acceptodds
Under review as a conference paper at ICLR 2027

Learning to Draw on the Curved Surface: A Riemannian Adaptive Moment Iterative Learning Control Approach

Abstract

Precise trajectory tracking on curved surfaces is essential in robotic operations such as drawing, polishing, and manufacturing. While classical Euclidean iterative learning control (ILC) improves repeatable tracking accuracy through trial-to-trial error correction, it ignores the underlying surface geometry, often producing command updates that violate geometric feasibility. This paper proposes R-ILC, a Riemannian adaptive moment ILC framework that embeds surface constraints directly into the command update rule via Riemannian optimization. Theoretical analysis establishes almost-sure convergence of the Riemannian gradient norm to zero, with avoidance of strict saddles. Experiments on curved-surface robotic drawing validate the effectiveness of our approach while demonstrating its advantages over classical Euclidean ILC in tracking accuracy and geometric feasibility.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Iterative Learning Control, Manifold Constraint, Manipulator

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_o7AHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work applies Riemannian optimization with adaptive moments to iterative learning control, where the goal is to satisfy the surface constraints by keeping the input while learning to improve the tracking error by incorporating past tracking errors. The authors propose to use an adaptive Riemannian method with momentum (Algorithm 1) applied to the input product manifold obtained from the system output being conditioned on staying on a fixed surface. Further the work established theoretical guarantees on the convergence of iterate-gradient of the proposed method (Theorem 1) and avoidance of saddle points (Theorem 2) under certain technical assumptions. Finally the authors provide experimental results in terms of simulations and a physical (real world) experiment to support their learning control method.

Critical strengths

None

Critical weaknesses
  1. Main issue: The authors assume in (2.4) to have access to a ideal (product) manifold M\mathcal{M} of inputs uu which guarantee that the outputs yy stay within a prescribed surface Sy\mathcal{S}_y. However, the entire reason to use ILC (as opposed to feedforward control) in practice is precisely because one does not have an accurate (enough) model of the system. The authors do not account for this possibility in their theory and experiments (both simulated and physical), i.e. they try to correct a tracking error mismatch due to imperfect information, but assume perfect information (by assuming access to M\mathcal{M}, its tangent spaces, etc). The simulated experiments use this ideal manifold, while the physical experiment does not show the true off-surface error.
    The reviewer cannot think of a real situation where M\mathcal{M} would be perfectly available, but a system model is too inaccurate to track a reference via feedforward control (solved offline) and plant inversion techniques, nor do the authors provide such an example. The physical system in this paper is not such an example, since we do not know M\mathcal{M} for the real world SO101 robot and surface (as opposed to its Mujoco model).

  2. The core contribution of this paper is a straightforward application of existing results in Riemannian Adaptive Optimization Methods to the well-established area of iterative learning control or a slight extension of these. While an application of one domain to another can lead to fruitful synergies, it is not the case here due to first issue (1) for this paper. In particular, the adaptive algorithm that the authors use was already proposed in [1] (cited by the paper itself) and some theoretical contributions (Theorem 1) have been already established for the Euclidean case in the work [2]. E.g. in [2, Theorem 11, Theorem 9] the Euclidean version of Theorem 1 was shown under a similar assumption [2, Assumption 1] as in this work. The Theorem 1 in this work seems to be a Riemannian extension of these due to the fact that scalar factors commute with parallel transport (see formula on top of page 17). The current work does not separate the technical novelty from [1, 2] sufficiently well.

  3. The reviewer fails to see how Theorem 1 and Theorem 2 are relevant for ILC. While avoidance of saddle points is important for optimization, for ILC it is crucial that the input sequence uu converges itself, which Theorem 1 does not establish. Therefore it is unclear in how far the current theory supports the experimental observations.

  4. The experiments are not sufficiently diverse and informative: They consist entirely of a single system with the same surface constraint. While this is understandable for the physical robot (where experiments are time-consuming and costly), the authors should have provided more diverse simulated situations (e.g. different robots and surfaces) and additionally distinguish between the input manifold M\mathcal{M} for the true system and the system model (see core issue (1)) for both simulation and physical experiment. Further, the physical experiment lacks the plot of the crucial off-surface error (which has to be measured somehow).

Questions for authors

Please see questions raised in "weaknesses"

1 comment

Thank you for asking what R-ILC needs to know about the plant.

Response to Weakness 1. Knowing M\mathcal M and its tangent spaces with an imperfect plant model.

Eq. (2.4) asks for inputs known to keep nominal outputs on the surface. A cylindrical guide fixes the pen radius RR mechanically. Its references (R,θ,z)(R,\theta,z) have tangent directions (0,δθ,δz)(0,\delta\theta,\delta z) for any angular servo delay dd. The pen reaches θt−d\theta_{t-d} at time tt. Two drives can share the exact manifold and tangents yet need different command advances to track on time. The measured lag tells us how far to advance the next stroke. On a guided CNC axis, Wang and Hsiao [4] used motor and table measurements to learn positioning corrections.

For our arm, nominal pen kinematics and the ellipsoid define M\mathcal M (Appendix E.1). If the modeled patch is within δS\delta_S of the workpiece and tip bias and execution noise are bounded by δF\delta_F and σ\sigma, the triangle inequality bounds true tip-to-workpiece distance by δS+δF+σ\delta_S+\delta_F+\sigma for every nominally feasible sample.

Response to Weakness 2. Technical novelty beyond [1, 2] and scalar commutation.

Transported momentum and factorwise scaling come from [1], whose regret bounds concern geodesically convex losses. Our nonconvex trial loss couples neighboring samples (Appendix E.2). Theorem 1 extends the Euclidean stationarity and rate results in the preprint of [2] (Theorems 9 and 11) to the trial-input manifold. Commutation transports the previous scale. Lemma 7 also bounds the momentum-alignment error from the changing gradient along the update geodesic by O(ηk−1)O(\eta_{k-1}). This and the scale-change error O(1−θk)O(1-\theta_k) (Lemma 5) are summable after weighting by ηk\eta_k.

For Theorem 2, [3] gives a Riemannian precedent for saddle avoidance. Lemma 12 absorbs the leading momentum term into an auxiliary iterate. Appendix D verifies that under Assumption 3 the noise excites the unstable subspace of the joint recursion for this iterate and the second moment.

Response to Weakness 3. Theorems 1 and 2, input convergence in ILC, and the trial data.

A single limiting input requires additional structure. Lemma 3 gives d(uk+1,uk)≤Bstepηk→0d(u_{k+1},u_k)\le B_{\rm step}\eta_k\to0. Theorem 1 drives the gradient norm to zero, so even the best feasible direction offers vanishing first-order gain in expected tracking loss. If the iterates stay in a compact set with continuous gradient and finitely many stationary inputs, these facts imply a single limit. Under its extra assumptions, Theorem 2 gives zero probability of settling at any member of a prescribed finite or countable set of strict saddles.

The recorded inputs settled as tracking improved. For the triangle, RMS change per joint and sample fell from 0.0329 rad (trials 1–2) to 0.000284 rad (9–10), while tracking RMSE fell from 20.24 to 0.106 mm. At all three starts the last input change was below 1% of the first.

Response to Weakness 4. Experimental breadth, model versus plant, and physical off-surface error.

We varied initialization on one SO-ARM101 and one ellipsoid. For the triangle, circle and scaled heart, final RMSE was 0.106, 0.0593 and 0.0499 mm with R-ILC versus 1.047, 0.812 and 0.754 mm with PD-ILC. Transfer to other robots or surfaces remains open. The same simulated kinematics define the manifold and evaluate the tip, so Pk<2.6×10−12P_k<2.6\times10^{-12} m over 30 trials measures nominal surface deviation.

On hardware the arm replayed the simulation-learned input as a fixed reference. Figure 3A shows the resulting heart. A physical off-surface plot requires 3-D tip positions registered to the workpiece throughout the stroke. Only the simulations recorded those positions.

We offset only the simulated base. From shared target IK, nine updates cut tracking RMSE EkE_k by about 76% at each offset. RMS shortest distance DkD_k to the fixed workpiece fell by 23.6–28.1%, to 0.155–0.650 mm at stroke 10.

Base translation / rotationIK E1E_1 (mm)R-ILC E10E_{10} (mm)IK D1D_1 (mm)R-ILC D10D_{10} (mm)
0.5 mm / 0.1°0.9645940.2311403510.2156288530.155127458
1.0 mm / 0.2°1.9286340.4625207730.4317763390.329947523
2.0 mm / 0.4°3.8550490.9253908770.8665618340.650337688

References

[1] Bécigneul and Ganea. Riemannian Adaptive Optimization Methods. ICLR, 2019.

[2] Liang et al. Machine Learning, 2025. Published article. 2023 version.

[3] Hsieh et al. Riemannian stochastic optimization methods avoid strict saddle points. NeurIPS, 2023.

[4] Wang and Hsiao. Sensors, 2024. Feed-drive experiments, §5.

Reviewer_5hCFOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1(−1)
Presentation
1(−1)
Soundness
2
Read review
Paper summary

The paper presents a method called R-ILC, a Riemannian adaptive moment ILC framework. It optimizes repeated robot command trajectories on a manifold that enforces nominal surface feasibility. It combines transported momentum, factorwise adaptive scaling, and geometric updates. The paper establishes conditional gradient-convergence and strict-saddle nonconvergence results, then evaluates curved-surface drawing in simulation and demonstrates physical execution of a simulation-learned trajectory.

Critical strengths
  1. The geometric formulation is reasonable. Preserving feasibility during optimization is valuable for surface-following tasks, and transporting momentum between tangent spaces is mathematically appropriate.

  2. This paper gives us theoretical development. The paper explicitly states assumptions and provides extensive arguments for convergence and avoidance of a prescribed finite or countable set of strict saddles.

  3. THe results are strong within the tested setting. The main simulation achieves approximately 0.106 mm final tracking RMSE, with feasibility maintained to numerical tolerance. Three initializations and equal-sized hyperparameter searches are reported.

Critical weaknesses
  1. There are strong assumptions for practical deployment for R-ILC, and the discussion of sim-to-real gaps is insufficent. Specifically, R-ILC assumes reliable Riemannian gradient estimation at every step, which requires analytic Jacobians from robot kinematics and perfectly known manifold geometry. In real industrial curved-surface tasks, workpiece geometry is often contaminated by modeling errors and noise. All simulations and hardware experiments in this paper use an analytic ellipsoidal surface with known ground-truth geometry. There is no robustness test under surface-model perturbations or parametric errors. If the geometric model is inaccurate, the underlying manifold definition becomes invalid, and the entire Riemannian update scheme degrades. This is a critical practical limitation that is barely discussed.

  2. The theorems rely on strong mathematical prerequisites including smoothness, bounded noise and well-defined normal neighborhoods. Real robots exhibit friction, hysteresis and non-smooth contact, which violate these assumptions. The paper does not analyze how the algorithm degrades when these theoretical conditions are violated in practice.

Questions for authors

It would be better to analyze failure modes. This may make the paper stronger.

1 comment

Thank you for distinguishing accuracy on the modeled ellipsoid from reliable motion on a real workpiece. Your friction and hysteresis examples helped us pin down when a previous stroke can guide the next one.

Response to Weakness 1. How do geometry and Jacobian errors affect the manifold, the trial gradient and sim-to-real transfer?

The drawing implementation uses nominal pen kinematics and an ellipsoid to construct the input manifold, and an analytic position Jacobian for the update (Appendices E.1–E.3). An inaccurate surface fit can still define a regular manifold. The practical question is how far a step tangent to that model moves the pen in the normal direction of the real workpiece.

Let ss be a joint correction tangent to the modeled constraint, so n^⊤J^s=0\hat n^\top\hat J s=0. Here J^,J\hat J,J are the modeled and actual tip Jacobians at the same command, and n^,n\hat n,n are unit normals at corresponding modeled and physical surface points. Writing Js=J^s+(J−J^)sJs=\hat J s+(J-\hat J)s gives the following first-order bound on unintended normal motion, where α\alpha is the angle between the normals.

∣n⊤Js∣≤sin⁡(α)∥J^s∥+∥J−J^∥op∥s∥.|n^\top J s|\le \sin(\alpha)\|\hat J s\|+\|J-\hat J\|_{\rm op}\|s\|.

With an accurate Jacobian, a tilt of α\alpha thus puts at most a fraction sin⁡(α)\sin(\alpha) of the predicted motion into the true normal direction. The second term bounds the additional effect of a Jacobian error. The same error can also bias the gradient used to reduce tracking loss. Theorem 1 requires bounded, conditionally unbiased estimates of the gradient on the chosen manifold. Analytic kinematics provide the sensitivity in our implementation. Zero-mean sensor noise alone cannot remove a systematic error in that sensitivity.

The added simulation isolates a base calibration error with the workpiece fixed. A base translation shifts every tip point but cancels from the displacement between two commands. With a base rotation RR, a predicted 3-D displacement vv becomes RvRv. At the largest tested rotation of 0.4°, their difference is at most 2sin⁡(0.2∘)∥v∥2\sin(0.2^\circ)\|v\|, about 0.7% of that displacement. Local motion can therefore remain predictable even when the entire stroke is displaced. A target's distance to the displaced ellipsoid lower-bounds the tracking error attainable there with nominally feasible commands.

The original simulation used matched geometry, and Figure 3A shows the hardware replay qualitatively. In the perturbed simulation we used the original parameters and no added random noise. After nine updates from the shared target-IK stroke, tracking RMSE EkE_k was about 76% lower at each offset. The RMS shortest distance DkD_k to the fixed workpiece decreased by 23.6–28.1%. At the largest offset it fell from 0.867 to 0.650 mm. Tracking thus improved even though the executed stroke retained a measurable surface error.

Base translation / rotationIK E1E_1 (mm)R-ILC E10E_{10} (mm)IK D1D_1 (mm)R-ILC D10D_{10} (mm)
0.5 mm / 0.1°0.9645940.2311403510.2156288530.155127458
1.0 mm / 0.2°1.9286340.4625207730.4317763390.329947523
2.0 mm / 0.4°3.8550490.9253908770.8665618340.650337688

Response to Weakness 2. How can learning degrade when friction, hysteresis or contact violate the assumptions?

Repeatable friction can be corrected while the response still has the right sign and size. Consider the signed error ee in one along-path coordinate at a bend. If the tip moves by Δy\Delta y there on the next stroke, its error becomes e+Δye+\Delta y. A movement opposite to ee reduces the absolute error whenever 0<∣Δy∣<2∣e∣0<|\Delta y|<2|e|. This explains why a consistent lag can be corrected without a friction coefficient. It also gives concrete failure modes. While the pen sticks, Δy\Delta y can be nearly zero and learning stalls. After it breaks loose [1], the same command change can carry it past the target. Beyond 2∣e∣2|e| the tracking error grows, even if model-based backtracking predicts improvement.

With hysteresis the same command can trace different paths depending on the initial contact state. If that state depends on earlier trials, the conditional mean gradient may no longer match the gradient of a fixed expected loss. The bounded-gradient assumption matters at breakaway. A large spike enters the second moment through (1−θk)∥gki∥2(1-\theta_k)\|g_k^i\|^2 and can damp subsequent corrections in that block. The normal-neighborhood assumption concerns the geodesic used for the command update. A short, well-defined step on the modeled manifold can still coincide with an abrupt change in physical contact.

References

[1] Q. Le Lidec et al. “Contact Models in Robotics: A Comparative Analysis.” IEEE Transactions on Robotics (2024). Paper

Reviewer_1RRhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed an iterative learning method for constrained optimization on a manifold.

Critical strengths

This paper is complete and provides an implementable algorithm.

Critical weaknesses
  1. In this setting, using IK to compute each reference joint angle is an easier solution. Why apply the proposed method?

  2. How do you compare the current methods with a shooting method or direct trajectory optimization? Essentially, the presented method and the aforementioned methods are solving the same problem using relatively bad initial joint angles. My understanding is that iterative learning is trying to tackle the uncertainty in dynamics, while the presented work is not doing this. The presented work seems more like rollout dynamics first, then correction. In this sense, how does this method compare with the SOTA manifold optimization method?

Questions for authors

Address the concerns in the weakness part.

1 comment

Thank you for suggesting one-shot inverse kinematics (IK). It makes the value of trial feedback testable even from an accurate starting trajectory.

Response to Weakness 1. Why use R-ILC if pointwise IK can generate the reference joint angles?

In the matched kinematic setting of §4, pointwise IK is sufficient if the target admits a continuous, joint-limit-feasible branch that closes at the last point. Hitting all 200 targets at their prescribed samples also matches every displacement between successive samples, so both terms of the loss in Appendix E.2 vanish. The three starting contours in §4 test how successive updates refine a feasible input that initially traces another shape.

The reason to update an IK reference is a repeatable error in the stroke it produces. A miscalibrated base can shift the heart even when IK has zero nominal residual. Solving the same nominal IK again returns the same command. R-ILC uses the measured error and nominal sensitivity to correct the component accessible along the modeled surface. Lee et al. [1] likewise used nominal robot kinematics and measured path deviations to correct joint references on an industrial arm.

Starting from target IK, nine R-ILC updates reduced 3-D tracking RMSE EkE_k by about 76% at each simulated base offset. The IK protocol specifies a continuous, closed branch within the joint limits and a nominal tip-residual tolerance of 10−610^{-6} m. Both methods shared the first stroke. Replaying the fixed IK command would reproduce that stroke because the offset was constant and no noise was added. R-ILC used the parameters from Table E.1 without retuning. At the prespecified 1.0 mm / 0.2° offset, RMSE fell from 1.929 to 0.463 mm. The RMS shortest distance DkD_k from the tip to the fixed workpiece also fell, from 0.432 to 0.330 mm.

Base translation / rotationIK E1E_1 (mm)R-ILC E10E_{10} (mm)IK D1D_1 (mm)R-ILC D10D_{10} (mm)
0.5 mm / 0.1°0.9645940.2311403510.2156288530.155127458
1.0 mm / 0.2°1.9286340.4625207730.4317763390.329947523
2.0 mm / 0.4°3.8550490.9253908770.8665618340.650337688

Response to Weakness 2. Is R-ILC simply optimizing a poor initial trajectory by model rollout, as in shooting or direct manifold optimization?

In §4, the simulated pen position is evaluated by nominal forward kinematics with no added noise. In that setting your interpretation is fair. Shooting or direct trajectory optimization can refine a poor starting path for the same nominal objective. A shooting method evaluates candidate commands by model rollout, and a direct method such as [2] uses derivatives of discretized rigid-body dynamics in a constrained trajectory solve. When execution differs from the model, a measured stroke supplies information a nominal rollout cannot provide. Two arms with the same nominal model can miss the same target point in opposite tangential directions. At exact target IK, any update based on the nominal tracking loss alone would see zero gradient for both arms. Their measured strokes give R-ILC opposite correction directions.

In one feasible coordinate, let uu be the nominal tip position, rr the target, and yk=u+b+εky_k=u+b+\varepsilon_k the observed tip, with fixed bias bb and independent zero-mean trial noise. At u=ru=r the nominal and observed gradients of the loss 12(y−r)2\tfrac12(y-r)^2 are

gnom=0,gk=yk−r=b+εk,E[gk]=b.g_{\rm nom}=0,\qquad g_k=y_k-r=b+\varepsilon_k,\qquad \mathbb E[g_k]=b.

An individual stroke can suggest the wrong direction, yet its mean gradient retains the bias. With zero noise and zero initial moments, the first R-ILC step points opposite to bb. For a plant with dynamics, §3.1 additionally needs sensitivities relating input changes to later outputs. Algorithm 1 combines the resulting trial gradient with transported momentum on the nominal manifold. Under Theorem 1's assumptions, including bounded and conditionally unbiased gradient estimates, the gradient norm of the expected tracking loss tends to zero almost surely and in mean square.

In the calibration simulation, the hidden base offset entered only at execution. During backtracking, the update adds the previous stroke's measured chart discrepancy Zkobs−Z0(Qk)Z_k^{\rm obs}-Z_0(Q_k) to each candidate's nominal prediction. Only the accepted command is executed on the shifted model. This supplied a correction even at target IK. At the largest offset, tracking RMSE fell from 3.855 to 0.925 mm after nine such updates.

References

[1] Y.-H. Lee et al. “Industrial robot accurate trajectory generation by nested loop iterative learning control.” Mechatronics 74, 102487 (2021). DOI

[2] S. Teng et al. “Riemannian Direct Trajectory Optimization of Rigid Bodies on Matrix Lie Groups.” Robotics: Science and Systems (2025). Paper

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.