Parallel Manipulators (Stewart-Gough, Delta, Hexapod)

1. At a glance

A parallel manipulator (also called a parallel kinematic machine, PKM) drives a single moving end-effector platform with two or more independent kinematic chains running in parallel from a fixed base. Every chain contributes a constraint; the platform pose is the simultaneous solution. This stands in contrast to the serial manipulator (manipulator-design), where a single open chain of links is stacked end-to-end and the last link carries the tool. The two architectures are mathematical duals: the things that are easy on a serial arm are hard on a parallel one, and vice versa.

Why anyone builds them. Stacking actuators in parallel rather than in series gives you four wins that a serial arm cannot match at the same cost:

Stiffness — load paths are short and triangulated. A serial 6R arm’s wrist deflection under a 10 kg load is millimetres; a Stewart platform’s is micrometres.
Precision — error stacking is sub-additive across parallel chains, super-additive across serial joints. A PI H-840 hexapod delivers 10 nm linear precision; the best serial arm hits ~10 µm.
Payload-to-mass ratio — actuators live at the base, moving mass is just the platform + leg upper portions. A delta robot at 1 kg moving mass can fling a 3 kg payload at 100 m/s².
Speed / acceleration density — same root cause; low moving mass, distributed actuation. ABB FlexPicker hits 10 g platform acceleration, 100+ picks/min.

Why most people don’t build them. Four cost factors:

Workspace is tiny relative to footprint. A Stewart with 1 m base diameter delivers ~0.3 m linear travel and ~30° angular. Serial arms reach 2× their base diameter.
Forward kinematics is hard. Inverse kinematics (pose → leg lengths) is trivial — you solve six independent scalar equations. Forward kinematics (leg lengths → pose) is a system of polynomial equations with up to 40 real solutions (Husty 1996). Numerical Newton-Raphson with a good seed, or interval analysis (Merlet), is the only way.
Singularities inside the workspace. Serial arms have singularities at the boundary (wrist-aligned, shoulder over-head). Parallel arms get Type II singularities in the interior where the platform gains an uncontrolled DOF — the platform can move under load with zero leg motion. Mission-critical.
Calibration is much more sensitive. Each leg’s manufacturing tolerance adds linearly to platform error.

Real-world archetypes. Three problems own this geometry:

Flight simulators (Stewart 6-DOF): CAE 7000XR, L3Harris RealitySeven. Hydraulic six-legs; 60° pitch/roll envelope; certified Level D simulators for Boeing 737, A320.
High-rate pick-and-place (delta 3T+1R): ABB IRB 360 FlexPicker, Fanuc M-1iA, Stäubli TP80, Adept Quattro. 100–200 picks/min in pharma and food packaging.
Sub-micron positioning (small Stewart hexapods): PI H-810 / H-840 / N-Series, Newport HXP1000, Symetrie BORA. Semiconductor metrology, beamline optics, astronomy secondary-mirror support.

Plus the niches: surgical robots (Mazor X spine, Brainlab Cirq), machine tools (Sprint Z3 head, Tricept T606), driving simulators (VI-grade DiM250, Cruden Hexatech), and a growing population of cable-driven parallel robots (cable-driven-robots).

2. First principles

Origin

Eric Gough (1947, Dunlop Tyres) built the first practical six-legged parallel mechanism — a universal tyre-testing rig that could apply six-DOF loads to a vehicle wheel. D. Stewart (1965, “A Platform with Six Degrees of Freedom”, Proc. IMechE 180:371) proposed the same topology for flight simulators, gave it a kinematic analysis, and the name Stewart platform (more accurately, Stewart-Gough platform) entered the literature. Cappel (1967) patented and built the first commercial flight-sim Stewart, and Klaus Cappel’s hexapod is the direct ancestor of every motion-base simulator made today.

The delta robot is much younger: Reymond Clavel at EPFL filed the patent in 1985 and presented it at ICRA 1988. The decisive idea was using three parallelograms to constrain the platform to pure translation regardless of leg angle — a much smaller workspace than Stewart but with closed-form forward kinematics and 10× the speed.

Degree-of-freedom count: Grübler-Kutzbach

The mobility of a spatial linkage is:

$F = 6 (n - 1) - \sum_{i = 1}^{j} (6 - f_{i})$

where $n$ is the number of links (including base), $j$ is the number of joints, and $f_{i}$ is the DOF of the $i$ -th joint (1 for R or P, 2 for U, 3 for S). For a generic 6-UPS Stewart-Gough:

$n = 14$ (1 base + 1 platform + 6 lower legs + 6 upper legs)
joints: 6 universal at base (U, $f = 2$ ) + 6 prismatic (P, $f = 1$ ) + 6 spherical at platform (S, $f = 3$ )
$F = 6 \cdot 13 - 6 (6 - 2) - 6 (6 - 1) - 6 (6 - 3) = 78 - 24 - 30 - 18 = 6$ ✓

For a 3-RUU delta (Clavel topology, three legs each R-U-U):

$n = 1 + 1 + 3 \cdot 4 = 14$ ; $j = 3$ actuated R + 3 U at top + 3 U at platform; but the parallelogram constraint reduces each leg’s free DOF
accounting for the parallelogram, $F = 3$ (pure translation) ✓

The exception (“over-constrained mechanism”) — when symmetry creates redundant constraints — is why Grübler-Kutzbach must be checked against a screw-system analysis (Hunt 1978; Davidson & Hunt 2004) before trusting the count.

Inverse kinematics — the easy half

Given the platform pose $T = (R, p) \in SE (3)$ , the leg lengths or joint angles are computed independently per leg. For a Stewart-Gough with base anchor $b_{i}$ (in base frame) and platform anchor $p_{i}$ (in platform frame), the leg vector is:

$l_{i} = p + R p_{i} - b_{i}$

The leg length is $ℓ_{i} = ∥ l_{i} ∥$ , and the unit direction $\hat{l}_{i} = l_{i} / ℓ_{i}$ . Six independent scalar equations, no iteration, microsecond-fast. This is why Stewart platforms run at >1 kHz servo rates — the controller does six norms per cycle.

Forward kinematics — the hard half

Given the six leg lengths, find the platform pose. Six unknowns (three translations + three rotations), six constraint equations $∥ l_{i} ∥^{2} = ℓ_{i}^{2}$ . After eliminating, the result is a polynomial system whose degree is famously 40 for the general 6-6 Stewart (Raghavan 1993; Husty 1996, “An algorithm for solving the direct kinematics of general Stewart-Gough platforms”, Mech. Mach. Theory 31:365). For special architectures the degree drops: 6-3 Stewart → 16 solutions (Innocenti & Parenti-Castelli 1990 ICRA), 3-3 Stewart → 16, planar 3-RPR → 6.

Three practical approaches:

Newton-Raphson from previous solution — for control. Cache the previous valid pose as the seed; the controller’s previous step is within microns of the current one. Converges in 2–3 iterations. Fails if you lose state (power-cycle without homing).
Multi-start Newton-Raphson — for off-line analysis. Randomly perturb the seed 50–100 times to find all real solutions.
Interval analysis (Merlet 2004) — guaranteed enclosure of all real solutions, the only method that gives certified results for safety-critical FK. Implemented in Merlet’s ALIAS library.

Jacobians and singularities

A parallel manipulator’s velocity equation is:

$J_{x} \dot{x} = J_{q} \dot{q}$

where $x$ is the platform pose and $q$ is the vector of actuated joint values. $J_{x}$ is the output Jacobian (how platform velocity maps into the constraints), $J_{q}$ is the input Jacobian (how actuator velocities map into the constraints). The conventional manipulator Jacobian is $J = J_{q}^{- 1} J_{x}$ (when the inverse exists).

Gosselin & Angeles (1990, “Singularity Analysis of Closed-Loop Kinematic Chains”, IEEE TRA 6:281) classified parallel singularities into three types based on which Jacobian degenerates:

Type	Determinant condition	Physical meaning	Severity
Type I (inverse kinematic / serial)	$det J_{q} = 0$	Actuator runs out of range; platform loses DOF at workspace boundary	Manageable; analogous to serial-arm reach limit
Type II (forward kinematic / parallel)	$det J_{x} = 0$	Platform gains an uncontrolled DOF — can move under load with zero actuator motion	Dangerous; structural collapse, control loss
Type III (combined / architecture)	both $= 0$	Architecture-dependent; mechanism temporarily becomes a different machine	Rare; design-out at synthesis

Type II is what makes parallel manipulators hard. A Stewart platform near a Type II singularity becomes a load amplifier — actuator forces blow up to maintain a load the platform can no longer resist. Real platforms reserve a 15–20% workspace margin and run online singularity proximity monitors based on $\sigma_\min(\mathbf{J}_x)$ , the smallest singular value.

3. Three worked examples

Example A — Delta inverse kinematics (ABB FlexPicker IRB 360 geometry)

The IRB 360 is a 3-RUU delta with three arms spaced 120° apart, each consisting of an actuated upper arm (length $L_{1} = 200$ mm) driving a parallelogram lower arm (length $L_{2} = 400$ mm) that connects to the moving platform. Base joint radius $R = 200$ mm, platform joint radius $r = 50$ mm.

Given: desired platform centre position $p = (x, y, z) = (100, 50, - 500)$ mm in base frame (z negative = downward toward conveyor).

Step 1 — reduce each arm to a planar problem. For arm $i$ at base angle $ϕ_{i} \in {0°, 120°, 240°}$ , rotate the world point into that arm’s plane:

$x_{i}^{'} = x cos ϕ_{i} + y sin ϕ_{i}$ $y_{i}^{'} = - x sin ϕ_{i} + y cos ϕ_{i}$ $z_{i}^{'} = z$

The arm’s shoulder joint sits at $(R, 0, 0)$ in its own plane; the platform anchor it drives sits at $(x_{i}^{'} + r, y_{i}^{'}, z_{i}^{'})$ — equivalently we shift the target by $- r$ to get effective shoulder-to-elbow target $(x_{i}^{'} - R + r, y_{i}^{'}, z_{i}^{'})$ in a frame where the shoulder is at origin.

Step 2 — solve the 2D two-link (shoulder + parallelogram). Let $u = x_{i}^{'} - R + r$ , $w = z_{i}^{'}$ . The elbow ride is constrained to distance $L_{2}$ from the platform anchor while the upper arm pivots about the shoulder. The closed-form expression for the upper-arm angle $θ_{i}$ (measured down from horizontal) is:

$θ_{i} = 2 arctan (\frac{- L _{1} - L _{1}^{2} + ( u ^{2} + w ^{2} + y _{i}^{'2} - L _{2}^{2} ) \cdot ( - u ^{2} - w ^{2} - y _{i}^{'2} + L _{2}^{2} + 2 L _{1} \cdot \dots )}{u ^{2} + w ^{2} + y _{i}^{'2} - 2 L _{1} u - L _{2}^{2} + L _{1}^{2}})$

(The full closed form takes a paragraph; the key point is that it exists and runs in microseconds.) For the spec target above with $ϕ_{1} = 0°$ : $u = 100 - 200 + 50 = - 50$ mm, $w = - 500$ mm, $y^{'} = 50$ mm — geometry feasible since $u^{2} + w^{2} + y^{'2} = 504$ mm < $L_{1} + L_{2} = 600$ mm and > $L_{2} - L_{1} = 200$ mm.

Step 3 — branch selection. The delta IK has $2^{3} = 8$ branches (each arm has elbow-up or elbow-down). For a pick-place delta, elbow-up is universally chosen — keeps actuators above the conveyor, gives upright workspace bounded above by the base plate. The controller hard-codes the branch; never iterate over branches mid-trajectory.

Example B — Stewart leg lengths for a 5° roll command (flight simulator)

Platform: 6-UPS Stewart with hydraulic legs, stroke ±150 mm about nominal $ℓ_{0} = 1200$ mm. Base anchor radius 1.5 m, platform anchor radius 1.0 m. Anchors paired: base 6-3 layout, platform 3-3 layout.

Command: $R = Rot_{x} (5°)$ , $p = (0, 0, h_{0})$ where $h_{0}$ is the nominal vertical height — pure roll.

Compute leg vectors using $l_{i} = p + R p_{i} - b_{i}$ for each of the six base–platform pairs. The two legs whose base anchors lie on the rolled-up side of the +x axis lengthen, the two on the rolled-down side shorten, the two on the roll axis change negligibly. Sample numerical result for one leg whose base anchor is at $(1.5, 0, 0)$ m and platform anchor at $(1.0, 0, 0)$ m (both on +x axis):

Rotated platform anchor in base frame: $R \cdot (1, 0, 0)^{T} = (1, 0, 0)^{T}$ (axis fixed) — leg unchanged on the roll axis. Good sanity check.

For a leg whose platform anchor is at $(0.5, 0.866, 0)$ m (the +y side):

Rotated platform anchor: $(0.5, 0.866 cos 5°, 0.866 sin 5°) = (0.5, 0.8627, 0.0755)$ m
Leg vector from base anchor at $(0.75, 1.299, 0)$ m: $l = (0 + 0.5 - 0.75, 0 + 0.8627 - 1.299, h_{0} + 0.0755) = (- 0.25, - 0.436, h_{0} + 0.0755)$
For $h_{0} = 1.2$ m: $∥ l ∥ = 0.0625 + 0.190 + 1.628 = 1.881 = 1.371$ m
Change from nominal 1.200 m: +171 mm — exceeds ±150 mm stroke; need taller nominal $h_{0}$ .

This 10–15 mm “rule of thumb” cited in the brief assumes the roll axis is not at platform level but well below — for a typical flight sim with the eye-point 1.5 m above the platform-anchor plane, a 5° roll induces ~130 mm of leg-length change.

Sanity check via Jacobian linearisation. For small rotations, $Δ ℓ_{i} \approx \hat{l}_{i} \cdot (ω \times R p_{i} + Δ p)$ . For pure roll $ω = (5° \cdot π /180, 0, 0) = (0.0873, 0, 0)$ rad and zero $Δ p$ : $Δ ℓ_{i} \approx \hat{l}_{i} \cdot (ω \times R p_{i})$ . For platform anchor at $p_{i} = (0.5, 0.866, 0)$ , $ω \times p_{i} = (0, 0, 0.0756)$ , projected onto a near-vertical leg direction $\approx 0.0756$ m = 76 mm. Half the exact answer because the leg is not perfectly vertical — the lateral component dominates the small-angle estimate. Lesson: for parallel-manipulator velocity-level analysis the Jacobian linearisation is cheap and informative, but final actuator-stroke specifications must use exact IK to avoid stroke-budget surprises at the workspace corners.

Example C-prelude — Platform stiffness from leg stiffness

A useful sanity-check during dimensional synthesis: compute the vertical (heave) stiffness of a 6-UPS Stewart at workspace centre from per-leg axial stiffness. With six identical legs of axial stiffness $k = 5 \times 1 0^{7}$ N/m each (a typical ball-screw electric Stewart), inclined at angle $β$ from vertical (commonly 30–40° at workspace centre):

$K_{z} = 6 k cos^{2} β$

For $β = 35°$ : $K_{z} = 6 \cdot 5 \times 1 0^{7} \cdot 0.671 = 2.0 \times 1 0^{8}$ N/m. The platform sags 1 µm under a 200 N load — three orders of magnitude stiffer than a serial 6R arm of equivalent footprint. The full stiffness matrix has off-diagonal couplings; the heave / surge / sway block is generally well-conditioned and the rotational block degrades faster near singularities.

Example C — Stewart Type II singularity, legs co-planar with platform

A clean illustration of Type II: position a 6-UPS Stewart at $p = (0, 0, 0)$ , i.e. platform plane coincident with base plane (legs horizontal). All six leg lines lie in the same plane as the platform.

The platform Jacobian $J_{x}$ has rows $\hat{l}_{i}^{T} [I - [R p_{i}]_{\times}]$ . The torque component of each row is $\hat{l}_{i} \times R p_{i}$ , which, because all $\hat{l}_{i}$ and $R p_{i}$ lie in the same plane, has only a single non-zero component (the out-of-plane axis). The rank of $J_{x}$ drops to 5; the platform has gained a sixth uncontrolled DOF — it can spin freely about the vertical axis with zero leg motion. $det J_{x} = 0$ . A real platform never reaches this configuration in service, but its proximity is monitored: the singularity-distance metric $\sigma_\min(\mathbf{J}_x)$ is logged at the controller and a fault is raised below 0.1 of its nominal centre-of-workspace value.

4. Families and topologies

Family	Topology	DOF	Originator (year)	Representative platforms
Delta	3-RUU parallelogram	3T (+1R wrist)	Clavel 1988 ICRA	ABB IRB 360 FlexPicker, Fanuc M-1iA / M-3iA, Stäubli TP80, Kawasaki YF003N
Stewart-Gough	6-UPS	6	Gough 1947, Stewart 1965	CAE 7000XR, L3Harris RealitySeven, PI H-840 hexapod, Mechtronix Ascent
Tricept	3-P + serial 3R wrist (hybrid)	5	Neumann 1988 (SMT)	Tricept T606, Loxin TR600 — machine-tool spindle head
Hexa	6-RUS	6	Pierrot 1991	Faster Stewart alternative; research prototypes
Quattro	4-RUU	4 (3T+1R)	Adept (now Omron) 2007	Adept Quattro s650H, s800H — fastest commercial delta (240 picks/min)
Sprint Z3 head	3-PRS	3 (1T+2R)	Sprint (DS Technologie / Starrag) 1998	Sprint Z3 spindle head — 5-axis aerospace machining
2-DOF planar	5R, 5-bar	2	Hunt 1978	Pen plotters, haptic interfaces (Novint Falcon = 3-RRR variant)
Cable-driven (CDPR)	$n$ -cable	up to 6	Albus & Bostelman 1992 (NIST RoboCrane)	IPAnema (Fraunhofer), CableRobot Simulator, FAST 500 m radio telescope feed cabin

(CDPRs are covered in detail under cable-driven-robots — they share the parallel-architecture math but with the wrinkle that cables can only pull, so the workspace is the set where positive cable tensions are feasible.)

Delta-vs-Stewart-vs-Tricept (the three most common)

Metric	Delta (3-RUU)	Stewart-Gough (6-UPS)	Tricept (hybrid)
DOF	3 translation (+ optional wrist)	6	5 (3 parallel + 2 wrist)
Forward kinematics	Closed-form (8th-degree poly, but reduces)	Up to 40 solutions; numerical	Closed-form for parallel stage
Workspace shape	Frustum, “umbrella”	Small, complex	Cylinder + wrist cone
Workspace / footprint	~10 %	~5 %	~25 % (best of the three)
Max acceleration	10–15 g	1–3 g	1–2 g
Stiffness	Moderate (parallelograms flex)	Very high	Very high
Best use	Pick-place, packaging	Sim, alignment, precision	Machining, milling
Cost (2026 commercial)	$30–80k	$100-500 k (s im :$ 2M+)	$200k+

5. Design and analysis methodology

Topology choice is the first design decision. Tsai (1999, Robot Analysis: The Mechanics of Serial and Parallel Manipulators) catalogues 3-RRR, 3-RUS, 3-RPS, 3-PRS, 6-UPS, 6-RUS, 6-PUS, and a dozen more. The convention reads left-to-right from base to platform: R revolute, P prismatic, U universal, S spherical, C cylindrical. The first joint in each leg is actuated (sometimes denoted by underlining in textbooks); the rest are passive. So 3-RUS means three legs, each consisting of one actuated revolute (at the base) followed by passive universal and spherical joints. Selection driven by:

Required mobility (which DOFs need to be active; which are constrained by design).
Workspace shape (translational vs orientational).
Actuator placement constraint (base-mounted = low moving mass; distributed = simpler joints).
Symmetry preferences (3-fold or 6-fold simplifies analysis and reduces calibration parameters).
Singularity locus shape (some topologies put Type II singularities only at the workspace boundary — e.g., the 3-RPS Sprint Z3 — at the cost of mobility).

Dimensional synthesis — once the topology is fixed, choose link lengths, joint placements, and anchor positions to maximise an objective: workspace volume, dexterity, manipulability, stiffness, or some weighted blend. Merlet (2006, Parallel Robots, 2nd ed., Springer) is the standard reference. Optimisation is non-convex; modern practice uses interval analysis or genetic algorithms with workspace + singularity constraints.

Singularity-aware stiffness analysis combines the Jacobian with actuator stiffness $k_{act}$ and structural stiffness of the legs to compute the platform stiffness matrix $K_{p} = J^{T} K_{leg} J$ . Stiffness drops to zero in the singularity direction. Maps of $\sigma_\min(\mathbf{K}_p)$ over the workspace identify “soft” zones for trajectory planning.

Workspace mapping — for an arbitrary topology and dimensional set, the boundary of the workspace is not analytic. Practical workflows discretise the configuration space, evaluate IK feasibility plus singularity proximity plus joint-limit feasibility at each grid point, and render a 3D volume.

Numerical forward kinematics in production code: Newton-Raphson seeded from the previous valid pose, fallback to multi-start with random perturbations. Merlet’s interval-analysis library ALIAS gives certified enclosures but is slower; appropriate for off-line trajectory verification.

Tooling. Open-source: GIM (UPV/EHU) — graphical kinematic analysis for parallel mechanisms. Robotran (UCLouvain) — symbolic multibody dynamics, exports C code. OpenSym. Commercial: Maple (with MapleSim), MATLAB Robotics System Toolbox, MSC Adams. For control: ROS 2 control has experimental parallel-manipulator support; most production controllers are custom (B&R Automation Studio, Siemens Sinumerik for machine tools).

Dynamic analysis of a parallel manipulator is harder than its serial counterpart because of the closed-loop constraints. Three formulations in common use:

Lagrangian with multipliers — explicit kinetic + potential energy plus constraint equations enforced via Lagrange multipliers. General but produces a differential-algebraic equation (DAE) system, index 3, needing specialised solvers (ODASSL, RADAU5).
Newton-Euler — write the equations of motion for each link, eliminate constraint forces via the principle of virtual work. Lower-level but more efficient; Khalil & Dombre (2002) gives the systematic recursive algorithm.
Natural-coordinate method (García de Jalón 1994) — pose each link by Cartesian coordinates of two reference points, all constraint equations become quadratic. Used internally by MSC Adams and Robotran.

For real-time control on production hardware, dynamic models are precomputed offline and run as look-up tables or polynomial fits indexed by platform pose; full inverse dynamics on the actual platform is reserved for high-performance applications (sports-driving simulators with motion-cueing washout, flight-sim G-onset cueing).

6. Real platforms and vendors

Domain	Vendor / model	Notes
Industrial pick-place delta	ABB IRB 360 FlexPicker	1998 launch; 100+ picks/min; payload up to 8 kg; 1.6 m/s end-effector
	Fanuc M-1iA / M-3iA	smaller (0.5 kg) / larger (12 kg) deltas
	Stäubli TP80	food-grade IP67; 80–200 picks/min
	Omron Adept Quattro s800H	4-arm parallelogram; 240 picks/min — fastest commercial
	Kawasaki YF003N	3 kg payload; semiconductor wafer handling
Flight simulator	CAE 7000XR	6-axis Stewart; 60° pitch/roll; certified Boeing 737/777, A320/A330/A350
	L3Harris RealitySeven	competing Level D platform; electric instead of hydraulic
	Mechtronix Ascent	mid-size cockpit Stewart
Driving sim	VI-grade DiM150 / DiM250	high-bandwidth Stewart for F1, OEM dynamics
	Cruden Hexatech	hexapod motion base for race / dev
Precision hexapod	PI H-810 / H-824 / H-840 / H-850 / H-206	nm precision, 10 nm minimum incremental motion; semicon and optics
	Newport HXP1000 / HXP50 / HXP100	optical-bench hexapods
	Symetrie BORA, NOTOX, PARO	astronomy + space simulation
	ALIO Industries hexapods	nanometre-class six-axis
Machine tools	Sprint Z3 (Starrag)	3-PRS spindle head, 5-axis aerospace milling
	Tricept T606 / T805	hybrid 5-axis; 200+ installed worldwide
	Exechon XT300 / XT700	improved Tricept geometry
Surgical	Mazor X / Mazor X Stealth (Medtronic)	spine surgery hexapod
	Brainlab Cirq	small-form-factor surgical arm with parallel sub-system
MEMS / nano hexapods	PI N-Series (N-310, N-381)	piezo-driven nano-hexapods, mm range, < 1 nm resolution

7. Control challenges

Singularity proximity monitoring

The standard online metric is the minimum singular value $\sigma_\min(\mathbf{J}_x)$ of the output Jacobian, computed every controller cycle via incremental SVD or a cheaper approximation (condition-number estimate). A normalised form — divide $\sigma_\min$ by the value at workspace centre — gives a 0–1 health indicator. Thresholds typically:

$\sigma_\min / \sigma_\min^\text{centre}$	Action
> 0.4	Normal operation
0.2 – 0.4	Reduce commanded velocity (often $v \to v \cdot \sigma_\min / 0.4$ )
0.1 – 0.2	Hold pose, raise warning
< 0.1	Safety-rated stop

Online singularity avoidance via redundancy resolution (adding a 7th leg, Merlet 2003) trades architecture cost for control simplicity but is rarely seen commercially.

Force / impedance control

A parallel platform’s impedance control inverts the same Jacobian used for kinematics. Cartesian impedance $F = K_{p} (x_{d} - x) + D_{p} (\dot{x}_{d} - \dot{x})$ maps to actuator forces by $τ = J^{T} F$ . Near Type II singularities $J^{T}$ amplifies platform force, so impedance gains must be reduced. Surgical hexapods (Brainlab Cirq, Mazor X) use admittance control instead — the platform reads external force and moves with it for hand-guiding modes.

Other items

FK numerical convergence. Production controllers cache the previous valid pose as seed; if state is lost (power cycle without homing), the controller must run a multi-start solve. Watchdog: if Newton-Raphson does not converge in <10 iterations, raise fault and stop motion — diverging FK in control is unsafe.
Singularity inside the workspace. Unlike serial arms where singularities are at the boundary, parallel arms have interior Type II singularities. Trajectories must be planned around them, typically by computing a no-go set in workspace space and constraining motion to a connected component. Online: monitor $\sigma_\min(\mathbf{J}_x)$ and reduce velocity / raise fault below threshold.
Synchronised actuator control. Stewart and hexapod platforms drive six actuators simultaneously; controller cycle rate is typically 1–10 kHz with sub-microsecond jitter across axes. Industrial implementations use EtherCAT or SERCOS III for deterministic distributed control.
Cross-coupled axes. Moving one leg changes the constraint on every other leg. A serial arm’s joint-1 controller can ignore joints 2–6 for stability; a parallel platform’s leg controllers cannot — Cartesian-space control with full Jacobian feedback is standard.
Joint clearances in passive joints. Each U-joint and S-joint has manufacturing clearance ~10 µm; these add (unfavourably) into platform pose error. Preloading reduces clearance but adds friction. PI’s high-precision hexapods use crossed-roller flexures instead of conventional U/S joints to eliminate clearance entirely.
Calibration sensitivity. A parallel manipulator has more geometric parameters than a serial arm of the same DOF (each leg’s base + platform anchor positions + nominal length = 7 parameters × 6 legs = 42 parameters for a Stewart). Calibration via CMM or laser tracker (Leica AT960, FARO Vantage) is mandatory for precision platforms — Daney 2003 (Mech. Mach. Theory) shows residuals drop 10× after self-calibration with redundant pose measurements.

Trajectory generation in Cartesian space

For a parallel manipulator, joint-space linear interpolation does not produce straight-line Cartesian motion — it produces curves in task space, and crucially can pass through Type II singularities even when the start and end poses are non-singular. Trajectory generation must be done in Cartesian space (interpolated rotation via SLERP, position via cubic / quintic / S-curve splines), with IK called per timestep to produce actuator setpoints. The IK call is cheap (microseconds); the design constraint is making sure the Cartesian trajectory stays inside the singularity-free workspace.

For high-rate delta robots, the trajectory generator is typically a time-optimal S-curve respecting per-axis velocity and acceleration limits, computed offline for fixed pick-place patterns and replayed at runtime. ABB calls this “QuickMove” on the IRB 360; Adept calls it “PackMaster”.

8. Edge cases and failure modes

Wrong forward-kinematics branch. Newton-Raphson with a poor seed can converge to a physically impossible solution (e.g., the platform “flipped through” the base). Symptom: control commands move the platform in the wrong direction. Fix: seed FK with the previous validated pose; if seed lost, run multi-start and select the branch matching commanded direction.
Type II singularity in service. Catastrophic. Platform gains a DOF, actuator forces saturate to maintain external load, structural collapse possible. Real platforms reserve 15–20% workspace margin and run online singularity proximity monitors.
Constraint redundancy at over-constrained geometries (e.g., parallel-parallelogram delta legs nominally over-constrained). Manufacturing tolerance → preload → thermal stress + accelerated wear. Either machine to tight tolerance and accept the preload, or design in compliance (rubber bushings on parallelogram bars).
Workspace boundary surprise. Computed workspace assumes ideal joint ranges; real U-joints typically yield only ±35–40°, not ±90°. A platform whose calculation assumed ±60° will collide with itself near corners. Always allow 20% workspace margin in spec.
Tool-to-base wiring. End-effector cables, hoses, vacuum lines must route through the parallel structure without snagging the legs. Service-loop design is harder than on a serial arm; many platforms use a slip-ring or rotating-union at the base and route through the central hollow.
Leg-leg collisions. Stewart platforms with large pose excursions can have neighbouring legs collide. Designed-in by minimum-angle constraints in the workspace map.
Backlash in actuators. Cycloidal or harmonic-drive actuators add joint backlash that compounds in the FK; direct-drive linear motors (PI ULC series, Aerotech ANT) are preferred where precision matters.
Thermal drift on long-running platforms. A hydraulic flight-sim Stewart heats the oil reservoir by 15–30 K over an 8-hour shift; leg length drifts by hundreds of micrometres. Solution: dual-channel temperature sensors per leg with first-order thermal compensation in firmware, plus closed-loop pose feedback from cockpit-mounted IMU during fast manoeuvres.
Resonance at structural eigenmodes. Stewart platforms typically have a vertical (heave) mode at 8–15 Hz, lateral modes at 4–8 Hz. Above these frequencies the platform low-passes commands; sim cueing washout filters are tuned below the lowest mode. PI nano-hexapods have first modes near 200 Hz — usable to ~50 Hz bandwidth.
Loss of one leg (Stewart). A single hydraulic actuator failure on a 6-UPS platform leaves it under-constrained — collapses under gravity. Safe-state design uses spring-return brakes on each leg that engage on power loss; passenger-carrying sim platforms have mechanical fallback supports that catch the platform within centimetres of free-fall.
Leg-length sensor failure on FK. If a leg-length sensor (LVDT, encoder, magnetostrictive) drifts or fails, FK gives a plausible-looking but wrong pose. Redundant pose measurement (an IMU on the platform, or external optical tracking) is used in safety-critical platforms.

Reference data — typical parameter ranges

Parameter	Industrial delta	Sim Stewart	Precision hexapod	Machine tool
Workspace diameter (mm)	800–1600	1500–3000	100–400	500–1500
Working height (mm)	200–400	±1200 surge/heave	±50–100	200–500
Payload (kg)	0.5–8 (Quattro: 15)	10000–15000	0.2–250	200–1000
Repeatability	±0.05–0.1 mm	±1 mm	±10–100 nm	±5–10 µm
Max acceleration (g)	5–15	0.6–1	0.1–1	0.2–0.5
Bandwidth (Hz)	10–20	5–8	10–200	5–15
Cost (2026 USD)	30k–80k	1M–10M	30k–150k	200k–800k
Controller cycle (kHz)	1–4	0.5–1	1–10	1–4
Joint angle range (passive U/S)	±30–45°	±35–50°	±15–30°	±20–35°

9. Case studies

ABB FlexPicker IRB 360 (1998-present)

The IRB 360 is the canonical industrial parallel manipulator: a 3-RUU delta with a fourth wrist axis, designed by ABB Robotics in collaboration with Demaurex S.A. (Reymond Clavel’s spin-off) and launched in 1998. As of 2026 it is in its eighth major generation. Spec sheet for the IRB 360-8/1130:

Topology: 3 parallelogram arms 120° apart + central wrist axis (4-DOF: x, y, z, rot-z)
Payload: 8 kg (1130 mm reach variant); 1 kg and 3 kg variants also available
Workspace: cylindrical, Ø 1130 mm × 200 mm working height
Repeatability: ±0.1 mm
End-effector speed: up to 10 m/s
End-effector acceleration: up to 100 m/s² (≈10 g)
Cycle time: 0.3 s (25-305-25 mm pick-and-place benchmark) → >120 picks/min
IP67 wash-down option for food and pharma
Controller: ABB OmniCore C30; supports vision-guided picking (Cognex Insight 7000) and conveyor tracking with sub-millisecond synchronisation

Why it works: the parallelogram lower arm constrains the platform to pure translation, simplifying both FK (closed-form) and kinematic calibration. The high acceleration comes from <1 kg moving mass against base-mounted servomotors, each driving an upper arm via a high-stiffness cycloidal reducer.

CAE 7000XR full-flight simulator

CAE 7000XR is the world’s most-installed Level D full-flight simulator (Boeing 737, 777, 787; Airbus A320, A330, A350). The motion system is a 6-axis Stewart platform with hydraulic actuators:

6 hydraulic legs, stroke ~1.5 m each, max velocity 1 m/s, peak force 100+ kN
Workspace: ±60° pitch / roll, ±25° yaw, ±1.2 m surge / sway / heave at platform centre
Payload: complete cockpit + visual + crew + instructor ≈ 12 000 kg
Bandwidth: 5–6 Hz (matches human vestibular sensitivity)
Certification: EASA / FAA / TCCA Level D — including motion-cueing fidelity tests

The whole simulator hooks into a real-time aerodynamic model (running on a CAE Tropos-6000XR image generator + flight model); washout-filtered motion commands are fed to the controller, which computes leg lengths via the easy IK direction and drives the servo valves. There is no online FK in service — the platform is run open-loop on leg-length commands with pressure feedback for force control.

PI Hexapod H-840

Physik Instrumente H-840 is a precision 6-UPS hexapod for optics alignment, semiconductor metrology, and beamline experiments:

Travel: ±50 mm linear, ±15° angular
Resolution: 40 nm linear, 2 µrad angular (open loop); 10 nm linear at minimum incremental motion (closed loop)
Load capacity: 250 N
Drive: brushless DC servos with backlash-free ball screws; cross-roller-flexure-style joints to eliminate clearance
Controller: PI C-887.52 hexapod controller; FK / IK in firmware; supports tool-center-point definition; SDK in C, Python, MATLAB, LabVIEW

The H-840 is the workhorse for semiconductor metrology (wafer alignment in inspection tools — KLA, Applied Materials), astronomy secondary-mirror support (ESO VLT auxiliary), and synchrotron beamline sample positioning. Its 10 nm precision is achieved by careful joint design (flexures eliminate backlash), low thermal mass, and on-board kinematic calibration from a built-in interferometer at each leg.

10. Cross-references

manipulator-design — serial-arm counterpart; topology trade-offs the parallel architecture inverts.
kinematics-dh — DH parameters work for serial chains; parallel manipulators need closed-loop-aware formulations (loop equations, screw theory).
cable-driven-robots — parallel architecture with cables (tension-only); shares Jacobian + singularity math, adds wrench-feasibility workspace.
end-effectors — what the platform actually carries; vacuum cups, jaw grippers, weld torches, optical mounts.
trajectory-generation — Cartesian-space planning is mandatory for parallel manipulators (joint-space interpolation does not yield smooth Cartesian motion).
manipulability-workspace — manipulability ellipsoid, dexterity index — both extend naturally to parallel architectures via $J_{q}^{- 1} J_{x}$ .
statics-fundamentals — load-path analysis through parallel chains; equilibrium with multiple unknowns.
structural-analysis — leg stiffness, beam buckling under compressive load (Stewart legs in compression can buckle).

11. Citations

Foundational textbooks

Merlet, J.-P. (2006). Parallel Robots, 2nd ed. Springer. ISBN 978-1-4020-4133-4. Canonical reference; covers kinematics, singularities, workspace, design.
Tsai, L.-W. (1999). Robot Analysis: The Mechanics of Serial and Parallel Manipulators. Wiley. ISBN 978-0-471-32593-2. Textbook treatment with extensive parallel-manipulator topology catalogue.
Hunt, K. H. (1978). Kinematic Geometry of Mechanisms. Oxford University Press. Screw-theory foundation used by every parallel-manipulator mobility analysis.
Briot, S., & Khalil, W. (2015). Dynamics of Parallel Robots — From Rigid Bodies to Flexible Elements. Springer Mechanism and Machine Science series.

Primary papers

Gough, V. E. (1947). Universal tyre-testing machine. (Dunlop internal report; later cited in Proc. 9th Int. Tech. Congr. F.I.S.I.T.A., 1962.) Origin of the 6-leg platform.
Stewart, D. (1965). “A Platform with Six Degrees of Freedom.” Proc. Institution of Mechanical Engineers 180:371–386. The naming paper.
Clavel, R. (1988). “DELTA, a fast robot with parallel geometry.” Proc. 18th Int. Symp. on Industrial Robots / ICRA 1988, 91–100. Delta debut.
Pierrot, F., Reynaud, C., & Fournier, A. (1991). “DELTA: a simple and efficient parallel robot.” Robotica 9:105–109. Plus Hexa robot.
Gosselin, C., & Angeles, J. (1990). “Singularity Analysis of Closed-Loop Kinematic Chains.” IEEE Transactions on Robotics and Automation 6:281–290. The three-type singularity classification.
Innocenti, C., & Parenti-Castelli, V. (1990). “Direct Position Analysis of the Stewart Platform Mechanism.” Mechanism and Machine Theory 25:611–621. Algebraic FK for 6-3 Stewart.
Husty, M. L. (1996). “An algorithm for solving the direct kinematics of general Stewart-Gough platforms.” Mechanism and Machine Theory 31:365–380. 40-solution FK polynomial.
Raghavan, M. (1993). “The Stewart Platform of General Geometry has 40 Configurations.” ASME J. Mech. Design 115:277–282.
Daney, D. (2003). “Kinematic calibration of the Gough platform.” Robotica 21:677–690.
Merlet, J.-P. (2004). “Solving the forward kinematics of a Gough-type parallel manipulator with interval analysis.” Int. J. Robotics Research 23:221–235.

Recent papers (post-2010)

Briot, S., & Bonev, I. (2010). “Singularity Analysis of Zero-Torsion Parallel Mechanisms.” Mech. Mach. Theory 45:712–724.
Bonev, I. (2014). “ParalleMIC — The Parallel Mechanisms Information Center.” Online catalogue of parallel topologies and properties, http://www.parallemic.org.
Carbone, G., & Ceccarelli, M. (2007). “Comparison of Indices for the Performance Evaluation of Parallel Manipulators.” Int. J. Mech. Sci. 49:1245–1255.
Cardona, M., et al. (2020). “A Survey on Industrial Applications of Parallel Robots.” Int. J. Industrial Engineering Computations 11:603–624.
Patel, Y., & George, P. M. (2012). “Parallel Manipulators Applications — A Survey.” Modern Mechanical Engineering 2:57–64.

Standards and assessment

ISO 9283:1998 — “Manipulating industrial robots — Performance criteria and related test methods.” Applies to parallel arms via the TCP-pose definitions.
ISO 10218-1:2025 / 10218-2:2025 — caged-industrial safety standards; apply to delta pick-place cells.
ASTM F3144 — performance metrics for parallel-kinematic machines (draft as of 2026).

Vendor literature

ABB. IRB 360 FlexPicker Product Specification, rev. 2024. ABB Robotics.
Physik Instrumente. H-840 Hexapod Datasheet, rev. 2025.
CAE. 7000XR Full Flight Simulator — Technical Description, 2024.
Fanuc. M-1iA / M-3iA Parallel-Link Robot Brochure, 2024.
Stäubli. TP80 FAST Picker Datasheet, 2024.
Newport Corporation. HXP1000 Hexapod User Manual, 2023.

Compendium

Explorer

Parallel Manipulators (Stewart-Gough, Delta, Hexapod) — Robotics Reference