Reactionless camera inspection with a free-flying space robot under reaction null-space motion control
Authors: Sone, Nenchev · Year: 2016 · Venue: Acta Astronautica
Raw: md
Summary
The paper applies Reaction Null-Space (RNS) motion control to an orientation-prevailing end-effector task — hand-held camera inspection — performed by a seven-DoF redundant manipulator on a satellite base. The authors argue that reactionless position tracking is severely limited (effectively one positioning DoF survives), so they redirect RNS toward wrist-orientation tasks, which fit the reactionless manifold far better. They contribute a structure-specific decomposition of reactionless motion into “predominant wrist motion” plus “elbow folding/unfolding,” a nonlinear-systems (fixed-point/bifurcation) analysis of dynamic singularities versus manipulator attachment offset, a prioritized null-space control law with two algorithmic-singularity treatments (DLS and Singularity Consistent), and a kinetic-energy argument that reactionless motion is essentially the instantaneous minimum-energy motion under zero base-attitude deviation.
Key Claims
- For a 7-DoF arm the reactionless motion set has 4 DoF (= 7 joints − 3 base-attitude DoF), decomposable into a 3-DoF predominant wrist motion pattern plus a 1-DoF elbow folding/unfolding pattern (Joints 2/3, with Joint 3 generating far smaller reaction than Joint 2).
- Under reactionless motion the effective positioning DoF of the end effector is essentially one, set mainly by the elbow folding/unfolding pattern — hence reactionless positioning is impractical and orientation tasks are the better candidate.
- The manipulator attachment offset
r(distance from base CoM) acts as a bifurcation parameter: at smallr(e.g. 0.5 m) the reactionless vector field has no fixed points; nearr ≈ 0.945m bifurcations appear; for largerr(1.5 m) two fixed points emerge and separate; asr → ∞they converge to the fixed-base kinematic singularities (θ₂ → 0 and θ₂ → ±π). - Fixed points of the reactionless vector field coincide with rank deficiency of the coupling inertia matrix (
det(M̃_ωm M̃_ωmᵀ) = 0), i.e. dynamic singularities;det(M̃_ωm M̃_ωmᵀ)is largest when the manipulator/base CoM distance is largest (extended-arm θ = 0). - Three singularity types are distinguished: kinematic (
det(J_ω J_ωᵀ)=0), dynamic (coupling-inertia rank loss), and algorithmic (det(J̄_ω J̄_ωᵀ)=0with full-rankJ_ωandM̃_ωm). Dynamic singularities do not corrupt the control law because only the RNS projector appears; algorithmic singularities (Joints 5/7 counter-rotating at high rate, akin to wrist/Euler-angle singularities) must be treated. - A modified DLS inverse (damping only on the minimum-singular-value term, with a smoothed damping schedule) suppresses joint-rate blowup but introduces direction error; the Singularity Consistent method preserves the commanded velocity direction and errs only in path speed, using “natural motion” with constant field magnitude
b. - Reactionless motion is almost identical to the instantaneous minimum kinetic-energy motion under zero base-attitude deviation: for the planar 2-DoF model the cost ratio
C_ratio = T_RNS/T_minaverages 1.002 over a 10 k-point joint-space mesh. - Simulations: under RNS the base attitude deviation is negligible during wrist inspection, while a conventional inverse-Jacobian + reaction-wheel controller incurs relatively large base attitude deviation despite using only the low-inertia wrist subchain.
Method
FFSR model = satellite base + serial n-DoF arm, optionally with 3 orthogonal reaction wheels. Momentum (zero initial) gives, after eliminating base linear velocity v_b, the angular-momentum conservation law (Eq. 2):
M̃_ω ω_b + M̃_ωm θ̇ + M̃_ωr φ̇ = 0,
with tilde quantities formed by the Schur-complement reduction M̃_ω = M_ω − M_vωᵀ M_v⁻¹ M_vω (analogously for M̃_ωm, M̃_ωr). The middle term is the coupling angular momentum; M̃_ωm is the coupling inertia matrix.
Regime note. The system is treated as free-floating (φ̇ = 0, base attitude uncontrolled, momentum conserved) until Section 6. Reactionless control deliberately avoids base actuation: it commands joint motion that keeps ω_b = 0 passively via M̃_ωm θ̇ = 0 (Eq. 3). The contrast case in Sections 4 and 6 is a free-flying-style approach where the ACS (reaction wheels) actively cancels the base disturbance — and the paper’s central message is that RNS removes the need to use that actuation. So although the title says “free-flying,” the manipulator regime under RNS is free-floating with passive base-attitude conservation; the base actuation only appears as the baseline being beaten.
Reactionless set (Eq. 4): θ̇ = P_RNS θ̇_a, with P_RNS = E − M̃_ωm⁺ M̃_ωm the projector onto the null-space (the Reaction Null-Space) of the coupling inertia matrix; θ̇_a is an arbitrary joint-velocity parameter. The authors note this captures the entire reactionless set, so RNS motions coincide with those from nonlinear-optimization methods.
Control law (Eq. 6), prioritized via consecutive null-space projection:
θ̇_ref = J̄_ω⁺ ω_e^ref + k_g P (J_w,vᵀ) Δp_w,
where J̄_ω = J_ω P_RNS is the restricted Jacobian for the wrist-orientation subtask (priority 1 = reactionless constraint, priority 2 = wrist orientation, priority 3 = wrist position stabilization), P = P_RNS (E − J̄_ω⁺ J̄_ω) is the combined null-space projector, Δp_w = p_w − p_w^init is wrist deflection, and the priority-3 term descends the potential V = ½‖Δp_w‖² (Eq. 7).
Two-DoF planar reduction (Eq. 5): θ̇ = b·n(θ), with n(θ) the lone null-space vector of the coupling inertia matrix and b a scalar field magnitude. Treated as an autonomous nonlinear system; fixed points ⇔ coupling-inertia rank loss. Bifurcation analysis sweeps attachment offset r.
Algorithmic-singularity treatments: modified DLS inverse (Eqs. 8–10) damps only the σ₃ term, with λ² switched on inside a singular region σ₃ ≤ ε via a smoothed schedule (the σ₃⁴/ε⁴ term added by the authors for C¹ continuity at the border); and the Singularity Consistent inverse (Eqs. 11–13) using cofactor-like weights μ_i = ∏_{j≠i} σ_j.
Kinetic-energy analysis (Sec. 6): from the EoM (Eq. 14) and zero base-attitude deviation, kinetic energy reduces to T = ½ θ̇ᵀ M̂ θ̇ (Eq. 18) with M̂ = M̃_m + I_r⁻¹ M̃_ωmᵀ M̃_ωm; the minimum-energy direction is the smallest-singular-value direction of M̂. Comparison via C_ratio = T_RNS/T_min (Eq. 19).
Relevance to thesis
This is a direct, high-value reference for the nominal guidance/control layer of a free-flying manipulator. It (1) sharply quantifies what RNS can and cannot do — orientation yes, positioning essentially no — which constrains any planning architecture that hopes to track end-effector pose reactionlessly; (2) exposes attachment-offset geometry as a design lever that creates or destroys dynamic singularities, relevant to both hardware layout and reachable-set/risk analysis; (3) gives a clean taxonomy (kinematic / dynamic / algorithmic) and concrete singularity-robustification recipes (DLS, SC) usable in our controller; and (4) supplies the energy-equivalence argument that motivates RNS as not merely disturbance-minimizing but near-energy-optimal — a useful nominal-optimality baseline before layering uncertainty. For our fully-actuated base, the conventional ACS-compensation case is the regime we actually live in, so this paper frames precisely the trade we must reason about: spend base actuation versus exploit redundancy.
Connections
Topics: reaction_null_space, coupling_inertia_matrix, dynamic_singularity, kinematic_redundancy
Key Equations / Quotes
“the positioning DoF of the end effector can be assumed to be just one. This DoF is determined mainly by the elbow folding/unfolding pattern. This is a severe limitation with regard to general positioning subtasks. This is the reason why we decided to focus on orientation-prevailing tasks here.” (Sec. 3.1)
“Note that until Section 6, the system is regarded as free-floating (φ̇ = 0).” (Sec. 2.1)
Reaction Null-Space projector (Eq. 4):
θ̇ = P_RNS θ̇_a,P_RNS = E − M̃_ωm⁺ M̃_ωm
Prioritized reactionless wrist-control law (Eq. 6):
θ̇_ref = J̄_ω⁺ ω_e^ref + k_g P (J_w,vᵀ) Δp_w, withJ̄_ω = J_ω P_RNS
Dynamic singularity condition (Sec. 5):
det(M̃_ωm M̃_ωmᵀ) = 0
Energy-equivalence result (Sec. 6.2.1):
“the average of C_ratio is 1.002 among all points. Hence, reactionless and instantaneous minimum energy motion are equivalent for this model.”
Open Questions
- The bifurcation/fixed-point analysis is carried out on a planar 2-DoF reduction of the elbow pattern; how do the fixed-point structure and attachment-offset bifurcations generalize to the full spatial 7-DoF case where joint-axis directions matter (the authors flag this but do not resolve it)?
- Algorithmic singularities are said to “occur occasionally only” — there is no characterization of where in configuration/task space they live, nor a guarantee that DLS/SC suffice in general.
- The energy equivalence is shown for one-DoF-reactionless models; the paper defers the 4-DoF spatial case — does
C_ratio ≈ 1survive higher reactionless dimensionality? - DLS parameters (
ε,λ_max) and the SC scalarb = 1.7are set empirically; no principled tuning rule is given.