Vibration, Damping & Chatter Control for Robot Arms
1. At a glance
Robot arms are not rigid. Every serial-link manipulator has structural compliance plus multi-mode vibration, and the resulting end-effector motion is typically far worse than the joint encoders suggest — the encoder sees the motor side of a flexible drivetrain, while the tool sees the load side at the end of a long lever. A 1 m reach arm with a 100 N·m/rad joint stiffness sees roughly 0.6° of static deflection per 1 N·m of tip load; under transient excitation that deflection rings.
Primary sources of vibration in industrial arms:
- Gearbox / harmonic drive flexibility — finite torsional stiffness, typically 1e4–1e5 N·m/rad
- Link bending — long aluminium / steel links act as cantilever beams
- Joint backlash — discrete impacts at zero-crossings produce broadband excitation
- Motor torque ripple — pole-pair × electrical frequency content
- Command bandwidth exceeding mechanical resonance — controller demanding 50 Hz on a 12 Hz structure
- External excitation — milling cut force, contact transients, payload swing
Consequences: poor surface finish in robotic machining, position overshoot and oscillation at the end-effector, accelerated wear on gearboxes, fatigue cracking on welded link joints, audible noise. Critical in: robotic machining (Comau Smart5 NJ, KUKA KMR Quantec, Stäubli RX260), semiconductor lithography stages, surgical and microsurgery (DLR MIRO, KUKA LBR Med, Smart Tissue Autonomous Robot — STAR), aerospace drilling, and very long-reach arms (KUKA KR 1000 Titan).
This note covers the physics, the canonical mitigation techniques (input shaping, active damping, tuned mass dampers, notch filtering, trajectory smoothing), and the industry tooling around them.
2. First principles
2.1 Lumped two-mass model
The simplest useful model for a single flexible joint is the two-mass / two-inertia system: motor inertia J_m connected through a torsional spring k (and damper c) to load inertia J_L.
[J_m] —k,c— [J_L]
| |
τ_m τ_ext
Equations of motion (small-signal, linearised):
J_m·θ̈_m + c·(θ̇_m − θ̇_L) + k·(θ_m − θ_L) = τ_m
J_L·θ̈_L + c·(θ̇_L − θ̇_m) + k·(θ_L − θ_m) = −τ_ext
This yields two characteristic frequencies (rad/s):
- Antiresonance (load fixed, motor moves):
ω_anti = √(k / J_L) - Resonance (free-free):
ω_res = √(k · (J_m + J_L) / (J_m · J_L))
Note ω_res ≥ ω_anti always, with equality only when J_m → ∞. The ratio ω_res / ω_anti = √(1 + J_L/J_m) sets how much the load inertia separates the modes; gear-reduced robots typically have reflected J_m comparable to or larger than J_L, so the two frequencies sit close together.
2.2 Modal analysis of a multi-link arm
For a 6- or 7-DOF arm with n flexible coordinates, mass M(q) and stiffness K(q) matrices, the linearised free-vibration eigenproblem at a configuration q₀ is
(K − ω²·M) · φ = 0
Eigenvalues ω_i² give modal frequencies; eigenvectors φ_i give mode shapes. Typical industrial 6-DOF arm: first mode 5–50 Hz, second mode 15–80 Hz, with configuration-dependent variation of up to 2× as the arm changes posture (compare [[Robotics/manipulator-design]] on posture-dependent stiffness, [[Engineering/vibration-dynamics]] for the general modal framework).
2.3 Damping ratios
| Mechanism | Damping ratio ζ |
|---|---|
| Welded steel structure (link) | 0.005 – 0.01 |
| Bolted joint (flanges, mounts) | 0.01 – 0.02 |
| Gearbox (harmonic / planetary, oil-bath) | 0.05 – 0.15 |
| Cycloid drive | 0.05 – 0.10 |
| Belt drive | 0.02 – 0.05 |
| Roller bearings (rolling friction) | 0.01 – 0.05 |
The bulk of structural damping in a robot comes from the gearboxes and bearings, not from the welded links. A “rigid base, flexible drive” model captures most of the observed behaviour at first-mode frequencies.
2.4 Frequency Response Function (FRF)
The canonical experimental characterisation is the bump test:
- Lock the arm in a representative posture
- Strike with an instrumented impulse hammer (PCB 086C03 or similar) at the end-effector
- Record acceleration at the TCP with a tri-axial accelerometer (PCB 356A15)
- Compute
H(jω) = A(jω) / F(jω)via FFT
Peaks in |H| are resonances; phase drops of 180° confirm structural modes. Coherence near 1 indicates a clean measurement. Tools: B&K LAN-XI front-end, Polytec PSV-500 laser Doppler vibrometer for non-contact, Siemens LMS Test.Lab for curve-fitting, ME’scope for animated mode-shape extraction.
2.5 Chatter (regenerative self-excited vibration)
In machining, chatter is a self-sustained oscillation driven by the regenerative cutting-force feedback: the current tooth pass cuts into a surface left wavy by the previous pass, modulating chip thickness and hence force, which excites the same vibration mode that produced the wave. The classical Tobias–Fishwick (1958) and later Tlusty / Altintas analysis gives a stability lobe diagram — a plot of axial cut depth a versus spindle speed n separating stable from unstable cutting.
Absolute stability limit (chatter-free at any speed) is approximately
a_lim = −1 / (2 · K_f · Re[G(jω_c)]_min)
where K_f is specific cutting energy, G(jω) is the directional TCP compliance FRF, and the minimum is taken over ω. Robotic machining: arm TCP compliance can be 100× that of a milling-machine spindle, so a_lim for a robot is typically 0.5–3 mm versus 5–20 mm for a CNC mill in the same material (see [[Engineering/machining]]).
2.6 Input shaping (Singer 1990)
A pre-filter that convolves the desired command with a short impulse train. For a mode at (ω_n, ζ), the Zero-Vibration (ZV) shaper is
┌ A₁ A₂ ┐ A₁ = 1/(1 + K) K = exp(−ζπ / √(1−ζ²))
│ t₁ t₂ │ A₂ = K/(1 + K)
t₁ = 0
t₂ = π / (ω_n · √(1−ζ²)) ≈ T_d / 2
Convolving the reference with [A₁ δ(t) + A₂ δ(t−t₂)] makes the net response at frequency ω_n vanish — the second impulse arrives 180° out of phase with the residual ring from the first. Cost: an added delay t₂ (half the damped period) to the trajectory. Variants:
- ZVD (Zero-Vibration-and-Derivative) — 3 impulses, twice the delay, robust to ω uncertainty up to ±10%
- EI (Extra-Insensitive) — 3 impulses, accepts ε% residual vibration for much wider tolerance band (±20% on
ω_n) - SI (Specified-Insensitivity) — designed to user-specified frequency band
2.7 Active damping via collocated acceleration feedback
If an accelerometer is collocated with an actuator (same physical point, same axis), the loop τ = −k_a · ẍ is unconditionally stable by passivity, regardless of the plant order. Used in Stewart platforms, semiconductor stages, and the DLR LWR III joints (where joint-torque sensors provide a collocated measurement of the structural deflection).
2.8 Tuned Mass Damper (TMD)
A passive auxiliary mass-spring-damper attached at a mode antinode. Optimal Den Hartog tuning for primary mass M and absorber mass m, mass ratio μ = m/M:
ω_TMD / ω_primary = 1 / (1 + μ)
ζ_TMD,opt = √(3μ / (8(1+μ)³))
Adding 2–5% mass typically raises the effective modal damping from ~0.01 to ~0.05–0.10 over a useful bandwidth.
2.9 Posicast (Smith 1957)
Predecessor to input shaping: a half-cycle delayed pulse used to cancel the first overshoot. Equivalent to a ZV shaper but originally framed in control-theoretic rather than signal-processing terms.
3. Worked examples
3.1 Example A — Input shaping a 6-DOF arm
A medium-payload arm (e.g. KUKA KR 16 R2010) shows a dominant mode at f_n = 15 Hz in joint 4 (wrist roll), with measured damping ζ = 0.05.
ω_n = 2π · 15 = 94.25 rad/s
ω_d = ω_n · √(1 − 0.05²) = 94.13 rad/s
T_d = 2π / ω_d = 66.7 ms
ZV shaper:
K = exp(−0.05 · π / √(1 − 0.05²)) = exp(−0.1574) = 0.854
A₁ = 1 / (1 + 0.854) = 0.539
A₂ = 0.854 / 1.854 = 0.461
t₂ = T_d / 2 = 33.4 ms
ZVD shaper (3 impulses):
A₁ = 1/(1+K)² = 0.290
A₂ = 2K/(1+K)² = 0.495
A₃ = K²/(1+K)² = 0.215
delays: 0, t₂, 2·t₂ = 0, 33.4, 66.8 ms
ZV cancels the mode exactly when ω_n is known precisely; ZVD adds 33 ms more delay but keeps residual vibration below 5% across ±10% uncertainty on ω_n. Choice depends on whether payload (and therefore ω_n) is known and constant.
3.2 Example B — Tuned mass damper at end-effector
A 100 kg-effective-modal-mass arm vibrating at 8 Hz with ζ = 0.01 (poorly damped, e.g. a long-reach pick-and-place arm). Add a 2 kg TMD.
μ = 2 / 100 = 0.02
ω_TMD,opt = 2π · 8 · (1 / (1 + 0.02)) = 49.27 rad/s → f = 7.84 Hz
k_TMD = m · ω_TMD² = 2 · 49.27² = 4855 N/m
ζ_TMD,opt = √(3 · 0.02 / (8 · 1.02³)) = √(0.00721) = 0.0849
c_TMD = 2 · m · ω_TMD · ζ_TMD = 2 · 2 · 49.27 · 0.0849 = 16.7 N·s/m
The two-DOF coupled system has effective primary-mode damping ~0.06 (six-fold increase), at the cost of a second nearby peak and a 2% mass penalty. For machining robots this is often built into the spindle mount.
3.3 Example C — Chatter stability lobe for a robotic mill
Arm milling aluminium at the TCP with measured directional compliance k_TCP = 5 × 10⁶ N/m, modal mass m = 100 kg, ζ = 0.02.
ω_n = √(k / m) = √(5e6 / 100) = 223.6 rad/s = 35.6 Hz
K_f (Al 6061, climb) ≈ 800 N/mm² = 8 × 10⁸ N/m²
Re[G(jω)]_min ≈ −1 / (4·k·ζ) at ω = ω_n·√(1−ζ²)
≈ −1 / (4 · 5e6 · 0.02) = −2.5 × 10⁻⁶ m/N
a_lim ≈ −1 / (2 · K_f · w_eff · Re[G]_min) (with w_eff =1 mm chip width)
≈ 1 / (2 · 8e8 · 1e-3 · 2.5e-6) = 0.25 mm
So at worst-case spindle speed the robot can cut only 0.25 mm axial depth chatter-free. Tuning spindle speed to a stability lobe (where regeneration phase shifts) can multiply this by 3–5× without changing the arm. For a 4-flute end mill running at n = 60 · ω_n / (Z · k) rpm (with k an integer lobe number), the first useful pocket sits near n ≈ 60·35.6/(4·1) ≈ 534 rpm, second pocket near 267 rpm, third near 178 rpm — counter-intuitively, slow speeds help.
4. Sources of vibration
| Source | Excitation frequency | Notes |
|---|---|---|
| Harmonic-drive gearmesh | 2 × pole pairs × ω_motor | Two-lobe wave generator, characteristic “two-per-rev” ripple |
| Planetary gearmesh | N_planet × ω_carrier | Adds high-frequency tonal content |
| Motor cogging | N_pole-pair × ω_electrical | Reduced by skewed magnets / fractional-slot windings |
| Motor torque ripple | 6 × ω_electrical (BLDC) | Inverter switching also injects PWM sidebands |
| Link bending mode | ω_n ∝ (1/L²)·√(EI/(ρA)) | Cantilever; first mode falls as 1/L² |
| Joint backlash | broadband | Limit-cycle at low speed, especially with integrator |
| Cable / harness drag | distributed, low-frequency | Mass distribution changes with arm pose |
| Tool / payload pendulum | √(g/L_eff) | Often <2 Hz; aliases into rate-loop bandwidth |
| Milling cut force | tooth-pass × spindle rpm | Regenerative; see chatter |
| Contact transients | impulse / broadband | Excites all modes simultaneously |
| Jerky trajectory | Fourier content of profile | Discontinuous accel rings every flexible mode |
A long thin link of length L = 1 m, square cross-section 50 × 50 mm, aluminium (E = 70 GPa, ρ = 2700 kg/m³): area A = 2.5e-3 m², second moment I = a⁴/12 = 5.21e-7 m⁴. First cantilever frequency f₁ = (1.875²/(2π))·√(EI/(ρAL⁴)) ≈ 23 Hz, well inside a typical 50 Hz position-loop bandwidth.
5. Mitigation techniques
| Technique | Type | Bandwidth | Cost | Notes |
|---|---|---|---|---|
| Trajectory smoothing (S-curve, min-jerk) | Open-loop | DC – f_n | Path planner CPU | Always cheap; first line of defence (see [[Robotics/trajectory-generation]]) |
| Input shaping (ZV / ZVD / EI / SI) | Open-loop pre-filter | Tuned to f_n | Added delay (T_d/2 .. T_d) | Singer 1990 / Singhose 1994 |
| Notch filter on torque command | Open-loop | At f_n | Phase loss near notch | Needs accurate ω_n |
| Active acceleration feedback | Closed-loop | DC – f_actuator | Sensor + DSP | Stable if collocated |
| Force-torque feedback | Closed-loop | DC – ~50 Hz | F/T sensor (ATI Mini45) | Enables impedance control (see [[Robotics/impedance-control]]) |
| Compliant gripper / wrist | Passive | ~50–500 Hz | Mech complexity | Absorbs collision transient |
| Variable-stiffness joints | Active mechanical | Slow | High complexity | Festo VS-Knee, DLR VS-Joint |
| Damped base mounts | Passive | At base mode | Footprint | Newport, Bilz, isolation pads |
| Posture optimisation | Path planning | Trajectory-level | None | Avoid low-stiffness configurations |
| Tuned Mass Damper at TCP | Passive | Tuned | Added mass | 2–5% mass penalty |
Trajectory smoothing is always done — replace trapezoidal velocity profiles with S-curves (limited jerk) or quintic / minimum-jerk profiles. The Fourier content of the command then falls off as 1/ω³ instead of 1/ω, suppressing high-frequency excitation by 40 dB at the first mode.
Notch filters look like (s² + 2ζ_n ω_n s + ω_n²) / (s² + 2ζ_p ω_n s + ω_n²) with ζ_n < ζ_p. They are simpler than input shaping but introduce phase loss near the notch and require accurate ω_n (within ±5%).
Variable-stiffness joints (DLR, IIT, Vienna) deliberately introduce mechanical compliance with electrically tunable stiffness, trading bandwidth for safety and energy storage (see [[Robotics/pid-control]] for the contrast with stiff position control).
6. Industry applications
| Application | Representative platform | Dominant problem |
|---|---|---|
| Robotic machining | Comau Smart5 NJ4, KUKA KR Quantec, Stäubli RX160 / RX260 | Chatter; arm stiffness ~100× lower than CNC mill |
| Aerospace drilling / fastening | KUKA KR 600 R2830, FANUC M-2000iA | Long reach + high payload; large structural modes |
| Semiconductor stage | PI H-840 hexapod, Aerotech ABL gantry | Nano-precision; passive + active isolation tables |
| Microsurgery | DLR MIRO, KUKA LBR Med, Intuitive da Vinci, STAR | Hand tremor + structural ring; need <0.1 mm RMS |
| Large reach | KUKA KR 1000 Titan, FANUC M-2000iA/2300 | First mode <5 Hz; severely limits bandwidth |
| Cobots | Universal Robots UR10e, FANUC CRX-10iA | Flexible-joint design choice for safety + lower stiffness |
Robotic machining is the hardest economic case: customers want CNC-like accuracy at robot prices, but the arm stiffness shortfall pushes chatter-limited material removal rate to 10–20% of a real mill’s. Comau’s chatter-aware path planner and ABB’s AppCanal (Application Channel) toolkit address this by adapting feed and spindle speed to a learned arm-FRF.
7. Tools and software
| Category | Tool | Use |
|---|---|---|
| Input shaping | MATLAB Singhose toolbox, Simulink, custom shaper plugins | Design and simulate shapers |
| Controller built-in | FANUC IFR (Intelligent Functions for Robots), KUKA SmartPAD vibration setting | Production-grade shaper deployment |
| Modal analysis HW | B&K LAN-XI, PCB hammers and accelerometers | Bump test acquisition |
| Modal analysis SW | ME’scope, Siemens LMS Test.Lab, DewesoftX | Curve-fit modes, animate shapes |
| Non-contact vibrometry | Polytec PSV-500, OptoMET SWIR | Mode shape on lightweight or hot structures |
| FEA | ANSYS Mechanical, Abaqus, MSC Nastran | Modal + harmonic analysis (see [[Engineering/fem-fea]]) |
| Multi-body | MSC Adams, Simpack, Robotran | Flexible-body manipulator simulation |
| Control prototyping | Simulink + dSPACE, ROS 2 ros2_control with custom shaper plugin | Real-time deployment |
| Chatter prediction | CutPro (UBC / MAL Inc.), MAL Inc. Tool Inspector, Liebherr Tool | Stability lobe generation from FRF |
For a ROS 2 deployment, the typical pattern is a controller_manager plugin that wraps the trajectory interface with a configurable shaper; payload changes are accommodated by re-identifying ω_n from a quick step-response test and updating the impulse-train coefficients online.
8. Edge cases and failure modes
- Shaper detuning under payload change — a 10 kg payload swing can shift
ω_nby 30%; ZV residual rises from 0% to 15%. Mitigation: switch to ZVD or EI shaper, or schedule shaper coefficients on payload. - Multiple closely spaced modes per axis — cascade two shapers tuned to each mode (additive delay), or use a single multi-mode SI shaper with notches at all frequencies.
- Sensor noise amplification — acceleration feedback differentiates noise; mandatory lowpass at 2–5× control bandwidth.
- Backlash limit cycle — integral action plus backlash creates limit cycles at frequencies
ω = √(K_I / J); mitigation via dead-band compensation or backlash-free harmonic drives. - Teach-pendant jog excitation — operators jogging at low speed often excite the first mode; controllers should jerk-limit jog commands.
- Cable harness mass redistribution — re-routing the harness shifts
ω_nby 5–15%; re-identify after maintenance. - Thermal warm-up — gearbox oil viscosity changes
ζandkover the first 30 min of operation; shapers retuned from cold are sub-optimal once warm. - Crash transient — collision impulses excite every mode; abort and re-home is the standard recovery.
- Hidden out-of-plane modes — uniaxial bump tests miss orthogonal modes; always do tri-axial FRF.
- Singular postures — near kinematic singularities, structural stiffness drops dramatically (small joint torque maps to large TCP displacement); path planners should add a stiffness-index cost term.
9. Case studies
9.1 NASA Canadarm SRMS — Singer’s input shaping (1990)
The Space Shuttle Remote Manipulator System (SRMS / Canadarm) was a 15 m, 410 kg, 6-DOF arm with first structural mode at 0.1–0.2 Hz when extended and loaded — too low to control by simple PID without persistent end-point oscillation. Neil Singer’s MIT thesis (1989, published 1990) formalised the convolution-with-impulse-train approach as input shaping, and the SRMS became its showcase application. Subsequent shaped-command operation of the Shuttle RMS cut grapple-fixture capture-envelope oscillation from ~0.3 m to <0.05 m at no extra hardware cost.
9.2 DLR LWR III / KUKA LBR iiwa — flexible-joint impedance control (Albu-Schäffer 2007)
The DLR Light-Weight Robot III, productised as the KUKA LBR iiwa, was explicitly designed with joint-torque sensors at every joint and intentionally compliant harmonic-drive gearboxes. Albu-Schäffer’s model-based controller treats the arm as a flexible-joint system (per-joint two-mass model from §2.1) and uses the joint-torque measurement to damp the structural mode actively. The result: a 7-DOF cobot with controlled Cartesian impedance, intrinsic compliance for human safety, and active suppression of the same flexibility it deliberately introduces (see [[Robotics/impedance-control]]).
9.3 Comau Smart5 NJ machining cell + ABB AppCanal
Comau’s robotic machining cells (Smart5 NJ4 series) couple a high-payload arm with an HSD spindle and a chatter-aware path planner. The system measures the configuration-dependent TCP FRF at job-setup time, computes stability lobes per posture, and adapts feed-per-tooth and spindle rpm along the toolpath to stay inside chatter-free regions. ABB’s AppCanal whitepapers describe a similar approach for the IRB 6700; both achieve ~3× improvement in chatter-free material removal rate versus naive fixed-feed strategies.
10. Cross-references
[[Robotics/manipulator-design]]— stiffness budgets, link sizing, posture-dependent stiffness[[Robotics/dynamics-rigid-body]]— Lagrangian formulation, M(q), C(q,q̇), G(q)[[Robotics/trajectory-generation]]— S-curve, minimum-jerk, quintic profiles[[Robotics/impedance-control]]— joint-torque-based active damping (LBR iiwa)[[Robotics/pid-control]]— bandwidth limits imposed by structural modes[[Engineering/vibration-dynamics]]— general modal framework, FRF interpretation[[Engineering/structural-dynamics]]— multi-DOF eigenproblem, mode shapes[[Engineering/machining]]— chatter in CNC milling and turning[[Engineering/fem-fea]]— ANSYS / Abaqus modal extraction
11. Citations
- Singer, N. C. (1990). Residual Vibration Reduction in Computer Controlled Machines. MIT PhD thesis (1989); ASME J. Dyn. Syst. Meas. Control 112(1):76–82. Foundational input-shaping paper.
- Singhose, W. E., Seering, W. P., Singer, N. C. (1994). “Residual vibration reduction using vector diagrams to generate shaped inputs.” ASME J. Mech. Design 116(2):654–659. EI shaper and robustness analysis.
- Smith, O. J. M. (1957). “Posicast control of damped oscillatory systems.” Proc. IRE 45(9):1249–1255. Pre-shaper precursor.
- Tobias, S. A., Fishwick, W. (1958). “Theory of regenerative machine tool chatter.” The Engineer, 199:199–203 / 238–239. Classical chatter formulation.
- Altintas, Y. (2012). Manufacturing Automation: Metal Cutting Mechanics, Machine Tool Vibrations, and CNC Design, 2nd ed. Cambridge. Canonical chatter / stability-lobe reference; underpins CutPro software.
- Albu-Schäffer, A., Ott, C., Hirzinger, G. (2007). “A unified passivity-based control framework for position, torque and impedance control of flexible joint robots.” Int. J. Robotics Research 26(1):23–39. DLR LWR III / iiwa control architecture.
- Singhose, W. (2015). “Command shaping for flexible systems: a review of the first 50 years.” Int. J. Precision Engineering & Manufacturing 16(13):2723–2731. Modern review.
- Tutter, A. et al. (2018). “Input-shaping for industrial robots: a benchmark.” IEEE/ASME Trans. Mechatronics. Comparative evaluation on KUKA and ABB platforms.
- Hassan, M. et al. / KUKA whitepapers. Vibration-suppression deployment notes for KR Quantec and LBR iiwa.
- Bruel & Kjaer / Siemens LMS application notes. Standard modal-analysis procedure, bump-test methodology, FRF curve-fitting (rational fraction polynomial, Polymax).
End of Tier 2 reference. Units: SI throughout (Hz, rad/s, N·m, kg, m, s). Originator + year cited for each foundational technique. Platforms named are real: KUKA KR 1000 Titan, KUKA LBR iiwa, KUKA KR 16 / KR 600 / KR Quantec, Stäubli RX160 / RX260, Comau Smart5 NJ4, FANUC M-2000iA / CRX-10iA, ABB IRB 6700, Universal Robots UR10e, DLR LWR III / MIRO, PI H-840 hexapod, Aerotech ABL.