Pith. sign in

REVIEW 4 major objections 6 minor 54 references

Updating both coefficients and bandwidth of a kernel disturbance model lets a quadrotor policy adapt online to wind, ground effects, and payloads, cutting tracking error by up to 70% in simulation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 11:34 UTC pith:D3UQXHZ3

load-bearing objection A real incremental contribution—online adaptation of RFF bandwidth plus coefficient—with credible but under-reported hardware results, undercut by a genuine representational gap between time-dependent disturbances and the state-only feature vector. the 4 major comments →

arxiv 2606.13842 v2 pith:D3UQXHZ3 submitted 2026-06-11 cs.RO

Efficient Domain-Adaptive Policy Learning via Kernel Representation with Application to Quadrotor Control under Non-Stationary Disturbances

Hongyu Zhou , Mingtian Tan , Vasileios Tzoumas This is my paper
classification cs.RO
keywords domain adaptationrandom Fourier featureskernel representationonline least squaresadaptive controlquadrotor controldifferentiable simulationnon-stationary disturbances
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the bottleneck in domain-adaptive control is the environment representation: it must be expressive enough to cover sim-to-real gaps offline and cheap enough to update in real time at deployment. Its answer is a random-Fourier-feature kernel model of the unknown disturbance in which both the coefficients α and the bandwidth σ are learnable, so the basis itself adapts online instead of staying fixed; the policy is trained by backpropagation through a differentiable simulator over randomized (α, σ) profiles in about 50 seconds, then deployed zero-shot with (α, σ) updated every step by online least squares. The authors report consistently lower position-tracking error than learning-based and model-based adaptive baselines in high-fidelity simulation (up to 60–70% improvement) and on a quadrotor under suspended payload, switching ground effects, and time-varying wind (up to 30% improvement). They attribute the gain to basis adaptation: ablations that freeze σ lose 10–30%, and ablations that freeze both parameters degrade far more. The paper notes as limitations that formal stability guarantees are absent, out-of-distribution kernel parameters are not handled, and bursting behavior may occur during rapid regime shifts.

Core claim

The central claim is that a control policy for a quadrotor can be made domain-adaptive to non-stationary disturbances by representing the unknown disturbance force h as a finite random-Fourier-feature approximation h(z) ≈ (1/M) Σ Φ(z, θᵢ) αᵢ, and by making both the coefficients α and the kernel bandwidth σ part of the policy's observation and updating them online through least-squares estimation at every control step. During offline training the coefficients and bandwidth are randomly sampled to generate a diverse family of disturbance profiles, and the policy is optimized with analytical gradients through the differentiable dynamics. At deployment the policy is frozen but the parameters (α,

What carries the argument

The load-bearing object is the random-Fourier-feature (RFF) kernel approximation of the disturbance, h(z) ≈ (1/M) Σ Φ(z, θᵢ) αᵢ, with Gaussian kernel features Φ(z, θ) = cos(σ wᵀ z + b); σ is the bandwidth that determines the base measure and hence the function space the features can represent. The paper's move is to treat both α and σ as learnable parameters of this approximation. Offline, differentiable simulation plus backpropagation-through-time trains the policy over randomized (α, σ); online, least-squares gradients update (α, σ) so the basis itself can adapt. The policy consumes current (α, σ) as part of its observation, which is how learned behavior reacts to the adapted disturbance m

Load-bearing premise

The method treats the entire one-step state-transition residual, x_{t+1} − f(x_t) − g(x_t)u_t, as the ground-truth disturbance h; if sensor noise, actuation delay, or unmodeled rotor dynamics corrupt that residual, the online (α, σ) estimates are biased and the adaptive advantage shrinks.

What would settle it

On a quadrotor under a known constant disturbance, add a controlled one-tick delay or white noise to the state measurement used in the residual calculation, then measure tracking error with online updates on versus off. If the estimated disturbance lags or the policy tracks worse with updates on, the residual-as-disturbance assumption is violated.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A policy can be deployed zero-shot to a real quadrotor after about 50 seconds of offline training and still adapt to disturbances not seen during training.
  • Updating kernel bandwidth along with coefficients is worth a 10–30% tracking-error reduction over a fixed basis, so basis adaptation is doing real work.
  • The same kernel representation can represent disturbances that exceed the training range of neural encoders, such as switching and quadratic-phase disturbances.
  • Online adaptation is fast enough for 50 Hz control with a small MLP policy and 25 random features.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: The 'adapt the basis' principle is not specific to quadrotors; any control-affine system with a differentiable simulator and a low-dimensional state/action space could use the same (α, σ) update, and the added online computation is one scalar bandwidth per feature.
  • Editorial extension: Since σ sets the feature frequencies, its online trajectory is a rough measure of the disturbance's dominant spatial or temporal scale; a testable prediction is that σ will grow under smooth large-scale wind and shrink under fast oscillations like rotor wash.
  • Editorial extension: The ablation pattern suggests the gain is less about kernel expressiveness in general and more about online basis adaptation; an oracle that knew the true disturbance frequency would be a stronger comparison than fixed-basis MPC.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a two-phase method for quadrotor trajectory tracking under non-stationary disturbances. Offline, an unknown disturbance is modeled with random Fourier features (RFF); coefficients α and bandwidth σ are randomized to create diverse training environments, and a neural policy is trained by backpropagation through a differentiable simulator in about 50 seconds. Online, the policy is deployed zero-shot and both α and σ are updated at every control step by one-step gradient descent on the residual between the observed next state and the nominal dynamics; the current estimates are included in the policy observation. The method is evaluated in simulation under aerodynamic effects plus sinusoidal, switching, and quadratic-phase sinusoidal disturbances, and on Crazyflie hardware under payload, wind, and ground-effect disturbances, against DATT and MPC baselines and ablations. The paper claims up to 60–70% position-tracking improvement in simulation and up to 30% on hardware.

Significance. If validated, the core claims would be significant: 50-second offline training, zero-shot sim-to-real transfer with online basis adaptation, and hardware demonstrations under several disturbance types. The ablation study is useful because removing the online α/σ updates degrades performance substantially, showing that the online mechanism is essential. The paper also states limitations honestly, including the absence of formal stability guarantees and possible bursting behavior. However, the empirical evidence is weakened by the absence of error bars or trial counts, and the representational gap between the RFF feature input and the time-dependent disturbances used in simulation undermines the interpretation that the method is learning the true disturbance. The dimensional inconsistency of the estimated residual compounds this concern. The contribution is potentially valuable but requires substantial revision before its central claims can be accepted.

major comments (4)
  1. [§4.1, §4.3, Algorithms 1–2] The RFF feature vector is defined in §4.1 as z=(R^T v_t, q_t, T_{t-1}/mg, ω_{t-1}), which contains no time or reference-phase information. However, all simulation disturbances in §4.3 are explicit functions of t (sinusoidal, switching at t=5, and quadratic-phase sinusoid). A function h(z_t) cannot represent a nontrivial function of t unless the closed-loop trajectory makes z_t injective in t; this is impossible for the periodic lemniscate reference and especially for the switching and chirp cases. Proposition 1 guarantees O(1/√M) uniform approximation only for h in F_2(B_h) over z, not for time-indexed disturbances. The paper reports no bound or experiment quantifying this representational gap. Consequently, the claimed 60–70% simulation improvement may be caused by the policy's feedforward/reference-tracking behavior or by α/σ acting as an arbitrary context code, rather than by the RFF
  2. [§2, Eq. (2), Algorithm 2 lines 5–6] Section 2 defines h∈R^3 as a disturbance force, but Algorithm 2 line 5 computes h(z_t)=x_{t+1}-f(x_t)-g(x_t)u_t, which is a full state residual (position, velocity, and quaternion components). The dimensions of f, g, and the RFF output are never reconciled: the residual is not projected onto the force subspace, and the effect of discretization/Euler integration and mass scaling is not addressed. On hardware this residual is corrupted by state-estimation noise, actuation delay, and unmodeled rotor dynamics; §4.1 only lowers the learning rate to η=0.02 and gives no noise or delay analysis. This biases the estimated α and σ and undermines the claim that the method estimates the physical disturbance. Please state the exact discrete dynamics used, specify the dimension of Φ, and provide a noise/delay analysis or a hardware validation of the residual estimator.
  3. [Tables 1–2, §4.3–4.4, Appendix D] The central empirical claim is based on a single average RMSE value per condition in Tables 1 and 2, and Appendix D shows one sample trajectory per hardware condition. No error bars, standard deviations, number of trials or random seeds, or statistical tests are reported. The hardware improvements over the best baseline are in the range of roughly 3–34%, and the ablation differences are 3–10%; without variability information these differences may not be statistically meaningful. Please report per-trial results, at least as mean±std, for all simulation and hardware experiments.
  4. [§3.3, Proposition 1] The online update in Algorithm 2 is gradient descent on the non-convex loss l_t(σ,α). The cited online-convex-optimization guarantee [27] applies to convex losses; the paper provides no convergence, regret, or stability analysis for the joint σ/α update. Proposition 1 is a uniform approximation result for a fixed base measure and does not cover σ adaptation or time-varying disturbances. If the contribution is intended to be empirical, this should be stated explicitly; if theoretical support is claimed, the missing analysis is a load-bearing gap.
minor comments (6)
  1. [§4.1] The policy observation is stated to have size 187, but the listed components sum to far less: body-frame velocity (3) + quaternion (4) + previous thrust (1) + previous body-rate command (3) + 10 future reference points (30 if 3D) + α (75) + σ (1) = 117. Please provide an itemized count of the observation vector.
  2. [§3.1, Example 1] The Gaussian-kernel RFF formula is not standard as written: with w~N(0,I) and feature cos(σ w^T z + b), the associated kernel is proportional to exp(-σ²||x-y||²/2), not exp(-||x-y||²/(2σ²)). Clarify the convention for σ, since the paper calls it the bandwidth.
  3. [§3.3, Algorithm 2] The method is called 'online least-squares estimation' but the update is a gradient-descent step on the current one-sample loss. Please align the terminology with the actual computation.
  4. [Algorithms 1–2] The algorithms project onto sets D_σ and D_α, but these projection bounds are not defined in §4.1; only the training sampling ranges for σ and α are given. Specify the online projection sets.
  5. [§4.4, Appendix D] The hardware protocol omits several details needed for reproducibility: how the wind disturbance was generated, the number of flights per condition, the state-estimation source, and how the ground-effect condition was calibrated. Please add these.
  6. [Throughout] There are several typographical errors, e.g., 'dynamcis' (§3.2), 'distrubance' (§4.3), 'restuls' (§4.4), 'trian' (§4.1). A careful proofread is needed.

Circularity Check

0 steps flagged

No significant circularity: the online RFF parameters are fit to state-transition residuals, while the reported tracking improvements are measured independently against reference trajectories.

full rationale

I walked the derivation and experiment chain and found no load-bearing step that reduces, by the paper's own equations or by self-citation, to its inputs. The disturbance estimate in Algorithm 2 is computed as the one-step state-transition residual h(z_t)=x_{t+1}-f(x_t)-g(x_t)u_t and the online least-squares updates of alpha and sigma minimize ||h(z_t)-hat_h(z_t)||^2, i.e. the estimation loss is defined on the residual, not on the tracking error. The reported metric (position RMSE against a reference trajectory) is therefore not the quantity being fitted. The offline training samples alpha and sigma from randomized distributions and trains the policy by BPTT; this is a standard differentiable-simulation pipeline, not a fit to the evaluation trajectories. Proposition 1 is a Monte-Carlo RFF approximation bound stated for the function class F_2(B_h) and cited from external work [20]; it does not by itself assert that the deployed disturbances lie in that class, so any time-dependence mismatch is a correctness/representational concern rather than a circular reduction. The paper cites several prior works by the same authors, but these are used for related-work context and for the MPC(RFF) baseline; the strongest improvements are also established against external DATT baselines, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the central claim. Accordingly, the derivation is self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 8 axioms · 0 invented entities

The algorithm is an empirical control pipeline; almost every performance-relevant number is chosen by hand rather than derived. The most load-bearing are the RFF feature count M, the training ranges for alpha and sigma, the online learning rates, and the reward weights. The axioms show the main unproved premises: the true disturbances must lie in the RKHS ball defined by the chosen kernel; the one-step residual in Algorithm 2 equals the true disturbance; the offline randomization covers deployment; and the non-convex online gradient updates converge fast enough without a stability guarantee.

free parameters (8)
  • Number of random Fourier features M = 25
    Controls approximation error O(1/sqrt(M)) and online update cost; no ablation reported.
  • Offline bandwidth sampling range D_sigma = [0.001, 1]
    Hand-chosen; online sigma is projected back into this range (Algorithm 2 line 9), so the policy never sees sigma outside this range.
  • Offline coefficient distribution for alpha = N(0, 1/2 I)
    Hand-chosen; defines the disturbance diversity and magnitude seen in offline training.
  • Online learning rate eta = 0.1 (simulation) / 0.02 (hardware)
    Tuned per domain; hardware uses smaller rate because residuals are noisy.
  • Policy architecture (3-layer MLP, hidden 512) = 512 hidden units
    Chosen without architecture ablation.
  • Reward weights = sim (2.5,0.5,0.1,0.01); hw (2.5,0.5,0.1,0.1)
    Hand-tuned; hardware penalizes control effort more to avoid aggressive behavior.
  • BPTT training schedule = 300 epochs, 500 envs, 250 rollout steps, dt=0.02s
    Chosen to fit the claimed 50-second training time on an RTX 4090.
  • Online initialization of alpha and sigma = alpha=0, sigma=0.5
    Start point for online adaptation; affects the early transient before estimates converge.
axioms (8)
  • domain assumption The unknown disturbance h lies in F_2(B_h) = {∫ Φν(·,θ)α(θ)dν(θ) : ‖α‖ ≤ B_h} for the chosen base measure ν.
    Section 3.1 assumes this to apply Proposition 1's O(1/sqrt(M)) error bound.
  • standard math Operator-valued Bochner's theorem and Proposition 1 of Boffi et al. [20] hold for the Gaussian RFF construction used here.
    Justifies that M random features approximate the RKHS function within O(1/sqrt(M)).
  • domain assumption Quadrotor dynamics are exactly control-affine per eq. (1): x_dot = f(x) + g(x)u + h with known f,g and input (T, omega).
    Used for BPTT gradients and for computing residuals in Algorithm 2.
  • domain assumption The feature vector z=(R^T v, q, T/mg, omega) is sufficient to represent wind, ground effect, payload swing, and aerodynamic disturbances.
    Ground effect depends strongly on height, which is not directly in z; the model must infer it from thrust, attitude, and velocity.
  • domain assumption The one-step residual x_{t+1}-f(x_t)-g(x_t)u_t equals the true disturbance h(z_t) with negligible noise, delay, and unmodeled dynamics.
    Algorithm 2 lines 4-5; on hardware this is approximate, and only a smaller learning rate is used to mitigate it.
  • domain assumption Deployment disturbances are covered by the offline randomization alpha~N(0,1/2 I) and sigma in [0.001,1], and online estimates stay in these sets due to projection.
    Section 5 states that out-of-distribution values are not handled.
  • ad hoc to paper Online gradient descent on the non-convex loss l_t converges sufficiently fast for real-time adaptation.
    Algorithm 2 uses gradient updates on a non-convex objective; no convergence proof is given, and bursting is acknowledged in Section 5.
  • domain assumption The differentiable simulator used for BPTT matches hardware dynamics well enough for zero-shot transfer.
    The policy is trained entirely in simulation; transfer depends on sim-to-real fidelity.

pith-pipeline@v1.3.0-alltime-deepseek · 15214 in / 17154 out tokens · 167774 ms · 2026-08-02T11:34:37.405089+00:00 · methodology

0 comments
read the original abstract

We present an algorithm for efficient domain-adaptive policy learning via kernel representations. Learning domain-adaptive policies is challenging since it requires an environment representation that is both sufficiently expressive to model complex sim-to-real gaps during offline training, and computationally efficient enough to support rapid online adaptation during deployment. For instance, a quadrotor may encounter time-varying, non-stationary disturbances, such as sudden gusts of wind, payload shifts, or transitions between distinct flight regimes with and without ground effects. To address these challenges, we model unknown disturbances using a differentiable kernel approximation based on random Fourier features. During the offline training phase, we randomly sample kernel coefficients and bandwidth parameters to generate a rich diversity of disturbance profiles. We then optimize the control policy via differentiable simulation with analytical gradients, a process that takes only 50 seconds of training time on an RTX 4090 GPU. During hardware deployment, the policy adapts to non-stationary environments in real time by updating both the kernel coefficients and bandwidth through online least-squares estimation. We evaluate our method on quadrotor trajectory tracking tasks across high-fidelity numerical simulations and hardware experiments using Crazyflie, subjected to various disturbances, including complex aerodynamic effects, wind, ground effects, and payload fluctuations.

Figures

Figures reproduced from arXiv: 2606.13842 by Hongyu Zhou, Mingtian Tan, Vasileios Tzoumas.

Figure 1
Figure 1. Figure 1: Efficient Domain-Adaptive Policy Learning via Kernel Representation for Quadrotor Control under Non-Stationary Disturbances. We leverage kernel representations for efficient domain-adaptive policy learning that achieves zero-shot sim-2-real transfer for quadrotor control under non-stationary disturbances: (a) unknown suspended payload that weighs 20% of drone mass and may swing, (b) on-and-off-switching gr… view at source ↗
Figure 2
Figure 2. Figure 2: Efficient Domain-Adaptive Policy Learning via Kernel Representation. The frame￾work is composed of offline and online phases. During the offline training phase, we randomly sample kernel coefficients and bandwidth parameters to generate a rich diversity of disturbance pro￾files. We then optimize the control policy via differentiable simulation with analytical gradients, a process that takes only 50 seconds… view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of σ Under Different Disturbance in Section 4.3. The value of σ lie within the range of [0.001, 1] used in training [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Evolution of α Under Different Disturbance in Section 4.3. All 75 elements of α are shown and each curve represents an averaged element of α. All elements of α lie within 3 times standard deviation of the Gaussian distribution N (0, 1 2 I) used in training. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results of Aerodynamic Secnario in the Numerical Simulations in Section 4.3. Top: Tracking error (m), with the inserted plot showing RMSE (m). Middle: Norm of residual predic￾tion error (m/s2 ), with the inserted plot showing the average error (m/s2 ). Bottom: Ground-truth disturbances (m/s2 ). 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results of Aerodynamic + Sinusoidal Secnario in the Numerical Simulations in Sec￾tion 4.3. Top: Tracking error (m), with the inserted plot showing RMSE (m). Middle: Norm of residual prediction error (m/s2 ), with the inserted plot showing the average error (m/s2 ). Bottom: Ground-truth disturbances (m/s2 ). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results of Aerodynamic + Switching Secnario in the Numerical Simulations in Sec￾tion 4.3. Top: Tracking error (m), with the inserted plot showing RMSE (m). Middle: Norm of residual prediction error (m/s2 ), with the inserted plot showing the average error (m/s2 ). Bottom: Ground-truth disturbances (m/s2 ). 17 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Results of Aerodynamic + Quadratic-Phase Sinusoidal Secnario in the Numerical Simulations in Section 4.3. Top: Tracking error (m), with the inserted plot showing RMSE (m). Middle: Norm of residual prediction error (m/s2 ), with the inserted plot showing the average er￾ror (m/s2 ). Bottom: Ground-truth disturbances (m/s2 ). 18 [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sample Trajectory of Hardware Experiment under Ground Effect in Section 4.4. Each subplot shows the x−, y−, z− position and the reference trajectory (dashed) over a 20s eval￾uation window. Our method achieves the best tracking performance in RMSE. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sample Trajectory of Hardware Experiment under Wind Disturbance in Sec￾tion 4.4. Each subplot shows the x−, y−, z− position and the reference trajectory (dashed) over a 20s evaluation window. Our method achieves the best tracking performance in RMSE. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Sample Trajectory of Hardware Experiment under Suspended Payload in Sec￾tion 4.4. Each subplot shows the x−, y−, z− position and the reference trajectory (dashed) over a 20s evaluation window. Our method achieves the best tracking performance in RMSE. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 9 linked inside Pith

  1. [1]

    Zhou and J

    K. Zhou and J. C. Doyle.Essentials of robust control, volume 104. Prentice hall Upper Saddle River, NJ, 1998

    1998

  2. [2]

    D. Q. Mayne, M. M. Seron, and S. Rakovi ´c. Robust model predictive control of constrained linear systems with bounded disturbances.Automatica, 41(2):219–224, 2005

    2005

  3. [3]

    Zhou and V

    H. Zhou and V . Tzoumas. Safe control of partially-observed linear time-varying systems with minimal worst-case dynamic regret. In2023 62nd IEEE Conference on Decision and Control (CDC), pages 8781–8787. IEEE, 2023

    2023

  4. [4]

    Martin, L

    A. Martin, L. Furieri, F. Dörfler, J. Lygeros, and G. Ferrari-Trecate. On the guarantees of minimizing regret in receding horizon.IEEE Transactions on Automatic Control, 2024

    2024

  5. [5]

    Liu and P

    J. Liu and P. Seiler. Robust regret optimal control.International Journal of Robust and Non- linear Control, 34(7):4532–4553, 2024

    2024

  6. [6]

    J.-J. E. Slotine. Applied nonlinear control.PRENTICE-HALL google schola, 2:1123–1131, 1991

    1991

  7. [7]

    Krstic, P

    M. Krstic, P. V . Kokotovic, and I. Kanellakopoulos.Nonlinear and adaptive control design. John Wiley & Sons, Inc., 1995

    1995

  8. [8]

    P. A. Ioannou and J. Sun.Robust adaptive control, volume 1. PTR Prentice-Hall Upper Saddle River, NJ, 1996

    1996

  9. [9]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, and J. Malik. Rma: Rapid motor adaptation for legged robots. arXiv preprint arXiv:2107.04034, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    T. Xue, A. Razmjoo, S. Shetty, and S. Calinon. Robust contact-rich manipulation through implicit motor adaptation.The International Journal of Robotics Research, 45(1):41–59, 2026

    2026

  11. [11]

    Zhang, A

    D. Zhang, A. Loquercio, J. Tang, T.-H. Wang, J. Malik, and M. W. Mueller. A learning-based quadcopter controller with extreme adaptation.IEEE Transactions on Robotics, 2025

    2025

  12. [12]

    Huang, R

    K. Huang, R. Rana, A. Spitzer, G. Shi, and B. Boots. Datt: Deep adaptive trajectory tracking for quadrotor control.arXiv preprint arXiv:2310.09053, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Hazan and K

    E. Hazan and K. Singh. Introduction to online nonstochastic control.arXiv preprint arXiv:2211.09619, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Agarwal, B

    N. Agarwal, B. Bullins, E. Hazan, S. Kakade, and K. Singh. Online control with adversar- ial disturbances. InInternational Conference on Machine Learning (ICML), pages 111–119, 2019

    2019

  15. [15]

    Zhao, Y .-X

    P. Zhao, Y .-X. Wang, and Z.-H. Zhou. Non-stationary online learning with memory and non- stochastic control. InInternational Conference on Artificial Intelligence and Statistics (AIS- TATS), pages 2101–2133. PMLR, 2022

    2022

  16. [16]

    Zhou and V

    H. Zhou and V . Tzoumas. Safe non-stochastic control of linear dynamical systems. In2023 62nd IEEE Conference on Decision and Control (CDC), pages 5033–5038. IEEE, 2023

    2023

  17. [17]

    Rahimi and B

    A. Rahimi and B. Recht. Random features for large-scale kernel machines.Advances in neural information processing systems, 20, 2007

    2007

  18. [18]

    Rahimi and B

    A. Rahimi and B. Recht. Uniform approximation of functions with random bases. In2008 46th annual allerton conference on communication, control, and computing, pages 555–561. IEEE, 2008. 9

    2008

  19. [19]

    Zhou and V

    H. Zhou and V . Tzoumas. Simultaneous system identification and model predictive control with no dynamic regret.IEEE Transactions on Robotics, 2025

    2025

  20. [20]

    N. M. Boffi, S. Tu, and J.-J. E. Slotine. Nonparametric adaptive control and prediction: Theory and randomized algorithms.Journal of Machine Learning Research, 23(281):1–46, 2022

    2022

  21. [21]

    Brault, M

    R. Brault, M. Heinonen, and F. Buc. Random fourier features for operator-valued kernels. In Asian Conference on Machine Learning, pages 110–125. PMLR, 2016

    2016

  22. [22]

    H. Q. Minh. Operator-valued bochner theorem, fourier feature maps for operator-valued ker- nels, and vector-valued learning.arXiv preprint arXiv:1608.05639, 2016

    Pith/arXiv arXiv 2016

  23. [23]

    Ghorbani, S

    B. Ghorbani, S. Mei, T. Misiakiewicz, and A. Montanari. Linearized two-layers neural net- works in high dimension. 2021

    2021

  24. [24]

    Jacot, F

    A. Jacot, F. Gabriel, and C. Hongler. Neural tangent kernel: Convergence and generalization in neural networks.Advances in neural information processing systems, 31, 2018

    2018

  25. [25]

    F. Bach. Breaking the curse of dimensionality with convex neural networks.Journal of Ma- chine Learning Research, 18(19):1–53, 2017

    2017

  26. [26]

    L. Metz, C. D. Freeman, S. S. Schoenholz, and T. Kachman. Gradients are not all you need. arXiv preprint arXiv:2111.05803, 2021

    Pith/arXiv arXiv 2021

  27. [27]

    Hazan et al

    E. Hazan et al. Introduction to online convex optimization.Foundations and Trends in Opti- mization, 2(3-4):157–325, 2016

    2016

  28. [28]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  29. [29]

    Z. Wu, S. Cheng, P. Zhao, A. Gahlawat, K. A. Ackerman, A. Lakshmanan, C. Yang, J. Yu, and N. Hovakimyan. L1 quad: L1 adaptive augmentation of geometric control for agile quadrotors with performance guarantees.arXiv preprint arXiv:2302.07208, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    J. Pan, J. Xing, R. Reiter, Y . Zhai, E. Aljalbout, and D. Scaramuzza. Learning on the fly: Rapid policy adaptation via differentiable simulation.IEEE Robotics and Automation Letters, 2026

    2026

  31. [31]

    J. Heeg, Y . Song, and D. Scaramuzza. Learning quadrotor control from visual features using differentiable simulation. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 4033–4039. IEEE, 2025

    2025

  32. [32]

    B. D. Anderson. Adaptive systems, lack of persistency of excitation and bursting phenomena. Automatica, 21(3):247–258, 1985

    1985

  33. [33]

    Tal and S

    E. Tal and S. Karaman. Accurate tracking of aggressive quadrotor trajectories using incre- mental nonlinear dynamic inversion and differential flatness.IEEE Transactions on Control Systems Technology, 29(3):1203–1218, 2020

    2020

  34. [34]

    R. Tao, P. Zhao, I. Kolmanovsky, and N. Hovakimyan. Robust adaptive mpc using uncertainty compensation. In2024 American Control Conference (ACC), pages 1873–1878. IEEE, 2024

    2024

  35. [35]

    Das and J

    E. Das and J. W. Burdick. Robust control barrier functions using uncertainty estimation with application to mobile robots.arXiv preprint arXiv:2401.01881, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Hanover, P

    D. Hanover, P. Foehn, S. Sun, E. Kaufmann, and D. Scaramuzza. Performance, precision, and payloads: Adaptive nonlinear mpc for quadrotors.IEEE Robotics and Automation Letters, 7 (2):690–697, 2021

    2021

  37. [37]

    N. M. Boffi, S. Tu, and J.-J. E. Slotine. Regret bounds for adaptive nonlinear control. In Learning for Dynamics and Control, pages 471–483. PMLR, 2021. 10

    2021

  38. [38]

    J. Jia, W. Zhang, K. Guo, J. Wang, X. Yu, Y . Shi, and L. Guo. Evolver: Online learning and prediction of disturbances for robot control.IEEE Transactions on Robotics, 2023

    2023

  39. [39]

    Sánchez-Sánchez and D

    C. Sánchez-Sánchez and D. Izzo. Real-time optimal control via deep neural networks: study on landing problems.Journal of Guidance, Control, and Dynamics, 41(5):1122–1135, 2018

    2018

  40. [40]

    Tobin, R

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 23–30. IEEE, 2017

    2017

  41. [41]

    Ramos, R

    F. Ramos, R. Possas, and D. Fox. Bayessim: Adaptive domain randomization via proba- bilistic inference for robotics simulators. InProceedings of Robotics: Science and Systems, FreiburgimBreisgau, Germany, June 2019. doi:10.15607/RSS.2019.XV .029

    doi:10.15607/rss.2019.xv 2019

  42. [42]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomo- tion over challenging terrain.Science robotics, 5(47):eabc5986, 2020

    2020

  43. [43]

    Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath. Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control.The International Journal of Robotics Research, 44(5):840–888, 2025

    2025

  44. [44]

    Mohri, A

    M. Mohri, A. Rostamizadeh, and A. Talwalkar.Foundations of machine learning. MIT press, 2018

    2018

  45. [45]

    Y . S. Abu-Mostafa, M. Magdon-Ismail, and H.-T. Lin.Learning from data, volume 4. AML- Book New York, 2012

    2012

  46. [46]

    Farahani, S

    A. Farahani, S. V oghoei, K. Rasheed, and H. R. Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877–894, 2021

    2020

  47. [47]

    Ben-David, J

    S. Ben-David, J. Blitzer, K. Crammer, and F. Pereira. Analysis of representations for domain adaptation.Advances in neural information processing systems, 19, 2006

    2006

  48. [48]

    P. Wu, A. Escontrela, D. Hafner, P. Abbeel, and K. Goldberg. Daydreamer: World models for physical robot learning. InConference on robot learning, pages 2226–2240. PMLR, 2023

    2023

  49. [49]

    F. Nan, H. Ma, Q. Guan, J. Hughes, M. Muehlebach, and M. Hutter. Efficient model-based reinforcement learning for robot control via online learning.arXiv preprint arXiv:2510.18518, 2025

    Pith/arXiv arXiv 2025

  50. [50]

    H. Zhou, Y . Song, and V . Tzoumas. Safe non-stochastic control of control-affine systems: An online convex optimization approach.IEEE Robotics and Automation Letters, 2023

    2023

  51. [51]

    H. Zhou, Z. Xu, and V . Tzoumas. Efficient online learning with memory via frank-wolfe optimization: Algorithms with bounded dynamic regret and applications to control. In2023 62nd IEEE Conference on Decision and Control (CDC), pages 8266–8273. IEEE, 2023

    2023

  52. [52]

    Tsiamis, A

    A. Tsiamis, A. Karapetyan, Y . Li, E. C. Balta, and J. Lygeros. Predictive linear online track- ing for unknown targets. InProceedings of the 41st International Conference on Machine Learning, pages 48657–48694, 2024

    2024

  53. [53]

    H. Zhou, X. Zhang, and V . Tzoumas. Adaptive legged locomotion via online learning for model predictive control.arXiv preprint arXiv:2510.15626, 2025

    arXiv 2025

  54. [54]

    Llanes, Z

    C. Llanes, Z. Kakish, K. Williams, and S. Coogan. Crazysim: A software-in-the-loop simulator for the crazyflie nano quadrotor. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12248–12254, 2024. doi:10.1109/ICRA57147.2024.10610906. 11 A Related Works We discuss work on robust and adaptive control, learning-based control, and o...

    arXiv 2024