betaFeedback

WorksComputer Science

Feedback-Linearised Trajectory Tracking via Model-Informed Robust Geometric DDPC

Abstract

Data-driven predictive control (DDPC) predicts system responses from recorded trajectories, but nonlinear dynamics can limit the accuracy of fixed linear predictors. This paper studies geometric robust DDPC for trajectory tracking of a fully actuated manipulator using model-based feedback linearisation. A mechanical model maps virtual acceleration commands into joint torques, while recorded data capture residual dynamics arising from imperfect model parameters. The controller minimises worst-case prediction error over an open Grassmannian ball centred at the fitted trajectory subspace. Simulations on a planar 2R manipulator under gravity compare the proposed method with double-integrator model predictive control (MPC) and direct regularised data-enabled predictive control (DeePC). Results show improved tracking over double-integrator MPC despite increased torque effort, while performance relative to DeePC depends on the reference trajectory.

Introduction

Predictive control chooses future inputs by predicting the system’s response, applies the first input, and repeats the optimisation after each new measurement. Model predictive control (MPC) predicts with a mathematical model. Data-driven predictive control (DDPC) predicts from recorded data instead, as in data-enabled predictive control (DeePC), which represents trajectories as linear combinations of recorded input–output data [4]. The Fundamental Lemma justifies this representation for controllable linear time-invariant (LTI) systems with noiseless, sufficiently exciting data [16].

Nonlinear systems with uncertain parameters challenge both: inaccurate mechanical models yield biased predictions, while linearly combining measured trajectories approximates nonlinear dynamics only locally. Model knowledge reduces the data burden. For fully actuated manipulators, feedback linearisation maps uses a mechanical model to map virtual acceleration commands to joint torques. With exact parameters and states, each joint becomes a double integrator from virtual input to position. Parameter, state-estimation and actuator errors produce residual dynamics learnable from virtual-input and position recordings.

Geometric robust DDPC [3] fits a low-dimensional trajectory subspace and handles fitting errors using an open chordal-distance ball of same-dimensional subspaces on the Grassmannian. It minimises worst-case weighted squared deviation from a target combining measured history and desired future trajectory. The mechanical model supplies the feedback-linearising torque map; DDPC predicts the virtual-input-to-position response. We test whether learning this response improves prediction and tracking beyond the known double-integrator model.

Related work. Regularised and robust formulations extend DeePC to noisy data [5, 9]. For nonlinear systems, Berberich et al. [2] obtain closed-loop stability guarantees by updating the data of a linear predictor online. For feedback-linearizable systems, Alsalti et al. [1] build a data-based predictor from input–output data and basis functions that approximate the unknown nonlinearities, with a fully actuated double inverted pendulum as example. Watson [13] incorporates partial model knowledge into DeePC by combining model-based and data-based representations. Tube MPC handles bounded disturbances around a model prediction [12], and learning-based MPC augments a model with learned corrections and uncertainty estimates [8 a].

Contributions. We apply geometric robust DDPC to trajectory tracking of a planar 2R manipulator under gravity. Prediction tests compare the learned feedback-linearised response with the ideal double integrator, for exact and inaccurate mechanical parameters. Transfer tests examine whether fixed parameter tunings remain effective when the measurement noise and arm parameters change. Tracking tests compare accuracy and torque effort with double-integrator MPC and with direct regularised DeePC, which predict from recorded joint torques and positions directly.

Preliminaries

Notation. Operator col⁡(⋅)\operatorname{col}(\cdot) stacks its arguments vertically. For Q⪰0Q \succeq0 we write ∥z∥Q2=zTQz\lVert z\rVert_{Q}^{2}=z^{\mathsf{T}}Qz. We use ∥⋅∥F\lVert\cdot\rVert_{\mathrm{F}} for the Frobenius norm, A†A^{\dagger} for the Moore–Penrose pseudoinverse, and im⁡A\operatorname{im} A for the column space. IrI_r and 1r\mathbf{1}_r are the identity matrix and the all-ones vector, and ⊗\otimes is the Kronecker product. For zj∈Rnzz_j \in\mathbb{R}^{n_z}, we write za:b=col⁡(za,…,zb)z_{a:b}=\operatorname{col}(z_a,\ldots,z_b). Superscripts d\mathrm{d}, ini\mathrm{ini} and ref\mathrm{ref} mark recorded data, initial histories and references. The measured output is y~k=yk+εky\widetilde{y}_k=y_k+\varepsilon_k^y, where εky\varepsilon_k^y is measurement noise. A record zd=z0:Td−1dz^{\mathrm{d}}=z_{0:T_{\mathrm{d}}-1}^{\mathrm{d}} of length TdT_{\mathrm{d}} defines, for 1≤s≤Td1\leq s\leq T_{\mathrm{d}}, the depth-ss block Hankel matrix

Hs(zd)=[z0dz1d⋯zTd−sdz1dz2d⋯zTd−s+1d⋮⋮⋱⋮zs−1dzsd⋯zTd−1d]∈Rnzs×(Td−s+1).(1)\mathcal{H}_s(z^{\mathrm{d}})= \begin{bmatrix} z_0^{\mathrm{d}} & z_1^{\mathrm{d}} & \cdots& z_{T_{\mathrm{d}}-s}^{\mathrm{d}}\\ z_1^{\mathrm{d}} & z_2^{\mathrm{d}} & \cdots& z_{T_{\mathrm{d}}-s+1}^{\mathrm{d}}\\ \vdots& \vdots& \ddots& \vdots\\ z_{s-1}^{\mathrm{d}} & z_s^{\mathrm{d}} & \cdots& z_{T_{\mathrm{d}}-1}^{\mathrm{d}} \end{bmatrix}\in\mathbb{R}^{n_zs\times(T_{\mathrm{d}}-s+1)}. \tag*{(1)}

Each column is a window of ss consecutive samples. The record is persistently exciting of order ss if this matrix has full row rank, that is, rank⁡Hs(zd)=nzs\operatorname{rank}\mathcal{H}_s(z^{\mathrm{d}})=n_zs.

Linear Systems and Model Predictive Control

Consider a controllable and observable discrete-time LTI system xk+1=Axk+Bukx_{k+1}=Ax_k+Bu_k, yk=Cxk+Duky_k=Cx_k+D u_k, with xk∈Rnxx_k\in\mathbb{R}^{n_x}, uk∈Rnuu_k\in\mathbb{R}^{n_u} and yk∈Rnyy_k\in\mathbb{R}^{n_y}. Its lag ℓ\ell is the smallest positive integer for which the observability matrix Oℓ=col⁡(C,CA,…,CAℓ−1)\mathcal{O}_{\ell}=\operatorname{col}(C,CA,\ldots,CA^{\ell-1}) has rank nxn_x.

MPC predicts over a horizon of NN samples. At step kk, let uk+i∣ku_{k+i|k} and yk+i∣ky_{k+i|k} denote the (k+i)(k+i)th prediction, stacked as uku_k and yky_k over i=0,…,N−1i=0,\ldots,N-1. With desired sequences ujref,yjrefu_j^{\mathrm{ref}},y_j^{\mathrm{ref}}, the input-output tracking cost is written as

Jk(uk,yk)=∑i=0N−1(∥yk+i∣k−yk+iref∥Q2+∥uk+i∣k−uk+iref∥R2).(2)J_k(u_k,y_k)=\sum_{i=0}^{N-1}\left(\left\lVert y_{k+i|k}-y_{k+i}^{\mathrm{ref}}\right\rVert_Q^2+\left\lVert u_{k+i|k}-u_{k+i}^{\mathrm{ref}}\right\rVert_R^2\right). \tag*{(2)}

where Q⪰0Q\succeq0 and R≻0R\succ0. MPC minimises this cost subject to the dynamics and an initial state estimate, then applies the first input.

Data-Based Prediction

For wj=col⁡(uj,yj)w_j = \operatorname{col}(u_j, y_j), the length-LL behaviour BL\mathcal{B}_L is the set of trajectories w=col⁡(w0,…,wL−1)\mathbf{w} = \operatorname{col}(w_0, \ldots, w_{L-1}) that the system generates from arbitrary initial states and inputs. For L≥ℓL \ge\ell, it is a subspace of dimension nuL+nxn_uL + n_x ([10]). Let wd\mathbf{w}^{\mathrm{d}} be a noiseless trajectory of the system with persistently exciting input of order L+nxL + n_x. Willems et al. [16]’s Fundamental Lemma states:

BL=im⁡HL(wd).(3)\mathcal{B}_L = \operatorname{im} \mathcal{H}_L(\mathbf{w}^{\mathrm{d}}). \tag*{(3)}

For prediction, choose a history length Tini≥ℓT_{\mathrm{ini}} \ge\ell and a prediction horizon NN, and set L=Tini+NL = T_{\mathrm{ini}} + N. In every column of a Hankel matrix, the first TiniT_{\mathrm{ini}} samples form the past (superscript p) and the remaining NN samples form the future (superscript f). The input and output Hankel matrices are therefore Hu=HL(ud)=col⁡(Up,Uf)\mathcal{H}_u = \mathcal{H}_L(\mathbf{u}^{\mathrm{d}}) = \operatorname{col}(U^{\mathrm{p}}, U^{\mathrm{f}}) and Hy=HL(yd)=col⁡(Yp,Yf)\mathcal{H}_y = \mathcal{H}_L(\mathbf{y}^{\mathrm{d}}) = \operatorname{col}(Y^{\mathrm{p}}, Y^{\mathrm{f}}). With measured outputs, the output matrix becomes H~y=HL(y~d)=col⁡(Y~p,Y~f)\widetilde{\mathcal{H}}_y = \mathcal{H}_L(\widetilde{\mathbf{y}}^{\mathrm{d}}) = \operatorname{col}(\widetilde{Y}^{\mathrm{p}}, \widetilde{Y}^{\mathrm{f}}). All blocks have ngn_g columns, with ng=Td−L+1n_g = T_d - L + 1 for a single record. The blocks UpU^{\mathrm{p}} and UfU^{\mathrm{f}} have nuTinin_uT_{\mathrm{ini}} and nuNn_uN rows, and the output blocks have nyTinin_yT_{\mathrm{ini}} and nyNn_yN rows.

Define the histories ukini=uk−Tini:k−1\mathbf{u}^{\mathrm{ini}}_k = u_{k-T_{\mathrm{ini}}:k-1} and ykini=yk−Tini:k−1\mathbf{y}^{\mathrm{ini}}_k = y_{k-T_{\mathrm{ini}}:k-1}, and let y~kini\widetilde{\mathbf{y}}^{\mathrm{ini}}_k denote the measured output history. Under the noiseless spanning conditions, (uk,yk)(u_k, y_k) is a valid continuation of this history if and only if some gk∈Rngg_k \in\mathbb{R}^{n_g} satisfies

col⁡(Up,Yp,Uf,Yf)gk=col⁡(ukini,ykini,uk,yk).(4)\operatorname{col}(U^{\mathrm{p}}, Y^{\mathrm{p}}, U^{\mathrm{f}}, Y^{\mathrm{f}})g_k = \operatorname{col}(\mathbf{u}^{\mathrm{ini}}_k, \mathbf{y}^{\mathrm{ini}}_k, u_k, y_k). \tag*{(4)}

Since Tini≥ℓT_{\mathrm{ini}} \ge\ell, the history determines the initial state, and the prediction is equivalent to model-based prediction with the same cost and constraints ([4]).

Regularised DeePC

Measurement noise perturbs both the recorded matrix and the online history, so (4) cannot be satisfied exactly. Regularised DeePC therefore allows a residual σky\sigma_k^y in the past-output equation and penalises the coefficient vector gkg_k ([5]). Penalising ∥gk∥2\lVert g_k\rVert_2 pulls the prediction towards the zero trajectory, which is unsuitable for a nonzero reference. Following Elokda et al. [6], we centre the penalty at a coefficient vector that reproduces the reference. Define

H~d=col⁡(Up,Y~p,Uf,Y~f),gkref=(H~d)†rk,\widetilde{\mathcal{H}}^{\mathrm{d}} = \operatorname{col}(U^{\mathrm{p}}, \widetilde{Y}^{\mathrm{p}}, U^{\mathrm{f}}, \widetilde{Y}^{\mathrm{f}}), \qquad g_k^{\mathrm{ref}} = \left(\widetilde{\mathcal{H}}^{\mathrm{d}}\right)^\dagger r_k,
rk=col⁡(uk−Tini:k−1ref,yk−Tini:k−1ref,uk:k+N−1ref,yk:k+N−1ref).(5)r_k = \operatorname{col}(u^{\mathrm{ref}}_{k-T_{\mathrm{ini}}:k-1}, y^{\mathrm{ref}}_{k-T_{\mathrm{ini}}:k-1}, u^{\mathrm{ref}}_{k:k+N-1}, y^{\mathrm{ref}}_{k:k+N-1}). \tag*{(5)}

The minimum-norm least-squares solution gkrefg_k^{\mathrm{ref}} is recomputed when the reference or the data change. It reproduces rkr_k exactly if rk∈im⁡H~dr_k \in\operatorname{im}\widetilde{\mathcal{H}}^{\mathrm{d}}, which does not imply physical feasibility.

The regularised problem is

min⁡gk, σkyJk(Ufgk,Y~fgk)+λy∥σky∥22+λg∥gk−gkref∥22\min_{g_k,\,\sigma_k^y} \quad J_k(U^{\mathrm{f}}g_k,\widetilde{Y}^{\mathrm{f}}g_k) + \lambda_y\lVert\sigma_k^y\rVert_2^2 + \lambda_g\lVert g_k-g_k^{\mathrm{ref}}\rVert_2^2
subject toUpgk=ukini,Y~pgk=y~kini+σky,(6)\text{subject to} \quad U^{\mathrm{p}}g_k = \mathbf{u}^{\mathrm{ini}}_k,\qquad\widetilde{Y}^{\mathrm{p}}g_k = \widetilde{\mathbf{y}}^{\mathrm{ini}}_k + \sigma_k^y, \tag*{(6)}

where λy,λg>0\lambda_y,\lambda_g > 0. Known inputs ukini\mathbf{u}^{\mathrm{ini}}_k are matched exactly, and the penalised past-output residual is σky=Y~pgk−y~kini\sigma_k^y = \widetilde{Y}^{\mathrm{p}}g_k - \widetilde{\mathbf{y}}^{\mathrm{ini}}_k. The future predictions are uk=Ufgku_k = U^{\mathrm{f}}g_k and yk=Y~fgky_k = \widetilde{Y}^{\mathrm{f}}g_k.

Geometric Robust DDPC

Geometric robust DDPC [3] accounts for uncertainty in the subspace that represents the trajectories. Let wj=col⁡(uj,y~j)w_j = \operatorname{col}(u_j,\tilde{y}_j) and namb=(nu+ny)Ln_{\mathrm{amb}} = (n_u+n_y)L. The leading d<nambd < n_{\mathrm{amb}} left singular vectors of HL(w~d)\mathcal{H}_L(\tilde{w}^{d}) form an orthonormal basis V^\hat{V} of the fitted subspace S^=im⁡V^\hat{\mathcal{S}} = \operatorname{im}\hat{V}. The dimension dd is motivated by the LTI behaviour dimension nuL+nxn_uL+n_x. Nonlinear effects and measurement noise raise the rank of the data matrix, S^\hat{\mathcal{S}} becomes a fixed-rank approximation. The data may be centred and scaled before fitting (Appendix § 9.§ 1).

The Grassmannian Gr⁡(d,namb)\operatorname{Gr}(d,n_{\mathrm{amb}}) is the set of dd-dimensional subspaces of Rnamb\mathbb{R}^{n_{\mathrm{amb}}}. For an orthonormal basis VV of S\mathcal{S}, the projector is PS=VVTP_{\mathcal{S}} = VV^{\mathsf{T}}. The principal angles θj∈[0,π/2]\theta_j \in[0,\pi/2] between two subspaces satisfy cos⁡θj=σj(V1TV2)\cos\theta_j = \sigma_j(V_1^{\mathsf{T}}V_2), where σj\sigma_j is the jjth singular value. The chordal distance and the open ball of radius ρ\rho around S^\hat{\mathcal{S}} are

dch(S1,S2)=∥PS1−PS2∥F2=(∑j=1dsin⁡2θj)1/2,(7)d_{\mathrm{ch}}(\mathcal{S}_1,\mathcal{S}_2) = \frac{\lVert P_{\mathcal{S}_1}-P_{\mathcal{S}_2}\rVert_{\mathrm{F}}}{\sqrt{2}} = \left(\sum_{j=1}^{d}\sin^2\theta_j\right)^{1/2}, \tag*{(7)}
Bρ(S^)={S∈Gr⁡(d,namb):dch(S,S^)<ρ},0<ρ<1.\mathbb{B}_{\rho}(\hat{\mathcal{S}}) = \{\mathcal{S}\in\operatorname{Gr}(d,n_{\mathrm{amb}}):d_{\mathrm{ch}}(\mathcal{S},\hat{\mathcal{S}})<\rho\},\quad0<\rho<1.

At update kk, the target bk∈Rnambb_k\in\mathbb{R}^{n_{\mathrm{amb}}} collects the measured past pairs col⁡(uj,y~j)\operatorname{col}(u_j,\tilde{y}_j) for j=k−Tini,…,k−1j=k-T_{\mathrm{ini}},\ldots,k-1 and the future reference pairs col⁡(ujref,yjref)\operatorname{col}(u_j^{\mathrm{ref}},y_j^{\mathrm{ref}}) for j=k,…,k+N−1j=k,\ldots,k+N-1. For a decision variable ξ∈Rnamb\xi\in\mathbb{R}^{n_{\mathrm{amb}}}, the projection PSξP_{\mathcal{S}}\xi onto a candidate subspace S\mathcal{S} is the corresponding predicted trajectory, with past and future inputs and outputs. The predictive problem is

min⁡ξ∈Rnamb  sup⁡S∈Bρ(S^)Fk(ξ,S),Fk(ξ,S)=∥PSξ−bk∥W2,(8)\min_{\xi\in\mathbb{R}^{n_{\mathrm{amb}}}\ }\ \sup_{\mathcal{S}\in\mathbb{B}_{\rho}(\hat{\mathcal{S}})} F_k(\xi,\mathcal{S}),\qquad F_k(\xi,\mathcal{S})=\lVert P_{\mathcal{S}}\xi-b_k\rVert_W^2, \tag*{(8)}

with W=diag⁡((1+γ)I(nu+ny)Tini,I(nu+ny)N)W=\operatorname{diag}((1+\gamma)I_{(n_u+n_y)T_{\mathrm{ini}}},I_{(n_u+n_y)N}) and γ>0\gamma>0. The past block of WW penalises deviations from the measured history, with γ\gamma controlling how strongly the prediction is tied to it, and the future block penalises tracking errors. The radius ρ\rho sets the subspace uncertainty. Figure 1 illustrates the projections.

Figure 1. Geometric uncertainty in Gr⁡(1,2)\operatorname{Gr}(1,2). The fitted subspace S^\hat{\mathcal{S}} is the horizontal line. The shaded double cone is the ball Bρ(S^)\mathbb{B}_{\rho}(\hat{\mathcal{S}}) of lines with ∣θ∣<θρ=sin⁡−1ρ|\theta|<\theta_{\rho}=\sin^{-1}\rho, (dashed boundary lines are excluded). Projecting ξ\xi onto S^\hat{\mathcal{S}} and onto an adverse subspace Sa\mathcal{S}^{a} gives different trajectories, the adverse with the larger residual to bkb_k.

The objective in (8) is convex in ξ\xi as a supremum of convex functions, while the search for the adverse subspace is nonconvex. Because FkF_k is continuous in S\mathcal{S}, the supremum over the open ball equals the maximum over its closure. For the returned ξ\xi, the adverse subspace Sa\mathcal{S}^{a} maximises Fk(ξ,⋅)F_k(\xi,\cdot) over the ball. The controller decodes the trajectory PSaξP_{S_a}\xi and applies its first future input. Both inputs and outputs vary with the subspace. The ball also contains subspaces without a causal LTI realisation, so its use here is evaluated experimentally.

Planar 2R Manipulator Control

The mechanical model feedback-linearises the arm, and geometric DDPC predicts the resulting virtual-input–position response.

Manipulator Model and Tracking Task

The base angle q1q_1 is measured from the positive horizontal axis; q2q_2 is relative to link 1, so link 2 has absolute angle q1+q2q_1+q_2. Link jj has length ljl_j, mass mjm_j, centre-of-mass distance rjr_j from its proximal joint, and inertia IjI_j about its centre of mass. Figure 2 shows this convention and the centre-of-mass positions pc1,pc2p_{c1},p_{c2}. For uniform rods, rj=lj/2r_j=l_j/2 and Ij=mjlj2/12I_j=m_jl_j^2/12.

Figure 2. Planar 2R Manipulator.

Let q=col⁡(q1,q2)q=\operatorname{col}(q_1,q_2), v=q˙v=\dot{q}, and x=col⁡(q,v)x=\operatorname{col}(q,v). The joint dynamics are

M(q)q¨+η(q,v)+g(q)+Dvv=τact,(9)M(q)\ddot{q}+\eta(q,v)+g(q)+D_vv=\tau^{\mathrm{act}}, \tag*{(9)}

where M(q)≻0M(q)\succ0 is the inertia matrix, η\eta contains Coriolis and centrifugal terms, gg is the gravity torque, Dv⪰0D_v\succeq0 is diagonal viscous damping, and τact\tau^{\mathrm{act}} is the applied torque. The expressions for MM and η\eta are given in Appendix A. For gravitational acceleration of magnitude g0g_0 in the negative vertical direction,

g(q)=g0[(m1r1+m2l1)cos⁡q1+m2r2cos⁡(q1+q2)m2r2cos⁡(q1+q2)].(10)g(q)=g_0 \begin{bmatrix} (m_1r_1+m_2l_1)\cos q_1+m_2r_2\cos(q_1+q_2)\\ m_2r_2\cos(q_1+q_2) \end{bmatrix}. \tag*{(10)}

Positions are measured at tk=kTst_k=kT_s as y~k=q(tk)+εky\tilde{y}_k=q(t_k)+\varepsilon_k^y, with independent εky∼N(0,Σy)\varepsilon_k^y\sim\mathcal{N}(0,\Sigma_y). Velocity is estimated. The task is to track qref(t)q^{\mathrm{ref}}(t), assessed by position error and applied torque. Predictive problems impose no joint or torque constraints.

Feedback Linearisation and State Estimation

Write h(q,v)=η(q,v)+g(q)+Dvvh(q,v)=\eta(q,v)+g(q)+D_vv. The supplied model uses M^(q)\widehat{M}(q) and h^(q,v)\widehat{h}(q,v), and its parameters may differ from the plant. For a held virtual acceleration μk∈R2\mu_k\in\mathbb{R}^2, the inverse dynamics map gives

τid(t)=M^(q^(t))μk+h^(q^(t),v^(t)).(11)\tau^{\mathrm{id}}(t)=\widehat{M}(\widehat{q}(t))\mu_k+\widehat{h}(\widehat{q}(t),\widehat{v}(t)). \tag*{(11)}

With exact parameters, state feedback, continuous torque evaluation and inactive actuator limits, substitution gives q¨=μk\ddot{q}=\mu_k. The feedback-linearised system then consists of two double integrators, with exact sampled model xk+1=Axk+Bμkx_{k+1}=Ax_k+B\mu_k, yk=Cxky_k=Cx_k, where

A=[I2TsI20I2],B=[12Ts2I2TsI2],C=[I20].(12)A= \begin{bmatrix} I_2 & T_sI_2\\ 0 & I_2 \end{bmatrix}, \qquad B= \begin{bmatrix} \frac{1}{2}T_s^2I_2\\ T_sI_2 \end{bmatrix}, \qquad C=\begin{bmatrix}I_2&0\end{bmatrix}. \tag*{(12)}

The ideal model is controllable, has observability lag two, and has dim⁡BL=2L+4\dim\mathcal{B}_L = 2L + 4 for L≥2L \ge2. A Kalman filter estimates position and velocity from measured positions and commanded virtual inputs [15], using (12). Figure 3 shows the loop; Appendix A gives the filter.

Figure 3. Geometric DDPC with feedback linearisation. The observer uses the current position measurement. The predictor uses completed (μj,y~j)(\mu_j,\widetilde{y}_j) pairs through j=k−1j = k - 1.

The difference between the actual and continuously evaluated model torque is Δτact=τact−τid\Delta\tau^{\mathrm{act}} = \tau^{\mathrm{act}} - \tau^{\mathrm{id}}. This includes torque holding and actuator limitations. Consequently, q¨=μk+ares\ddot{q} = \mu_k + a^{\mathrm{res}}, with

ares=M(q)−1[(M^(q^)−M(q))μk+h^(q^,v^)−h(q,v)+Δτact].(13)a^{\mathrm{res}} = M(q)^{-1}\left[\left(\widehat{M}(\widehat{q}) - M(q)\right)\mu_k + \widehat{h}(\widehat{q},\widehat{v}) - h(q,v) + \Delta\tau^{\mathrm{act}}\right]. \tag*{(13)}

With exact state feedback and unrestricted continuous torque application, the dynamics become q¨=G(q)μk+fr(q,v)\ddot{q} = G(q)\mu_k + f_r(q,v), where G=M−1M^G = M^{-1}\widehat{M} and fr=M−1(h^−h)f_r = M^{-1}(\widehat{h}-h). The recorded response contains these model errors together with estimation and actuator effects, giving the learned predictor information beyond the ideal double integrator.

Offline Data and Reference Initialisation

A PD tracking controller with acceleration feedforward and random excitation generates offline data. Its sinusoidal reference qcol(t)q^{\mathrm{col}}(t) varies in centre, amplitude, phase and period between episodes (Appendix B).

For the feedback-linearised arm, uk=μku_k = \mu_k is held for TsT_s, while inverse dynamics is evaluated from the propagated estimate every TτT_\tau. For direct data, uk=τkcmdu_k = \tau_k^{\mathrm{cmd}} is computed once at tkt_k and held for TsT_s. The actuator enforces torque magnitude and slew limits at each TτT_\tau update in both cases. The mechanical model and state estimates are used to generate both offline datasets. Windows are concatenated without crossing episode boundaries.

Tracking references are joint sinusoids with a smooth entry ramp. Sampled qkrefq_k^{\mathrm{ref}}, vkrefv_k^{\mathrm{ref}}, μkref\mu_k^{\mathrm{ref}} satisfy the ideal dynamics through a piecewise-constant-acceleration interpolant. Torque and torque-rate requirements are checked throughout the motion (Appendix B).

Geometric DDPC for the Feedback-Linearised Arm

A rank-d=2L+4d = 2L + 4 fit to records of (μk,y~k)(\mu_k,\widetilde{y}_k) defines S^\widehat{\mathcal{S}}. This dimension follows the ideal four-state model; the affine fitting coordinates are specified in Appendix § 9.§ 1.

At update kk, the target bkb_k contains the preceding TiniT_{\mathrm{ini}} measured pairs and the next NN reference pairs (μjref,qjref)(\mu_j^{\mathrm{ref}},q_j^{\mathrm{ref}}), expressed in the predictor’s coordinates. The controller solves (8) over the open ball in Gr⁡(2L+4,4L)\operatorname{Gr}(2L + 4,4L), with fixed ρ\rho and γ\gamma. A scenario-exchange algorithm returns an approximate adverse trajectory, whose first future virtual input is applied through (11). Appendix § 9.§ 3 gives the solve and stopping rules.

Comparison Controllers

Double-integrator MPC. This controller predicts with (12), using the same horizon, observer, feedback-linearising torque map and actuator as geometric DDPC. It minimises (2) with u=μu=\mu and y=qy=q, using the same future tracking weights as geometric DDPC in physical coordinates. Its first virtual input is applied through the shared torque map.

Direct regularised DeePC. Direct DeePC uses the recorded torque and position pairs to predict future motion [6, 14]. In Figure 4, the desired motion defines the torque reference τkref=M^(qkref)μkref+h^(qkref,vkref)\tau_k^{\mathrm{ref}}=\widehat{M}(q_k^{\mathrm{ref}})\mu_k^{\mathrm{ref}}+\widehat{h}(q_k^{\mathrm{ref}},v_k^{\mathrm{ref}}). Set uref=τrefu^{\mathrm{ref}}=\tau^{\mathrm{ref}} and yref=qrefy^{\mathrm{ref}}=q^{\mathrm{ref}} in both the tracking cost and grefg^{\mathrm{ref}} (5). At each update, the measured history initialises (6), whose first predicted torque is applied directly. The model supplies references and collection torques; online feedback uses measured input–output history. Appendix § 9.§ 5 gives the quadratic solve.

Direct regularised DeePC

Figure 4. Direct regularised DeePC. The desired torque and position sequences define the reference, while completed torque–position pairs update the history and online data bank.

Simulation Studies

We examine prediction accuracy (A), transfer of predictive tuning (B), and cyclic tracking (C). Controllers receive noisy positions; true positions are used only to evaluate errors.

Simulation Setup and Evaluation

Table 1 defines the mechanical conditions. Experiments A and C retain a fixed model while changing the plant parameters. Experiment B uses exact models of two known arms. In all cases, gravity acts downward and position noise has covariance Σy=σy2I2\Sigma_y=\sigma_y^2 I_2 during collection and control.

ConditionStudiesLengths [m]True masses [kg]Model masses [kg]
Exact parametersA, C(1.0, 0.8)(0.9, 0.72)(0.9, 0.72)
Parameter errorsA, C(1.0, 0.8)(1.0, 0.8)(0.9, 0.72)
Arm AB(1.0, 0.8)(1.0, 0.8)(1.0, 0.8)
Arm BB(0.75, 0.55)(1.3, 0.6)(1.3, 0.6)

Table 1. Mechanical conditions. Model lengths equal true lengths. True damping is (0.04, 0.02) N m s rad−1^{-1} in the parameter-error condition and zero elsewhere; model damping is zero throughout.

Controller settings are fixed before evaluation. Direct-DeePC weights are selected on separate Arm A trials at baseline noise. The data-driven methods share the collection procedure and budget, with their respective input conventions. Trials use paired measurement-noise sequences across controllers, and predictive computation is instantaneous. Each run covers a smooth entry ramp and two reference cycles. Appendix § 10 gives the settings and evaluation protocol.

Tables report median [minimum, maximum] periodic position RMSE EcycE_{\mathrm{cyc}}, median whole-run applied-torque RMS τrms\tau_{\mathrm{rms}}, and completed trials ncn_{\mathrm{c}}. Both measures average over the two joints. Incomplete trials enter completion counts but are excluded from full-duration error and effort statistics.

Prediction Accuracy After Feedback Linearisation

Experiment A compares the double integrator with the predictor fitted to virtual-input–position data. On independent test trajectories, both predict future positions from the same prescribed inputs and past measured positions ([11]. Future positions enter only the error calculation. Table 2 covers exact and inaccurate mechanics at three noise levels.

Mechanicsσy\sigma_y [rad]Double integratorLearned behaviour
Exact parameters0.0010.0098 [0.0068, 0.0110]0.0058 [0.0051, 0.0070]
0.0050.0157 [0.0118, 0.0191]0.0129 [0.0094, 0.0147]
0.0100.0251 [0.0213, 0.0279]0.0200 [0.0191, 0.0242]
Parameter errors0.0010.2096 [0.1831, 0.2496]0.0153 [0.0131, 0.0194]
0.0050.1801 [0.1510, 0.2207]0.0167 [0.0104, 0.0183]
0.0100.1598 [0.1292, 0.1916]0.0234 [0.0187, 0.0270]

Table 2. Position prediction RMSE [rad]. Entries are medians [minimum, maximum] across independently fitted models, after taking the median error over each model’s test windows.

Learning gives a modest improvement with exact parameters and a much larger improvement with parameter errors, indicating that data capture systematic departures from the ideal dynamics. The learned predictor remains more accurate at each noise level, although its error grows with noise. Noise also changes the feedback-controlled collection trajectories, so comparisons across noise levels include this effect.

Transfer Across Noise and Arm Changes

Experiment B uses the base cycle with unchanged predictive tuning. Noise is increased on Arm A, then a known Arm B with shorter links and different masses is tested. Each condition uses newly collected data and the stated observer noise covariance; the mechanical model is updated for Arm B. The collection procedure and predictive settings stay fixed.

Tracking error increases with noise for both methods (Table 3). Geometric DDPC retains lower median error and completes every trial. At high noise, its torque variation and actuator activity increase (Figure 5); direct DeePC has a wider error spread and one incomplete trial. On Arm B, both methods retain accuracy close to their Arm A baselines with reduced torque effort. Geometric DDPC remains more accurate at similar effort.

ConditionControllerEcycE_{\mathrm{cyc}}τrms\tau_{\mathrm{rms}}nc/6n_c/6
Arm A, σy=0.001\sigma_y = 0.001Geometric DDPC0.0042 [0.0038, 0.0054]2.9846/6
Direct DeePC0.0245 [0.0236, 0.0266]2.9616/6
Arm A, σy=0.005\sigma_y = 0.005Geometric DDPC0.0138 [0.0116, 0.0164]2.9776/6
Direct DeePC0.0370 [0.0285, 0.0545]2.9556/6
Arm A, σy=0.010\sigma_y = 0.010Geometric DDPC0.0280 [0.0210, 0.0345]3.5606/6
Direct DeePC0.0662 [0.0369, 0.2517]2.9495/6
Arm B, σy=0.001\sigma_y = 0.001Geometric DDPC0.0048 [0.0039, 0.0055]1.9576/6
Direct DeePC0.0257 [0.0227, 0.0323]1.9486/6

Table 3. Transfer with fixed predictive settings. Errors are in radians and torque RMS in N m; nc/6n_c/6 gives completed/attempted trials.

Figure 5. Noise transfer with geometric DDPC on Arm A: second-joint position and applied first-joint torque in the first trial, after the ramp. Legend values give σy\sigma_y in radians.

Trajectory Tracking

Experiment C compares geometric DDPC, double-integrator MPC and direct regularised DeePC. The base reference is tested with exact mechanical parameters; all three references in Table 4 are tested with parameter errors. The mechanical model stays fixed, and offline data are reused across references within each plant condition. Sampled reference torques satisfy the actuator limits.

ReferenceCentre [rad]Amplitudes [rad]Period [s]
Base(−π/2,0.4)(-\pi/2, 0.4)(0.35,0.50)(0.35, 0.50)6
Shifted(−π/2+0.05,0.48)(-\pi/2 + 0.05, 0.48)(0.42,0.58)(0.42, 0.58)6
Faster(−π/2,0.4)(-\pi/2, 0.4)(0.30,0.42)(0.30, 0.42)4.5

Table 4. Joint references, with phases (0,π/3)(0, \pi/3) and a two-second entry ramp.

Table 5 summarises tracking accuracy and effort. Figure 6 shows the base-cycle trajectories with parameter errors.

Condition, cycleControllerEcycE_{\mathrm{cyc}}τrms\tau_{\mathrm{rms}}nc/6n_c/6
Exact parameters, baseDouble-integrator MPC0.0050 [0.0046, 0.0053]2.7036/6
Geometric DDPC0.0050 [0.0047, 0.0064]2.6906/6
Direct DeePC0.0253 [0.0213, 0.0292]2.6656/6
Parameter errors, baseDouble-integrator MPC0.1091 [0.1090, 0.1093]2.8106/6
Geometric DDPC0.0650 [0.0482, 0.0749]3.0306/6
Direct DeePC0.0826 [0.0636, 0.0848]2.5306/6
Parameter errors, shiftedDouble-integrator MPC0.1152 [0.1150, 0.1154]3.4566/6
Geometric DDPC0.0592 [0.0412, 0.0744]3.6816/6
Direct DeePC0.1365 [0.1089, 0.1411]3.0436/6
Parameter errors, fasterDouble-integrator MPC0.1068 [0.1068, 0.1069]2.2096/6
Geometric DDPC0.0779 [0.0648, 0.0866]2.3816/6
Direct DeePC0.0731 [0.0594, 0.0771]1.9776/6

Table 5. Cyclic tracking at σy=0.001\sigma_y = 0.001 rad. Errors are in radians and torque RMS in N m; nc/6n_c/6 gives completed/attempted trials.

Figure 6. Base-cycle tracking with mass and damping errors. Both joint positions are shown for the first trial of each controller, together with the reference.

With exact mechanics, double-integrator MPC and geometric DDPC give similar accuracy and effort, both outperforming direct DeePC in tracking. With parameter errors, geometric DDPC improves on double-integrator MPC in every paired trial at greater effort. It also has lower median error than direct DeePC on the base and shifted cycles, while direct DeePC has lower error and effort on the faster cycle. All tracking trials complete with inactive actuator limits.

Discussion and Conclusions

This study examined geometric robust DDPC for a feedback-linearised planar 2R arm. The results identify a useful role for data when the mechanical model captures the main structure but leaves systematic prediction errors. Learning the resulting virtual-input–position response improves prediction and tracking under parameter errors. With exact mechanical parameters, the known double integrator already provides an effective controller. Performance relative to direct regularised DeePC depends on the reference, with different trade-offs between accuracy and torque effort.

The predictive settings also remained effective across the tested noise levels and arm parameters. This transfer used newly collected data and the corresponding mechanical-model and observer updates. Higher measurement noise increased tracking error and actuator activity, although every geometric-DDPC trial completed.

Generative AI Use

Generative AI tools assisted with literature review, preparation of TikZ illustrations, simulation-code optimisation, and auditing the mathematical exposition/results.

Mechanical Model, State Estimation and Torque Application

Inertia and velocity coupling. For the coordinates and link parameters in § 3, the inertia and velocity-coupling terms are

M(q)=[α+2βcos⁡q2δ+βcos⁡q2δ+βcos⁡q2δ],η(q,v)=βsin⁡q2[−2v1v2−v22v12].(14)M(q)=\begin{bmatrix}\alpha+2\beta\cos q_{2} & \delta+\beta\cos q_{2}\\ \delta+\beta\cos q_{2} & \delta\end{bmatrix},\qquad\eta(q,v)=\beta\sin q_{2}\begin{bmatrix}-2v_{1}v_{2}-v_{2}^{2}\\ v_{1}^{2}\end{bmatrix}. \tag*{(14)}

Here α=I1+I2+m1r12+m2(l12+r22)\alpha=\mathcal{I}_{1}+\mathcal{I}_{2}+m_{1}r_{1}^{2}+m_{2}(l_{1}^{2}+r_{2}^{2}), β=m2l1r2\beta=m_{2}l_{1}r_{2}, and δ=I2+m2r22\delta=\mathcal{I}_{2}+m_{2}r_{2}^{2}.

The estimator uses the double-integrator matrices AA, BB, CC in (12) and measured joint positions.

Position-based state estimation. A Kalman filter on the ideal double-integrator model estimates position and velocity [15]. Let x^k−\widehat{x}_{k}^{-} and x^k+\widehat{x}_{k}^{+} be the prior and corrected estimates, with covariances Σk−\Sigma_{k}^{-} and Σk+\Sigma_{k}^{+}. The current measurement gives

Kk=Σk−CT(CΣk−CT+Σy)−1,x^k+=x^k−+Kk(y~k−Cx^k−),K_{k}=\Sigma_{k}^{-}C^{\mathsf{T}}(C\Sigma_{k}^{-}C^{\mathsf{T}}+\Sigma_{y})^{-1},\qquad\widehat{x}_{k}^{+}=\widehat{x}_{k}^{-}+K_{k}(\widetilde{y}_{k}-C\widehat{x}_{k}^{-}),
Σk+=(I4−KkC)Σk−(I4−KkC)T+KkΣyKkT.(15)\Sigma_{k}^{+}=(I_{4}-K_{k}C)\Sigma_{k}^{-}(I_{4}-K_{k}C)^{\mathsf{T}}+K_{k}\Sigma_{y}K_{k}^{\mathsf{T}}. \tag*{(15)}

After selecting μk\mu_{k}, propagate x^k+1−=Ax^k++Bμk\widehat{x}_{k+1}^{-}=A\widehat{x}_{k}^{+}+B\mu_{k} and Σk+1−=AΣk+AT+QKF\Sigma_{k+1}^{-}=A\Sigma_{k}^{+}A^{\mathsf{T}}+Q^{\mathrm{KF}}, where QKF=BΣaBTQ^{\mathrm{KF}}=B\Sigma_{a}B^{\mathsf{T}} and Σa\Sigma_{a} is a chosen acceleration-error covariance. The covariance models uncertainty in the observer prediction; the plant has no added process noise.

Between measurements, inverse dynamics uses x^(tk+s)=A(s)x^k++B(s)μk\widehat{x}(t_{k}+s)=A(s)\widehat{x}_{k}^{+}+B(s)\mu_{k} for 0≤s<Ts0\leq s<T_{s}, with A(s)A(s), B(s)B(s) obtained by replacing TsT_{s} in (12) by ss. The initial estimate uses the reset position and zero velocity. Commanded virtual inputs drive the filter during collection, startup and tracking.

Torque application. Let TτT_{\tau} divide TsT_{s}. At tk,m=tk+mTτt_{k,m}=t_{k}+mT_{\tau}, for m=0,…,Ts/Tτ−1m=0,\ldots,T_{s}/T_{\tau}-1, compute τk,mcmd=τid(tk,m)\tau_{k,m}^{\mathrm{cmd}}=\tau^{\mathrm{id}}(t_{k,m}). Given the preceding applied torque τprev\tau^{\mathrm{prev}}, positive magnitude bounds τmax⁡\tau_{\max}, and slew bounds rτr_{\tau}, apply

τk,msat=clip⁡(τk,mcmd,−τmax⁡,τmax⁡),\tau_{k,m}^{\mathrm{sat}}=\operatorname{clip}(\tau_{k,m}^{\mathrm{cmd}},-\tau_{\max},\tau_{\max}),
τk,mact=τprev+clip⁡(τk,msat−τprev,−Tτrτ,Tτrτ).(16)\tau_{k,m}^{\mathrm{act}}=\tau^{\mathrm{prev}}+\operatorname{clip}(\tau_{k,m}^{\mathrm{sat}}-\tau^{\mathrm{prev}},-T_{\tau}r_{\tau},T_{\tau}r_{\tau}). \tag*{(16)}

Here clip⁡(z,a,b)\operatorname{clip}(z,a,b) truncates each component of zz to its interval [aj,bj][a_{j},b_{j}]. The torque is held until the next update. Torque holding and actuator modification enter Δτact\Delta\tau^{\mathrm{act}} in (13).

Data Collection and Reference Construction

Data collection. A PD tracking controller with acceleration feedforward follows an auxiliary sinusoid qcol(t)q^{\mathrm{col}}(t) and adds random virtual excitation:

μkd=q¨col(tk)+Kp(qcol(tk)−q^k+)+Kd(q˙col(tk)−v^k+)+ekd.(17)\mu_{k}^{d}=\ddot{q}^{\mathrm{col}}(t_{k})+K_{p}(q^{\mathrm{col}}(t_{k})-\widehat{q}_{k}^{+})+K_{d}(\dot{q}^{\mathrm{col}}(t_{k})-\widehat{v}_{k}^{+})+e_{k}^{d}. \tag*{(17)}

The diagonal gains Kp,Kd≻0K_{p},K_{d}\succ0 are fixed. The sinusoid varies between episodes, and ekde_{k}^{d} is independent, zero-mean Gaussian excitation. The complete command also depends on measured feedback.

For the feedback-linearised dataset, the recorded command is μkd\mu_{k}^{d}. For the direct dataset, it is τkd=M^(q^k+)μkd+h^(q^k+,v^k+)\tau_{k}^{d}=\widehat{M}(\widehat{q}_{k}^{+})\mu_{k}^{d}+\widehat{h}(\widehat{q}_{k}^{+},\widehat{v}_{k}^{+}). Both use the supplied model and estimated state, with the input-holding conventions of § 3.3.

Position is sampled before the new command acts. The recorded pair is (uk,yk)(u_{k},y_{k}), where uk=μku_{k}=\mu_{k} for the feedback-linearised system and uk=τkcmdu_{k}=\tau_{k}^{\mathrm{cmd}} for direct DeePC. Inputs are stored before actuator modification. An episode of Te≥LT_{e}\geq L samples supplies Te−L+1T_{e}-L+1 columns, giving ng=∑e(Te−L+1)n_{g}=\sum_{e}(T_{e}-L+1) offline windows. All windows remain within their episodes.

Reference construction. The evaluation references are

qjref(t)=qc,j+ajχ(t/Tr)sin⁡(ωt+ϕj),j=1,2,(18)q_j^{\mathrm{ref}}(t) = q_{c,j} + a_j \chi(t/T_r)\sin(\omega t + \phi_j), \qquad j = 1,2, \tag*{(18)}

with centre qcq_c, amplitudes aja_j, phases ϕj\phi_j, angular frequency ω\omega and ramp duration TrT_r. The ramp χ(z)=10z3−15z4+6z5\chi(z) = 10z^3 - 15z^4 + 6z^5 for 0≤z≤10 \le z \le1, extended by zero before and one after this interval, gives continuous position, velocity and acceleration.

The desired position samples are qkref=qref(tk)q_k^{\mathrm{ref}} = q^{\mathrm{ref}}(t_k). A virtual-input reference consistent with the ideal sampled dynamics is obtained from v0ref=q˙ref(0)v_0^{\mathrm{ref}} = \dot{q}^{\mathrm{ref}}(0) and

μkref=2Ts2(qk+1ref−qkref−Tsvkref),vk+1ref=vkref+Tsμkref.(19)\mu_k^{\mathrm{ref}} = \frac{2}{T_s^2}(q_{k+1}^{\mathrm{ref}} - q_k^{\mathrm{ref}} - T_s v_k^{\mathrm{ref}}), \qquad v_{k+1}^{\mathrm{ref}} = v_k^{\mathrm{ref}} + T_s\mu_k^{\mathrm{ref}}. \tag*{(19)}

This defines a piecewise-constant-acceleration interpolant through the desired positions. Its discrete velocity can differ from the analytic reference derivative. Torque and torque-rate demands of this interpolant are checked through the ramp and periodic motion.

Initial history and sample timing. Before predictive tracking, the arm is held near the initial reference for max⁡(Tini,12)\max(T_{\mathrm{ini}}, 12) sampling intervals, using the collection controller with a constant reference and zero added excitation. The final plant and observer states of this hold initialise the tracking run.

At update kk, the current measurement corrects the observer while the predictor uses completed pairs through k−1k - 1. The selected command is applied immediately in simulated time. After the interval, (uk,y~k)(u_k,\tilde{y}_k) replaces the oldest history pair.

For the direct coefficient reference, extend the desired history before time zero by qjref=q0refq_j^{\mathrm{ref}} = q_0^{\mathrm{ref}}, vjref=0v_j^{\mathrm{ref}} = 0, and τjref=h^(q0ref,0)\tau_j^{\mathrm{ref}} = \hat{h}(q_0^{\mathrm{ref}},0) for j<0j < 0. At nonnegative samples, use the torque reference in § 3.5 with the sampled velocity and acceleration in (19). The measured history enters the predictive consistency terms separately.

Geometric Representation and Predictive Solves

Affine, Dimensionless Trajectory Coordinates

Let H~wd∈R4L×ng\widetilde{H}_{w}^{d} \in\mathbb{R}^{4L \times n_g} be the concatenated Hankel matrix of measured pairs col⁡(μk,y~k)\operatorname{col}(\mu_k,\widetilde{y}_k), ordered by sample. Fix positive scales sμ,sy∈R2s_\mu, s_y \in\mathbb{R}^2 for virtual input and position, and set Ds=IL⊗diag⁡(sμ,1,sμ,2,sy,1,sy,2)D_s = I_L \otimes\operatorname{diag}(s_{\mu,1},s_{\mu,2},s_{y,1},s_{y,2}). Define the mean window cc, scaled centred matrix ZZ, and fitted basis by

c=1ngH~wd1ngT,Z=Ds−1(H~wd−c1ngT),c = \frac{1}{n_g}\widetilde{H}_{w}^{d}\mathbf{1}_{n_g}^{\mathsf{T}}, \qquad Z = D_s^{-1}(\widetilde{H}_{w}^{d} - c\mathbf{1}_{n_g}^{\mathsf{T}}),
d=2L+4,V^=svd⁡d(Z),S^=im⁡V^.(20)d = 2L + 4, \qquad\widehat{V} = \operatorname{svd}_d(Z), \qquad\widehat{S} = \operatorname{im}\widehat{V}. \tag*{(20)}

Here svd⁡d\operatorname{svd}_d returns the leading dd left singular vectors. Centring limits the data rank to at most ng−1n_g - 1, so the fit requires ng>dn_g > d and an adequately excited ddth direction. This rank follows the ideal four-state model and approximates the implemented response.

Each candidate SS describes the physical affine set c+DsSc+D_sS. The mean, scales and fitted subspace remain fixed during a run. Scaling sets both the relative input/output weights and the geometry used to compare subspaces.

Let w~j=col⁡(μj,y~j)\widetilde{w}_j=\operatorname{col}(\mu_j,\widetilde{y}_j) and wjref=col⁡(μjref,qjref)w_j^{\mathrm{ref}}=\operatorname{col}(\mu_j^{\mathrm{ref}},q_j^{\mathrm{ref}}). Stack the measured past and desired future, then transform them into the fitting coordinates:

b‾k=col⁡(w~k−Tini:k−1,wk:k+N−1ref),bk=Ds−1(b‾k−c).\overline{b}_k=\operatorname{col}\left(\widetilde{w}_{k-T_{\mathrm{ini}}:k-1},w_{k:k+N-1}^{\mathrm{ref}}\right),\qquad b_k=D_s^{-1}(\overline{b}_k-c).

For the returned trajectory parameter and adverse subspace, decode

wka=c+DsPSkaξk,μk=Sμwka,Sμ=[02×4TiniI202×(4N−2)].(21)w_k^{\mathrm{a}}=c+D_sP_{S_k^a}\xi_k,\qquad\mu_k=S_\mu w_k^{\mathrm{a}},\qquad S_\mu=\begin{bmatrix}0_{2\times4T_{\mathrm{ini}}}&I_2&0_{2\times(4N-2)}\end{bmatrix}. \tag*{(21)}

The selector extracts the first future virtual input. Physical future tracking weights are the inverse squared channel scales.

Prediction for Prescribed Inputs

For the prediction tests, write a fitted trajectory as w(ζ)=c+DsVζw(\zeta)=c+D_sV\zeta, with ζ∈Rd\zeta\in\mathbb{R}^d. Let SuS_u select all 2L2L input entries and Sy,pS_{y,p} the 2Tini2T_{\mathrm{ini}} past output entries. Given the complete prescribed input stack utestu^{\mathrm{test}} and past measured positions y~test,p\widetilde{y}^{\mathrm{test},p}, compute

min⁡ζ∈Rd∥Sy,pw(ζ)−y~test,p∥Qp2subject toSuw(ζ)=utest,Qp=ITini⊗diag⁡(sy,1−2,sy,2−2).(22)\begin{aligned} \min_{\zeta\in\mathbb{R}^d}\quad&\left\|S_{y,p}w(\zeta)-\widetilde{y}^{\mathrm{test},p}\right\|_{Q_p}^2\\ \text{subject to}\quad&S_uw(\zeta)=u^{\mathrm{test}},\qquad Q_p=I_{T_{\mathrm{ini}}}\otimes\operatorname{diag}(s_{y,1}^{-2},s_{y,2}^{-2}). \tag*{(22)} \end{aligned}

The future output entries are the forecast. For the known double integrator, use its exact behaviour and zero affine mean in the same solve. Future positions enter only the error calculation.

Sine Chart and Scenario-Exchange Optimisation

The scenario-exchange method alternates a convex master solve with an adverse-subspace search. The master minimises the largest cost over its current subspaces. The adverse search adds a subspace with a larger cost at the returned ξ\xi. The master uses the bound ∥ξ∥2≤(bkTWbk+∥bk∥2)/ρ\|\xi\|_2\leq(\sqrt{b_k^{\mathsf{T}}Wb_k}+\|b_k\|_2)/\rho, which contains the sublevel set whose worst-case cost is no greater than the cost at ξ=0\xi=0.

If V^⊥\widehat{V}_{\perp} is an orthonormal basis of the complement of S^\widehat{S}, the sine chart is

V(X)=V^(Id−XTX)1/2+V^⊥X,X∈R(4L−d)×d,∥X∥F≤ρ.(23)V(X)=\widehat{V}(I_d-X^{\mathsf{T}}X)^{1/2}+\widehat{V}_{\perp}X,\qquad X\in\mathbb{R}^{(4L-d)\times d},\qquad\|X\|_{\mathrm{F}}\leq\rho. \tag*{(23)}

For ρ<1\rho<1, the positive-definite square root exists and the resulting basis is orthonormal. Its chordal distance from the fitted subspace is dch(im⁡V(X),S^)=∥X∥Fd_{\mathrm{ch}}(\operatorname{im}V(X),\widehat{S})=\|X\|_{\mathrm{F}}. Multistart local searches operate on this closed chart ball. Its extremal value equals the open-ball supremum; the selected chart point is restricted to ∥X∥F≤(1−10−7)ρ\|X\|_{\mathrm{F}}\leq(1-10^{-7})\rho before command extraction, keeping the executed projection strictly inside the open ball.

Warm starts and an invertible preconditioning of the master improve its numerical solution without changing the objective. Let JkUJ_k^{\mathrm{U}} be the numerical upper estimate of the worst-case cost at the returned parameter, JkLJ_k^{\mathrm{L}} the master-based lower estimate of the optimum, and FkaF_k^{\mathrm{a}} the cost at the returned adverse projection. The recorded relative gaps are

δkout=max⁡{0,JkU−JkL}1+∣JkL∣,δkadv=max⁡{0,JkU−Fka}1+∣JkU∣.(24)\delta_k^{\mathrm{out}}=\frac{\max\{0,J_k^{\mathrm{U}}-J_k^{\mathrm{L}}\}}{1+\lvert J_k^{\mathrm{L}}\rvert},\qquad \delta_k^{\mathrm{adv}}=\frac{\max\{0,J_k^{\mathrm{U}}-F_k^{\mathrm{a}}\}}{1+\lvert J_k^{\mathrm{U}}\rvert}. \tag*{(24)}

These floating-point estimates assess the numerical solve. A finite candidate can be applied at the iteration limit with an open gap, and its termination status is retained in the results. Table 6 lists the numerical settings.

Numerical settingValue
Exchange iteration limit20
Adverse searchThree starts, up to 80 iterations each
Master / spectral-search limits500 / 80 iterations
Relative outer / adverse gap targets10−410^{-4} / 10−410^{-4}
Inner / master tolerances10−610^{-6} / 10−910^{-9}
Extracted chart radiusAt most (1−10−7)ρ(1 - 10^{-7})\rho

Table 6. Numerical settings for geometric DDPC. Warm starts are enabled.

Across the 6540 geometric updates, 404 meet both gap targets and 6136 reach the exchange-iteration limit. The median relative outer gap is 0.453, with maximum 0.765; the median adverse-search gap is 1.19×10−71.19 \times10^{-7}, with maximum 0.491. These bound separations leave the suboptimality of the executed decisions unresolved. Every executed subspace lies strictly inside the radius-0.01 ball.

Double-Integrator MPC

Let ON=col⁡(C,CA,…,CAN−1)\mathcal{O}_N = \operatorname{col}(C, CA, \ldots, CA^{N-1}). Define TN\mathcal{T}_N by its (i,j)(i,j) block: CAi−j−1BCA^{i-j-1}B for $0 \le j < i < N$, and zero otherwise. The predicted output is y=ONx^k+TNμ\mathbf{y} = \mathcal{O}_N\hat{x}_k+\mathcal{T}_N\boldsymbol{\mu}. With Qˉ=IN⊗Q\bar{Q}=I_N\otimes Q and Rˉ=IN⊗R\bar{R}=I_N\otimes R, the minimiser satisfies

(TNTQˉTN+Rˉ)μk=TNTQˉ(qkref−ONx^k)+Rˉμkref.(25)(\mathcal{T}_N^\mathsf{T}\bar{Q}\mathcal{T}_N+\bar{R})\mu_k=\mathcal{T}_N^\mathsf{T}\bar{Q}(q_k^{\mathrm{ref}}-\mathcal{O}_N\hat{x}_k)+\bar{R}\mu_k^{\mathrm{ref}}. \tag*{(25)}

The positive-definite coefficient matrix is factored once; the reference and observer prior update the right-hand side.

Reference-Centred Direct DeePC

Let H=HˉdH=\bar{H}^{\mathrm{d}} be the current direct data matrix, including online windows. The reference centre is gkref=H†rkg_k^{\mathrm{ref}}=H^\dagger r_k, with rkr_k defined in (5).

For a wide matrix, take a thin QR factorisation HT=QcRcH^\mathsf{T}=Q_cR_c, with QcTQc=IQ_c^\mathsf{T}Q_c=I, and set Hc=RcTH_c=R_c^\mathsf{T}. Then H=HcQcTH=H_cQ_c^\mathsf{T}. Decompose g=Qcζ+g⊥g=Q_c\zeta+g_\perp, where QcTg⊥=0Q_c^\mathsf{T}g_\perp=0. Since gkrefg_k^{\mathrm{ref}} belongs to the row space of HH, it can be written as QcζkrefQ_c\zeta_k^{\mathrm{ref}}, with ζkref=Hc†rk\zeta_k^{\mathrm{ref}}=H_c^\dagger r_k. Hence

Hg=Hcζ,∥g−gkref∥22=∥ζ−ζkref∥22+∥g⊥∥22.(26)Hg=H_c\zeta,\qquad\left\|g-g_k^{\mathrm{ref}}\right\|_2^2=\left\|\zeta-\zeta_k^{\mathrm{ref}}\right\|_2^2+\left\|g_\perp\right\|_2^2. \tag*{(26)}

The omitted component changes no prediction and is zero at the optimum. Thus the reduction is exact for the reference-centred squared penalty. When compression is unnecessary, use Qc=IQ_c=I and Hc=HH_c=H. The pseudoinverse is evaluated by a singular-value decomposition with relative cutoff 10−1210^{-12}.

Partition Hc=col⁡(Ucp,Ycp,Uf,Yf)H_c = \operatorname{col}(U_c^p, Y_c^p, U_f, Y_f) and set Qˉ=IN⊗Q\bar{Q} = I_N \otimes Q, Rˉτ=IN⊗Rτ\bar{R}^{\tau} = I_N \otimes R^{\tau}. Let τkref=col⁡(τkref,…,τk+N−1ref)\tau_k^{\mathrm{ref}} = \operatorname{col}(\tau_k^{\mathrm{ref}}, \ldots, \tau_{k+N-1}^{\mathrm{ref}}). The condensed quadratic has coefficient matrix and linear term

Ag=(Ycf)⊤QˉYcf+(Ucf)⊤RˉτUcf+λy(Ycp)⊤Ycp+λgI,(27)A_g = (Y_c^f)^\top\bar{Q}Y_c^f + (U_c^f)^\top\bar{R}^{\tau}U_c^f + \lambda_y(Y_c^p)^\top Y_c^p + \lambda_g I, \tag*{(27)}
hk=(Ycf)⊤Qˉqkref+(Ucf)⊤Rˉττkref+λy(Ycp)⊤ykini+λgζkref.h_k = (Y_c^f)^\top\bar{Q}q_k^{\mathrm{ref}} + (U_c^f)^\top\bar{R}^{\tau}\tau_k^{\mathrm{ref}} + \lambda_y(Y_c^p)^\top y_k^{\mathrm{ini}} + \lambda_g\zeta_k^{\mathrm{ref}}.

The stationarity and equality equations are

\begin{bmatrix} A_g & (U_c^p)^\top\\ U_c^p & 0 \end{bmatrix} \begin{bmatrix} \zeta_k \\ \nu_k \end{bmatrix} = \begin{bmatrix} h_k \\ u_k^{\mathrm{ini}} \end{bmatrix}, \tag*{(28)}\end{bmatrix}

where νk\nu_k is a scaled equality multiplier. Cholesky factors and a Schur complement solve this system. Scaled equality and stationarity residuals are checked before accepting the command. Data updates change the factors; reference and history updates change the linear term.

Simulation Parameters and Evaluation

Fixed Settings and Data Collection

All links are uniform rods with g0=9.81 m s−2g_0 = 9.81\ \mathrm{m}\,\mathrm{s}^{-2}. The plant and model parameters are given in Table 1. Table 7 lists the shared controller and simulation settings. The observer uses Σy=σy2I2\Sigma_y = \sigma_y^2 I_2, Σa=σa2I2\Sigma_a = \sigma_a^2 I_2, and diagonal initial covariance obtained from the listed standard deviations.

SettingValue
Predictive / torque-update periodsTs=0.1 sT_s = 0.1\ \mathrm{s}, Tτ=0.01 sT_\tau= 0.01\ \mathrm{s}
History, horizon and rankTini=4T_{\mathrm{ini}} = 4, N=8N = 8, L=12L = 12; d=28d = 28 in R48\mathbb{R}^{48}
Offline data per method and trialEight episodes of 60 samples; 392 windows
Collection gains and excitationKp=12I2K_p = 12I_2, Kd=7I2K_d = 7I_2; excitation standard deviation 0.7 rad s−2\mathrm{rad}\,\mathrm{s}^{-2} per joint
Baseline position noiseσy=0.001\sigma_y = 0.001 rad per joint
Observer uncertaintyσa=1 rad s−2\sigma_a = 1\ \mathrm{rad}\,\mathrm{s}^{-2}; initial position and velocity standard deviations 0.01 rad and 0.1 rad s−1\mathrm{rad}\,\mathrm{s}^{-1}
Geometric settingsγ=300\gamma= 300, ρ=0.01\rho= 0.01; position scale 0.1 rad and virtual-input scale 2 rad s−2\mathrm{rad}\,\mathrm{s}^{-2} per joint
Direct DeePC settingsQ=100I2Q = 100I_2, Rτ=0.3I2R^\tau= 0.3I_2, λy=3×103\lambda_y = 3 \times10^3, λg=100\lambda_g = 100; centre gkrefg_k^{\mathrm{ref}}
Torque magnitude / slew limits(15,5)(15, 5) N m / (300,100)(300, 100) N m s−1\mathrm{s}^{-1}
Plant integrationFourth-order Runge–Kutta, 0.005 s steps

Table 7. Shared simulation settings. The transfer study changes the specified arm and sensor-noise covariance while retaining predictive tuning and the collection procedure.

The auxiliary sinusoids have uniformly sampled centres in (−π/2±0.06,0.4±0.08)(-\pi/2 \pm0.06, 0.4 \pm0.08) rad, amplitudes in [0.25,0.45][0.25, 0.45] and [0.35,0.60][0.35, 0.60] rad, periods in [5,7.5][5, 7.5] s, and phases in [0,2π)[0, 2\pi). Episodes start at zero velocity with an independent 0.02 rad Gaussian position perturbation. Direct DeePC retains all offline columns and up to 392 completed online windows.

Direct-DeePC weights were selected on four separate Arm A trials with the base reference and baseline noise, favouring completed runs with low cyclic error. The selected weights, torque reference and reference-centred penalty are then used in every evaluation condition. No condition-specific retuning is performed.

References and Performance Measures

The references in Table 4 have phases (0,π/3)(0, \pi/3) and a two-second ramp. Evaluation follows a 1.2-second stabilising hold and lasts for the ramp and two cycles: 14 seconds for the base and shifted references, and 11 seconds for the faster reference. Each condition uses six independent data/noise pairs, with measurement noise paired across controllers.

For a completed run of KK intervals, define ek=q(tk)−qkrefe_k = q(t_k) - q_k^{\mathrm{ref}} and the periodic indices Kc={k:0≤k<K, tk≥Tr}\mathcal{K}_{\mathrm{c}} = \{k : 0 \le k < K,\, t_k \ge T_{\mathrm{r}}\}. The reported measures are

Ecyc=12∣Kc∣∑k∈Kc∥ek∥22,τrms=12KTs∫0KTs∥τact(t)∥22 dt.(29)E_{\mathrm{cyc}} = \sqrt{\frac{1}{2|\mathcal{K}_{\mathrm{c}}|} \sum_{k \in\mathcal{K}_{\mathrm{c}}} \lVert e_k \rVert_2^2}, \qquad\tau_{\mathrm{rms}} = \sqrt{\frac{1}{2KT_{\mathrm{s}}} \int_{0}^{KT_{\mathrm{s}}} \lVert\tau^{\mathrm{act}}(t) \rVert_2^2 \,\mathrm{d}t}. \tag*{(29)}

The factor two averages over the joints. Torque effort includes the actual held torque after actuator modification. A trial stops when either position error exceeds 2 rad, either speed exceeds 10 rad s−110\ \mathrm{rad}\,\mathrm{s}^{-1}, or a numerical failure occurs. The terminal state is checked as well. Incomplete runs enter completion counts but not full-duration error or effort statistics.

Prediction Tests and Numerical Accuracy

For each test window, the error is ∥qpred−qtrue∥2/2N\lVert\mathbf{q}^{\mathrm{pred}} - \mathbf{q}^{\mathrm{true}} \rVert_2/\sqrt{2N}, where the vectors stack future positions. Each independently fitted predictor uses three separate 60-sample test episodes, giving 15 non-overlapping windows. Take the median over windows for each fit, then the median and range across six fits. Data are recollected for each noise/mechanics condition, and both predictors solve (22) with identical inputs and past measurements.

The dimension d=28d = 28 is retained throughout. Supporting fits with d=30d = 30 and 32 gave larger prediction errors and poorer conditioning on the same test records. A numerical-integration check at half the step size changed saved-interval positions by at most 1.42×10−9 rad1.42 \times10^{-9}\,\mathrm{rad}, with unchanged actuator paths. Optimisation settings and termination statistics are reported in Appendix § 9.§ 3.

References

  1. [1]Mohammad Alsalti, Victor G. Lopez, Julian Berberich, Frank Allgöwer, and Matthias A. Müller. Data-driven nonlinear predictive control for feedback linearizable systems. IFAC-PapersOnLine, 56(2):617–624, 2023. doi: 10.1016/j.ifacol.2023.10.1636.
  2. [2]Julian Berberich, Johannes Köhler, Matthias A. Müller, and Frank Allgöwer. Linear tracking MPC for nonlinear systems—part II: The data-driven case. IEEE Transactions on Automatic Control, 67(9):4406–4421, 2022. doi: 10.1109/TAC.2022.3166851.DOI
  3. [3]Shreyas Bharadwaj, Bamdev Mishra, Cyrus Mostajeran, Alberto Padoan, Jeremy Coulson, and Ravi N. Banavar. Robust least-squares optimization for data-driven predictive control: A geometric approach. In Gaurav Sukhatme, Lars Lindemann, Stephen Tu, Adam Wierman, and Nikolay Atanasov, editors, Proceedings of The 8th Annual Learning for Dynamics and Control Conference, volume 331 of Proceedings of Machine Learning Research, pages 511–531. PMLR, 2026. URL https://proceedings.mlr.press/v331/bharadwaj26a.html.arxiv.org/abs/2511.09242
  4. [4]Jeremy Coulson, John Lygeros, and Florian Dörfler. Data-enabled predictive control: In the shallows of the DeePC. In 2019 18th European Control Conference (ECC), pages 307–312, 2019a. doi: 10.23919/ECC.2019.8795639.DOI
  5. [5]Jeremy Coulson, John Lygeros, and Florian Dörfler. Regularized and distributionally robust data-enabled predictive control. In 2019 IEEE 58th Conference on Decision and Control (CDC), pages 2696–2701, 2019b. doi: 10.1109/CDC40024.2019.9028943.DOI
  6. [6]Ezzat Elokda, Jeremy Coulson, Paul N. Beuchat, John Lygeros, and Florian Dörfler. Data-enabled predictive control for quadcopters. International Journal of Robust and Nonlinear Control, 31(18):8916–8936, 2021. doi: 10.1002/rnc.5686.DOI
  7. [8]Lukas Hewing, Kim P. Wabersich, Marcel Menner, and Melanie N. Zeilinger. Learning-based model predictive control: Toward safe learning in control. Annual Review of Control, Robotics, and Autonomous Systems, 3:269–296, 2020b. doi: 10.1146/annurev-control-090419-075625.DOI
  8. [9]Linbin Huang, Jianzhe Zhen, John Lygeros, and Florian Dörfler. Robust data-enabled predictive control: Tractable formulations and performance guarantees. IEEE Transactions on Automatic Control, 68(5):3163–3170, 2023. doi: 10.1109/TAC.2023.3241282.DOI
  9. [10]Ivan Markovsky and Florian Dörfler. Behavioral systems theory in data-driven analysis, signal processing, and control. Annual Reviews in Control, 52:42–64, 2021. doi: 10.1016/j.arcontrol.2021.09.005.DOI
  10. [11]Ivan Markovsky and Paolo Rapisarda. Data-driven simulation and control. International Journal of Control, 81(12):1946–1959, 2008. doi: 10.1080/00207170801942170.DOI
  11. [12]D. Q. Mayne, M. M. Seron, and S. V. Raković. Robust model predictive control of constrained linear systems with bounded disturbances. Automatica, 41(2):219–224, 2005. doi: 10.1016/j.automatica.2004.08.019.DOI
  12. [13]Jeremy D. Watson. Hybrid data-enabled predictive control: Incorporating model knowledge into the DeePC. IEEE Open Journal of Control Systems, 4:549–561, 2025. doi: 10.1109/IOCSYS.2025.3625364.arxiv.org/abs/2502.12467
  13. [14]Felix Wegner. Data-enabled predictive control of robotic systems. Master’s thesis, ETH Zürich, April 2021.DOI
  14. [15]Greg Welch and Gary Bishop. An introduction to the Kalman filter. Technical report TR 95-041, University of North Carolina at Chapel Hill, November 2000. URL https://www.cs.yale.edu/homes/hudak-paul/CS474S01/kalman.pdf. Revised edition, November 13, 2000.
  15. [16]Jan C. Willems, Paolo Rapisarda, Ivan Markovsky, and Bart L. M. De Moor. A note on persistency of excitation. Systems & Control Letters, 54(4):325–329, 2005. doi: 10.1016/j.sysconle.2004.09.003.DOI

Paper details

vlnc:W_ph2jwhu5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Generative AI tools

View statement
Looking for related papers…
Contents