating pressure: 200-500 kPa.
• DEA breakdown field for VHB 4910: 200 V/μm; operate at 100 V/μm for margin.
• HASEL operating voltage: 5-10 kV; current draw mA-class at steady state.
• Nitinol transformation temperature (typical): 30-90 °C; pick variant for application.
• Silicone density: ~1.05 g/cm³ for Ecoflex; ~1.15 for tin-cure variants.
• Permittivity of free space ${\epsilon }_0=8.854\times 10^{-12}$ F/m.
• Relative permittivity VHB 4910: 4.7; silicone elastomers: 2.5-3.5.
• Young’s modulus, Ecoflex 00-30: 0.05-0.1 MPa.
• Young’s modulus, Dragon Skin 10: 0.15-0.2 MPa.
• Young’s modulus, human skin: 0.5-2 MPa (depends on direction + age).
• Maximum strain at break, Ecoflex 00-30: ~900%.
• Tear strength, Dragon Skin 30: 26 N/mm.

13. Glossary

• Compliance — inverse of stiffness; the ability of a body to deform under load.
• Continuum robot — a robot with a structurally compliant backbone but discrete actuation.
• Cosserat rod — a continuum-mechanics model treating a flexible body as a 1D curve with attached frames.
• DEA — Dielectric Elastomer Actuator — voltage-driven Maxwell-stress actuator using elastomer + compliant electrodes.
• FEA — Fluidic Elastomer Actuator — pressure-driven soft actuator with internal channels.
• HASEL — Hydraulically Amplified Self-healing ELectrostatic actuator; hybrid hydraulic + electrostatic.
• Hyper-redundant — having many more DoF than the task requires; usually rigid with many discrete joints.
• Jamming — granular / layer / fiber-based stiffness transition via vacuum.
• LCE — Liquid Crystal Elastomer — elastomer with embedded mesogens; contracts on heating.
• McKibben muscle — braided pneumatic muscle; high force-to-weight contractile actuator.
• Morphological computation — the body’s mechanical dynamics doing useful computation, replacing controller logic.
• PCC — Piecewise Constant Curvature — kinematic approximation of continuum body as connected arcs.
• PneuNet — Pneumatic Network — Whitesides-style molded silicone with internal channels.
• SMA / SMP — Shape Memory Alloy / Polymer — thermal-triggered shape recovery actuators.
• Soft robot — robot whose primary structure is intrinsically compliant material.
• Stigmergy — coordination via the environment, not direct communication.
• Synergy (postural) — low-dimensional manifold of finger postures used in underactuated hands.
• Tendon-driven — actuated via cables routed through the body to a remote motor.

Further reading

• Rus D., Tolley M.T. (2015) “Design, fabrication and control of soft robots.” Nature 521.
• Laschi C., Mazzolai B., Cianchetti M. (2016) “Soft robotics: technologies and systems pushing the boundaries of robot abilities.” Science Robotics 1.
• Polygerinos P., Correll N., Morin S.A., Mosadegh B., Onal C.D., Petersen K., Cianchetti M., Tolley M.T., Shepherd R.F. (2017) “Soft robotics: review of fluid-driven intrinsically soft devices.” Advanced Engineering Materials 19.
• Pelrine R., Kornbluh R., Joseph J. (2000) “High-speed electrically actuated elastomers with strain greater than 100%.” Science 287.
• 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 transducers harnessing the Hertzian dipole.” Science 359.
• Wehner M., Truby R.L., Fitzgerald D.J., Mosadegh B., Whitesides G.M., Lewis J.A., Wood R.J. (2016) “An integrated design and fabrication strategy for entirely soft, autonomous robots.” Nature 536.
• Brown E., Rodenberg N., Amend J., Mozeika A., Steltz E., Zakin M.R., Lipson H., Jaeger H.M. (2010) “Universal robotic gripper based on jamming of granular material.” PNAS 107.
• Webster R.J., Jones B.A. (2010) “Design and Kinematic Modeling of Constant Curvature Continuum Robots: A Review.” International Journal of Robotics Research 29.
• Della Santina C., Duriez C., Rus D. (2023) “Model-Based Control of Soft Robots: A Survey of the State of the Art and Open Challenges.” IEEE Control Systems Magazine 43.
• Marchese A.D., Komorowski K., Onal C.D., Rus D. (2014) “Autonomous soft robotic fish capable of escape maneuvers using fluidic elastomer actuators.” Soft Robotics 1.
• 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.
• Lum G.Z., Ye Z., Dong X., Marvi H., Erin O., Hu W., Sitti M. (2016) “Shape-programmable magnetic soft matter.” PNAS 113.
• Cianchetti M., Ranzani T., Gerboni G., Nanayakkara T., Althoefer K., Dasgupta P., Menciassi A. (2014) “Soft robotics technologies to address shortcomings in today’s minimally invasive surgery: the STIFF-FLOP approach.” Soft Robotics 1.
• Mazzolai B., Mondini A., Tramacere F., et al. (2022) “Roadmap on soft robotics: multifunctionality, adaptability and growth without borders.” Multifunctional Materials 5.
• Whitesides G.M. (2018) “Soft Robotics.” Angewandte Chemie International Edition 57.
• Bishop-Moser J., Krishnan G., Kim C., Kota S. (2012) “Design of soft robotic actuators using fluid-filled fiber-reinforced elastomeric enclosures in parallel combinations.” IROS.
• Ware T.H., White T.J. (2015) “Programmed liquid crystal elastomers.” Macromolecules 48.
• Hu Y., Liu J., Spielberg A., Tenenbaum J.B., Freeman W.T., Wu J., Rus D., Matusik W. (2019) “ChainQueen: A real-time differentiable physical simulator for soft robotics.” ICRA.

Adjacent

• materials-polymers
• mechanics-of-materials
• fem-fea
• biomechanics
• microfluidics
• polymers-taxonomy
• vibration-dynamics
• bioinstrumentation