Iterative learning control (ILC) lets a robot improve at a repeated motion by replaying it. After each trial, the robot uses the full recorded tracking error to correct the pre-planned motor command, the "feedforward" signal, for the next trial. It works well: on a soft pneumatic arm, error fell from 13° to under 2° within 30 repetitions (Hofer et al.). The catch is that everything it learns is tied to one reference path. A new path starts from scratch, and the first few trials on a quadrotor or an industrial arm are the dangerous ones. The obvious fix is to reuse what was learned on other paths. Pereida et al. do this with a single linear map from desired trajectory to required input, and report a 74.12% average first-trial error reduction over 36 quadrotor transfer pairs. Their map assumes transfer always helps.
Multi-task learning has a name for what happens when that assumption fails: negative transfer. Sharing across tasks makes a given task worse than learning it alone (Wang et al.). This note argues that in ILC, negative transfer has a specific physical shape. Low-frequency parts of the feedforward, such as gravity, inertia and rigid-body motion, transfer well across paths. High-frequency parts, such as resonances, saturation and regime-specific friction, often transfer badly, and which frequencies go bad depends on the new path. We propose to transfer the learned command band by band. Each frequency band is kept or discarded according to a held-out test of whether transfer beat the nominal model in that band, for targets resembling the new one.
Core Claim
Given a library of paths on which ILC has already converged, a predicted starting command for a new path should be merged with the nominal-model command separately in each frequency band. The mixing weights depend on the new path. Let $u_{\text{nom}}$ be the command from the nominal model, $\hat u$ the command predicted by any transfer method, and $P_b$ a filter that isolates frequency band $b$. The first-trial command is then
$$u_0 = u_{\text{nom}} + \sum_{b} g_b(y_d)\, P_b\,(\hat u - u_{\text{nom}}),$$
where $g_b(y_d) \in [0,1]$ is a gate that depends on the target reference $y_d$.
The gate is trained on leave-one-path-out cases from the library. For each held-out path $j$ with converged command $u^*_j$, we ask whether transfer beat the nominal model in band $b$:
$$\|P_b(\hat u_{-j} - u^*_j)\| < \|P_b(u_{\text{nom},j} - u^*_j)\|.$$
This is the negative-transfer condition of Wang et al., evaluated one frequency band at a time. Here $\hat u_{-j}$ is the prediction made without path $j$. The gate then learns to predict this outcome from features of the target, chiefly how far the target's energy in band $b$ lies outside what the library covers.
After the first real trial, the measured error in each band gives direct evidence. A band whose error exceeds the nominal model's predicted error is reset to nominal before ILC continues.
The claim has two parts. First, negative transfer in multi-path ILC is concentrated in a target-dependent subset of frequency bands. Second, gating those bands removes most of the harm while keeping most of Pereida-style gains. The gate works on top of whichever predictor produces $\hat u$; it is not a new predictor.
Relation to Prior Work
Three lines of work motivate the proposal.
- Transfer in ILC without harm detection. Pereida et al. and Pereida et al. 2017 learn a transfer map from six seconds of flight on a single source path. They assume the map holds globally, and they have no way to detect when it hurts. We reuse their map as one candidate predictor and add the check they lack.
- The iteration-domain fix, single-path only. Balta et al. combine norm-optimal ILC with a Gaussian process fitted to model mismatch on one octagonal path (70.9% hardware improvement over plain feedforward). We use a Gaussian-process predictor of the same kind, fitted across paths, as the second candidate for $\hat u$.
- Detecting harmful sharing in machine learning.
- Wang et al. filter harmful source samples with a discriminator. Our gate filters frequency bands of a signal, and it is trained against physical outcomes rather than domain labels.
- Liu et al. reweight task losses by learning progress. They also report how many tasks end up worse than single-task training (42 of 128 with their method, over 100 with GradNorm). We borrow that count as our main harm metric.
Other work constrains the design.
- Standley et al. show that pairwise transfer affinities predict joint-training compatibility poorly. That is why we calibrate the gate on held-out outcomes for each target rather than on pairwise path similarity.
- Bonilla et al. show that multi-task Gaussian processes collapse to independent models when all tasks are observed at the same inputs without noise. ILC error profiles are sampled on the same time grid for every path, so that is exactly our situation. We therefore treat paths as inputs to one regression, not as separate Gaussian-process tasks.
- Clustering approaches assign each system to exactly one group. Examples are robust clustered identification for adaptive control (Fallah et al.) and optimization-based grouping (Towards Principled Task Grouping). Our hypothesis is that one path can share its low frequencies with every other path while sharing its high frequencies with none. Under that hypothesis, whole-path clustering is the wrong unit of sharing, and we test this directly.
Research Program
Libraries of paths. We build libraries of 10 to 30 reference paths on three testbeds, each covering several dynamical regimes.
- A simulated quadrotor that mirrors the six-path, six-second protocol of Pereida et al., extended with aggressive paths that saturate the motors.
- A dual-axis motion stage simulation with a lightly damped structural resonance. The octagon of Balta et al. is included along with circles, spirals and sharp-cornered contours at several speeds.
- A physical dual-axis stage or quadrotor if one is available.
On each path we run norm-optimal ILC to convergence to obtain $u^*_j$. These converged commands serve as ground-truth labels.
Predictors and gate.
- Two predictors produce $\hat u$: the relative-degree linear map of Pereida et al., and a Gaussian process mapping reference derivatives to the correction on top of the nominal model.
- Frequency bands come from a fixed filter bank of 6 to 8 roughly logarithmically spaced bands.
- The gate is a small logistic model per band. Its inputs are the target's energy in that band, its peak amplitude relative to the library, and the predictor's own uncertainty where the predictor provides one.
Analysis. We ask where negative transfer lives in the frequency domain. We also ask whether it grows with ILC iteration count, as Standley et al.'s assumption of fixed task relationships would miss. Finally, we ask whether a harmful starting command costs extra trials or only raises early peak error. Under the usual monotone-convergence conditions, a bad start should cost trials but not stability. We will check this empirically rather than assume it.
Baselines
Naive reference points.
- Single-path ILC started from the nominal-model command. By definition it has zero negative transfer, and it sets the floor that any transfer method must clear on every path, not just on average.
- An oracle that starts from the converged $u^*$. It sets the ceiling.
Strongest existing method. This is the Pereida et al. linear map, fitted on all library paths except the target. Its reported figure is a 74.12% average first-iteration tracking error reduction versus no transfer on the 36-pair quadrotor benchmark (Pereida et al.). We re-run it on our testbeds with identical library data, nominal model and ILC update law. The pooled Gaussian-process predictor, ungated, is run as the second strong competitor.
Cluster-then-transfer. Library paths are clustered by regime in the style of Fallah et al., and a separate map is fitted per cluster. The target is assigned to the nearest cluster. This baseline represents the grouping line of work.
Ablations isolating the key ingredient.
- A single global gate: one scalar blend between $\hat u$ and $u_{\text{nom}}$ chosen by the same leave-one-out procedure. This tests whether splitting by frequency matters.
- A fixed low-pass filter on $\hat u - u_{\text{nom}}$, with its cutoff tuned on the library. This tests whether dependence on the target matters.
- Gating without the first-trial reset. This tests whether correcting with measured error adds anything.
Fairness and success criterion. Every method gets the same library, the same nominal model and an equal tuning budget of held-out folds. Beating the baselines means two things. The gated method must cut the number of target paths where the first trial is worse than nominal, relative to the best ungated predictor. And it must retain at least 90% of that predictor's mean first-trial error reduction.
Why not just…?
Why not just run single-path ILC and skip transfer? It is safe, but the first trials are the costly ones: 13° of error before learning on the soft arm (Hofer et al.), and crash risk on quadrotors. The 74.12% first-trial reduction of Pereida et al. shows how much is left on the table. Our metrics report trials-to-tolerance and early peak error, so the value of transfer is measured, not assumed.
Why not just use a more expressive shared model, so that negative transfer disappears? This is the most serious threat. If the harm is simply misspecification of a linear map, a flexible pooled model may remove it. We test this directly by running the ungated Gaussian-process predictor. If it shows no paths worse than nominal on saturating and resonant targets, the gate is unnecessary and we will say so. We expect that it will not remove the harm, because targets outside the library's frequency coverage are an extrapolation problem, not a capacity problem. The gate is applied to both predictors so the two effects can be separated.
Why not just low-pass the transferred command, as ILC robustness filters already do? A fixed cutoff is exactly our second ablation. If a tuned fixed cutoff matches the target-dependent gate, the finding simplifies to "negative transfer in ILC is high-frequency, so filter it." That would still be worth publishing, and it would supersede our method. The target-dependent gate should only win when the harmful bands move with the target, as with resonances excited at particular speeds. Speed sweeps on the stage are designed to reveal this.
Why not cluster paths by regime and transfer within clusters? Clustering discards the shared low-frequency content between clusters. The cluster-then-transfer baseline measures how much that costs on targets that sit between regimes.
Proposed Experiments
First decisive check: localization. Using leave-one-path-out on each library, we compute per-band negative-transfer rates for both ungated predictors. The core premise fails if negative transfer is spread evenly across frequencies, or does not vary with the target.
Second decisive check: gating. We compare all methods on held-out targets. Reported quantities are:
- the fraction of targets whose first-trial RMS error exceeds nominal, following Liu et al.'s way of counting harmed tasks;
- mean and worst-case first-trial error reduction;
- peak error and saturation time over the first three trials;
- the number of ILC trials needed to reach a fixed tolerance.
Third check: robustness to the library. We repeat the comparison with libraries deliberately biased toward slow paths, then tested on aggressive paths, where extrapolation is at its worst.
On the quadrotor protocol, we check that the ungated linear map reproduces a first-trial gain comparable to the reported 74.12%. No other reported figure on Lacuna covers multi-path ILC transfer on these testbeds.
References
- Iterative Learning Control for Fast and Accurate Position Tracking with an Articulated Soft Robotic Arm
- Data-Efficient Multirobot, Multitask Transfer Learning for Trajectory Tracking
- Characterizing and Avoiding Negative Transfer
- Efficient Multi-Task and Multi-Robot Transfer with Continued Learning
- Learning-Based Repetitive Precision Motion Control with Mismatch Compensation
- Loss-Balanced Task Weighting to Reduce Negative Transfer in Multi-Task Learning
- Which Tasks Should Be Learned Together in Multi-task Learning?
- Multi-task Gaussian Process Prediction
- Adversarially Robust Multitask Adaptive Control
- Towards Principled Task Grouping for Multi-Task Learning