Manipulability Ellipsoid and Kinematic Singularities of Robot Arms

What it is and where it is used

Every non-redundant robot arm maps a vector of joint velocities q̇ to an end-effector twist v through the geometric Jacobian: v = J(q)·q̇. If you take every unit-norm joint-velocity vector (‖q̇‖ = 1) and push it through J, the resulting set of achievable end-effector velocities is an ellipsoid in task space — the velocity manipulability ellipsoid. A fat, round ellipsoid means the tip can move fast and equally well in all directions; a long thin sliver means the arm is fast one way and sluggish another.

• Motion planning: steer trajectories toward high-manipulability postures so servos never saturate.
• Redundancy resolution: a 7-DOF arm (KUKA LBR iiwa, Franka Emika) uses its null space to maximize w(q) while tracking the task.
• Mechanism design: Yoshikawa showed a planar 2R arm's manipulability is maximal when link lengths are equal (l₁ = l₂) and the elbow sits at 90°.
• Teleoperation and surgery: visualizing the ellipsoid warns operators before they drive into a singularity.

How it works: from the Jacobian to the ellipsoid

The ellipsoid is built directly from the singular value decomposition J = U·Σ·Vᵀ. The columns of U (left singular vectors) give the directions of the ellipsoid's principal axes in task space; the singular values σ₁ ≥ σ₂ ≥ … ≥ σₘ ≥ 0 on the diagonal of Σ give the lengths of the corresponding semi-axes.

Equivalently, the ellipsoid is defined by the symmetric matrix M = J·Jᵀ. The eigenvectors of M are the principal axes, and its eigenvalues are the squared singular values λᵢ = σᵢ². The set {v : vᵀ (JJᵀ)⁻¹ v ≤ 1} is exactly the velocity ellipsoid.

Yoshikawa's scalar manipulability measure is the ellipsoid's volume (up to a constant): w(q) = √det(J·Jᵀ) = σ₁·σ₂·…·σₘ. For a square Jacobian this simplifies to w = |det J|. Because it is a product of singular values, w collapses to zero the instant any single σᵢ hits zero — the precise algebraic signature of a singularity, where the ellipsoid flattens into a lower-dimensional disk.

Key quantities and a worked example

Take the classic planar 2R arm with link lengths l₁, l₂ and joint angles θ₁, θ₂. Its position Jacobian has determinant det(J) = l₁·l₂·sin(θ₂), so the manipulability measure is w = |l₁·l₂·sin(θ₂)|.

• With l₁ = l₂ = 0.5 m and elbow θ₂ = 90°: w = 0.5·0.5·sin(90°) = 0.25 m² — the maximum, an isotropic near-circular ellipsoid.
• Elbow at θ₂ = 30°: w = 0.25·sin(30°) = 0.125 m² — half the dexterity, ellipsoid elongated.
• Fully extended, θ₂ = 0° (or 180°): sin = 0, so w = 0. The arm is at the boundary of its workspace — the notorious elbow (stretch) singularity. The ellipsoid becomes a line segment; the tip cannot move radially outward at all.

The condition number κ = σ_max/σ_min quantifies the same story numerically. κ = 1 is a perfect sphere (isotropic); at θ₂ = 90° above, κ ≈ 2.4 for a typical 2R; approaching extension κ → ∞. Practitioners often keep κ below about 10 to guarantee well-behaved inverse-kinematics numerics.

Design and operation in practice

Because the velocity ellipsoid has a dual force ellipsoid defined by (JJᵀ)⁻¹, the two share the same axes but have reciprocal lengths. This kinetostatic duality is a hard trade-off engineers exploit: directions of high speed are directions of low force, and vice versa. A robot polishing a surface wants stiffness (short velocity axis, long force axis) normal to the part but speed tangent to it — so you orient the arm to align the ellipsoid accordingly.

• Redundant arms use gradient projection: q̇ = J⁺·v + (I − J⁺J)·∇w, adding null-space motion that climbs the manipulability gradient without disturbing the task.
• Damped least-squares (Levenberg–Marquardt) inverse: q̇ = Jᵀ(JJᵀ + λ²I)⁻¹ v. The damping factor λ (often scheduled from σ_min) trades tracking accuracy for bounded joint speeds near singularities — the standard fix so servos don't demand thousands of deg/s.
• Workcell layout: place tasks in the region where w is largest; avoid the shoulder-, elbow-, and wrist-singularity surfaces that carve the workspace.

Comparison to related measures and components

The manipulability ellipsoid is one member of a family of Jacobian-based indices, and choosing among them matters:

• Manipulability w vs. condition number κ: w measures ellipsoid volume (can be large even if the shape is a thin cigar), while κ measures eccentricity. A pose can have healthy w yet a terrible κ, so many controllers monitor σ_min directly as the most honest singularity alarm.
• Velocity vs. dynamic manipulability: Yoshikawa later defined the dynamic manipulability ellipsoid using the inertia matrix M(q), √det(J M⁻¹ M⁻ᵀ Jᵀ), which weights directions by how easily the arm can accelerate, not just move at constant speed.
• Units problem: for a 6-DOF arm the Jacobian mixes m/s (linear) and rad/s (angular), so the ellipsoid isn't dimensionally homogeneous. Fixes include the characteristic length normalization and the homogeneous power manipulability index.

Unlike forward kinematics, which only tells you where the tip is, the ellipsoid tells you how well the tip can move from there — a first-order, velocity-level property.

Failure modes, trade-offs, and significance

Kinematic singularities are the dominant failure mode the ellipsoid predicts. Classic 6-DOF anthropomorphic arms like the PUMA 560 exhibit three families:

• Shoulder singularity: wrist center directly above the base axis — the arm can't move the tip in one horizontal direction.
• Elbow singularity: arm fully extended; radial motion is lost (the 2R example above).
• Wrist singularity: two wrist axes become collinear (parallel), losing one rotational DOF — the most common in practice and why welding/spray paths jerk violently when a tool orientation lines up the wrist.

Near any of these, the inverse J⁻¹ or J⁺ demands unbounded joint rates: a commanded 10 cm/s tip velocity can require 5,000 deg/s at a joint — impossible, so the robot either faults, stalls, or overshoots. The deep significance is that singularities are an intrinsic property of the mechanism's geometry, not a control bug: no amount of clever software removes them from a 6R arm's workspace. The manipulability ellipsoid is the engineer's map of where those cliffs lie and how much dexterity remains before the arm falls off one.

Jacobian-based performance indices for robot manipulators

Index	Definition	Range / ideal value	What it tells you
Yoshikawa manipulability w	√det(JJᵀ) = Πσᵢ	0 → large; 0 at singularity	Ellipsoid volume; overall distance-from-singularity
Minimum singular value σ_min	smallest σ of J	0 → large; 0 at singularity	Worst-case velocity gain; tightest bottleneck
Condition number κ	σ_max / σ_min	1 (isotropic) → ∞ (singular)	Ellipsoid eccentricity; numerical ill-conditioning
Isotropy index (1/κ)	σ_min / σ_max	1 (best) → 0 (singular)	How sphere-like the ellipsoid is
Dynamic manipulability	√det(J M⁻¹ M⁻ᵀ Jᵀ)	0 → large	Acceleration capability accounting for inertia M