Inverse Jacobians and Singularity Avoidance for Redundant Free-Floating Space Manipulators

A methods survey synthesized from the deep-research run of 2026-06-09

Provenance and health warning. This report was assembled from the salvaged journal of a deep-research workflow (run wf_7c0f5633-e15) that fetched 33 web sources, extracted 131 falsifiable claims, and adversarially verified 57 of them (0 refuted) before the run was halted prior to its own synthesis step. Every citation here is web-derived, not grounded in sources/Markdown_Papers. Per the workspace rules, treat the whole document as a research lead: anything destined for the thesis must be re-grounded against the local corpus and given a page locator first. Source URLs were not preserved in the journal; sources are identified below by the venue/author/date the fetch agents recorded, and are marked [UNVERIFIED-SRC] where even that was missing. Scope is the methods only — no inspection-planning framing, as requested.

1. The inverse map: why a generalized Jacobian is needed

For a fixed-base arm the differential kinematics ẋ = J(q)q̇ invert by the Moore-Penrose pseudoinverse q̇ = J⁺ẋ, with redundancy (n > m) resolved by adding a null-space term [S2]. A free-floating base breaks this directly: with attitude and position control switched off, any arm motion induces reaction forces and moments that move the base, so the end-effector velocity depends on base motion as well as joint rates [S15, S24]. Umetani & Yoshida (1989) resolved this by folding the momentum-conservation law into the kinematic formulation, producing the Generalized Jacobian Matrix (GJM) — a Jacobian in generalized form specific to free-flying multi-link systems — and built a resolved-motion-rate control law on top of it, applicable to both endpoint manipulation and base-attitude control [S15, S24].

The modern explicit form, for zero base momentum, is

J*_m = J_m − J_b M_b⁻¹ M_bm,

i.e. the manipulator Jacobian minus a base/coupling-inertia correction term [S14, S30]. The control law inverts this operator: q_des(t) = ∫ J*_m⁻¹ ν_e,des dt, which is well-defined only while J*_m stays invertible [S30]. Source [S13] generalizes the construction to constant non-zero conserved momentum and formulates the kinematics and dynamics directly on SE(3) to avoid pose-parametrization singularities, treating the GJM as the system’s input-output decoupling matrix.

A key structural result for singularity analysis comes from the inverse-chain (inverted-links) factorization [S8, S30]: J*_m = −M_ee⁻¹ M_em, so det(J*_m) = −det(M_ee⁻¹)·det(M_em). Because M_ee is an inertia and always invertible, all dynamic singularities arise solely from degeneracy of the coupling matrix M_em, which also proves they are independent of the last link’s inertia and last joint coordinate — reducing the singularity search space by one dimension.

2. Two kinds of singularity

Kinematic singularities are the familiar geometric, configuration-only rank losses of the (fixed-base) Jacobian: rank(J) < m, with det(J) = 0 when m = n, at which the condition number κ(J) = σ_max/σ_min diverges and the Yoshikawa manipulability w(q) = √det(JJᵀ) vanishes [S2, S5].

Dynamic singularities are unique to free-floating systems and are the central phenomenon of this literature. Papadopoulos & Dubowsky (1993) introduced them as configurations where the end-effector cannot be moved in some inertial direction even though the arm is not kinematically singular, because the uncompensated base reaction collapses the achievable end-effector velocities onto a single direction [S10, S19, S26, S31]. Their defining properties, corroborated across many sources:

• They correspond to loss of invertibility of the GJM, not the fixed-base Jacobian [S10, S6, S32]; J*_m must stay full-rank throughout the motion.
• Their existence and location depend on the system’s inertial/mass and momentum properties, so they cannot be predicted from kinematics alone [S10, S12, S19, S31].
• They are path-dependent: a given Cartesian pose may or may not be singular depending on the path taken to reach it [S6, S7, S26].
• Jacobian-inverse controllers (resolved-rate, resolved-acceleration) fail at them, and plain pseudoinverse solutions produce large errors there [S12].

This motivates the Path-Independent Workspace (PIW) vs. Path-Dependent Workspace (PDW) partition [S10, S21, S32]: dynamic singularities can only occur in the PDW; the PIW is dynamic-singularity-free. Under non-zero angular momentum the workspace usable for sustained tasks shrinks to the PIW [S21].

A sharp caveat from the RSS 2020 “Singularity Maps” work [S8, S30]: for spatial free-floating robots the PIW/PDW distinction loses effectiveness because nearly the entire workspace becomes path-dependent, and singularities are not meaningfully representable in Cartesian space (projection makes almost everything look singular). Because J*_m depends solely on q, however, the singularity loci are fixed in configuration space — so the right move is to characterize them there, not in the workspace.

3. Singularity-avoidance and handling methods

The literature splits cleanly into reactive numerical handling (condition the inverse at/near a singularity) and predictive planning (don’t go there).

3.1 Singularity-robust inversion (reactive)

The damped least squares (DLS) inverse J⁺ = Jᵀ(JJᵀ + λ²I)⁻¹ regularizes the pseudoinverse, minimizing ‖ẋ − Jq̇‖² + λ²‖q̇‖² and bounding joint rates near singularities at the cost of task-space error; λ = 0 recovers Moore-Penrose [S2, S5]. Adding λ²I drives the condition number of (JJᵀ + λ²I) toward 1, guaranteeing a well-conditioned inverse even where the smallest singular value is zero [S5]. Fixed damping is suboptimal, so several variable schedules appear: damping proportional to squared task error (Nakamura & Hanafusa) so it vanishes away from singularities [S5]; and, for the free-floating case specifically, damping varied as a function of det(GJM) to preserve continuity of the computed joint velocities [S3]. Sources [S4, S9] handle GJM dynamic singularities with a singular-value-filtering (SVF) decomposition that conditions the small/zero singular values, and [S13] derives a DLS-based singularity-robust inverse that prevents excessive joint torques near singular configurations.

3.2 Configuration-space singularity maps (predictive, global)

Manipulability maximization m(q) = |det(J*_m(q))| at each via-point is the classical local avoidance heuristic, but it has a documented failure mode: it can miss an infeasible step where the determinant switches sign across a singularity, because it lacks global knowledge of singularity locations [S30]. The S-Map [S8, S30] fixes this by precomputing a signed distance to the nearest singularity set in configuration space; this turns singularity avoidance into a collision-avoidance-style inequality constraint usable in gradient-based trajectory planning, with global knowledge the manipulability measure lacks.

3.3 Initial-configuration selection (predictive, offline)

Because dynamic singularities are path-dependent and momentum-dependent, Papadopoulos and co-workers show they can be avoided before motion by choosing the initial spacecraft attitude and manipulator configuration so the prescribed trajectory never crosses a singular surface [S6, S26, S32]. The admissible set of initial configurations is itself a function of the accumulated angular momentum; for systems with momentum, keeping both initial and final end-effector locations inside the PIW suffices [S6, S32, S10]. Nanos & Papadopoulos (2015) extend this to planar and spatial systems, with or without initial angular momentum, for arbitrary position-and-attitude trajectories [S6, S32].

3.4 Forward-kinematics-only planning (avoids inversion entirely)

A distinct strategy sidesteps the inverse map altogether [S7, S33]: parameterize joint trajectories by polynomial/sinusoidal functions, forward-integrate the differential kinematics, and cast Cartesian path planning as a parameter search (solved with Particle Swarm Optimization). Because the Jacobian is never inverted, both kinematic and dynamic singularities are avoided by construction, and quaternion attitude representation removes representation singularities [S7, S33].

3.5 Reaction Null-Space family (redundancy used to decouple base and arm)

For kinematically redundant systems (n > 6) the Reaction Null-Space (RNS) is the kernel of the coupling-inertia matrix M_bm: joint velocities drawn from it are reactionless, imposing no momentum on the base, so arm motion is completely decoupled from base motion [S22]. Redundancy resolution then yields a generalized-inverse term (mapping reaction momentum) plus an RNS-projected term of rank n−6 [S22]. The same structure appears as a reaction-torque null-space, shown mathematically equivalent to the RNS, usable as an optimization index inside GJM-based Cartesian path tracking to minimize base disturbance [S16, S28]. Task-priority RNS schemes assign base-attitude adjustment as the primary task and the end-effector task to the null space, exploiting coupling to save propellant, and remove an algorithm singularity present in the conventional RNS [S4, S9]. The Zero-Reaction Maneuver (ZRM) is the RNS applied to lift the base-reaction velocity limit; crucially, for a non-redundant 6-DOF arm a ZRM barely exists — kinematic redundancy is required for useful reactionless freedom — and it was demonstrated on orbit on ETS-VII [S22, S29]. The RNS also underlies dynamically decoupled control of flexible-base manipulators [S18].

3.6 Base-attitude restoration / bidirectional planning

For dual-arm free-floating robots, an “enhanced bidirectional” planner solves the nonholonomic redundancy-resolution problem to eliminate the net base-attitude change over a maneuver, using an acceleration-level state equation and a Lyapunov-based controller for smooth, restorable joint trajectories [S11].

4. Synthesis of the method landscape (my inference, not a single source)

The literature is organized around one fact — the operative operator for a free-floating arm is the GJM, not the geometric Jacobian — and then forks on how to keep that operator invertible. Reactive methods (DLS, SVF, variable det-based damping) keep control numerically stable when a singularity is unavoidable but pay task-error/torque costs. Predictive methods (initial-config selection, configuration-space S-Maps, forward-kinematics optimization) try to never encounter one, and the field’s trajectory runs from local manipulability indices toward globally-aware configuration-space maps, explicitly because the local index can silently cross a sign change. Redundancy is the enabler that ties the two together: it is what makes RNS/ZRM decoupling, task-priority base control, and null-space-projected conditioning possible at all — a non-redundant 6-DOF free-floater has almost no reactionless freedom [S29]. The one genuinely inversion-free branch [S7, S33] is the cleanest way to dodge both singularity classes but trades the closed-form inverse for a global optimization.

Source key (web-derived; re-ground before citing)

Adversarial verification covered 57 of the 131 extracted claims before the run was halted; all 57 survived (0 refuted). Unverified-but-extracted claims are still only as good as their web source — re-ground against sources/Markdown_Papers before use.