Pith. sign in

REVIEW 3 major objections 4 minor 270 references

A self-adaptive learning controller tracks unknown, switching target dynamics with expected regret O(T^{3/4}), so average regret vanishes.

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

2026-08-01 17:15 UTC pith:UOXZK7JX

load-bearing objection The algorithm and experiments are worthwhile, but the central no-regret theorem has a broken step in its proof, so the theoretical claim as written does not stand. the 3 major comments →

arxiv 2607.26370 v1 pith:UOXZK7JX submitted 2026-07-29 cs.RO cs.LGcs.SYeess.SY

Self-Adaptive Learning and Model Predictive Control for Tracking Unknown Dynamics with No Regret

Atharva Navsalkar , Hongyu Zhou , Vasileios Tzoumas This is my paper
classification cs.RO cs.LGcs.SYeess.SY
keywords online learning for controladaptive model predictive controldynamic regretrandom Fourier featurestarget trackingmixture of expertsnon-stationary dynamicsquadrotor control
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 proposes a self-adaptive online learning controller for a robot tracking a target whose dynamics are unknown and possibly switching among structured, random, and adversarial behaviors. It runs several lightweight predictors in parallel, each learning the target motion from data via random Fourier features, and uses a bandit-style expert selector to pick the predictor that has tracked best recently; the chosen predictor feeds a model predictive controller. The main theoretical claim is that the expected cumulative tracking cost stays within O(T^{3/4}) of the cost of an optimal non-causal controller that knows the target's future, so average regret vanishes. With nonzero learning error or switching, the regret degrades gracefully and remains sublinear in many cases. If true, this lets a robot track unpredictably moving objects without offline training or per-scenario retuning.

Core claim

On the paper's own terms: the method achieves expected dynamic regret bounded by O(T^{3/4}) under four assumptions — bounded robot and target states, Lipschitzness of the value function, an MPC perturbation bound, and representability of the target dynamics by one of the finite random-Fourier-feature predictors. Since average regret is Regret_T / T, this goes to zero, so in expectation the self-adaptive learner asymptotically matches a non-causal controller that knows the target's future. With nonzero approximation error ε_t, the bound becomes O(T^{3/4} + sqrt(T Σ_t ||ε_t||^2)); with κ switches of the best predictor, it becomes O(κ^{1/4}T^{3/4} + sqrt(T Σ_t ||ε_t||^2) + Õ(sqrt(κT))). The pro

What carries the argument

The central mechanism is a composition of three modules: (i) isolated online learning (ISO), which maintains P independent predictors of the target's N-step-ahead motion, each a finite random-Fourier-feature representation updated by efficient online gradient descent; (ii) adaptive selection (AS), which scores each predictor by its recent prediction error over a sliding window and keeps an exponentially weighted (Hedge) probability distribution over predictors, with a forgetting factor to handle switching; and (iii) model predictive control (MPC), which simulates the robot and the selected predictor over a horizon to produce control inputs. The proof's load-bearing inequality is A^⋆_t(u_t; x

Load-bearing premise

The no-regret theorem assumes that the target's true dynamics are exactly expressible by one of the finite random-Fourier-feature predictors with bounded weights; if representability fails, the guarantee becomes an additive term proportional to the total prediction error and average regret vanishes only as that error goes to zero.

What would settle it

Simulate a non-switching target whose dynamics h is exactly representable by one of the RFF predictors, run the algorithm over a long horizon, and measure average regret; if Regret_T / T does not decay roughly as T^{-1/4} — or total regret grows faster than O(T^{3/4}) — the claimed bound is contradicted.

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

If this is right

  • For exactly representable, non-switching targets, average tracking cost converges to that of the optimal non-causal controller at rate O(T^{-1/4}), so long missions incur no asymptotic penalty for learning online.
  • When the learned model has bounded error ε_max, the algorithm's average regret stays within O(ε_max) of the clairvoyant controller, so even imperfect predictors give near-optimal tracking.
  • Under switching dynamics, the regret bound grows as κ^{1/4}T^{3/4} plus Õ(sqrt(κT)); in the worst case κ = O(T), the average regret remains bounded up to logarithmic factors, so the method need not be re-tuned when the target's behavior changes.
  • The full pipeline runs at 50 Hz with about 9 ms per step for four predictors, making the scheme usable in real-time drone tracking.
  • The architecture — a portfolio of cheap online predictors plus a bandit selector — generalizes the earlier single-predictor RFF-MPC approach to non-stationary targets and removes the need to know which predictor will work best in advance.

Where Pith is reading between the lines

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

  • Editorial inference: the literal no-regret statement holds only for targets inside the chosen random-Fourier-feature class; outside that class, the operative guarantee is the additive error term, so practitioners should read the advertised 'no regret' as 'regret proportional to representability.'
  • Editorial inference: the architecture is a general recipe — a portfolio of cheap predictors plus a bandit selector — and the same proof structure should transfer to other function classes, such as online deep learners or Gaussian processes, provided each learner has a per-step regret bound.
  • Editorial inference: the forgetting factor γ in the selector defines a time scale for switching; a natural extension is to make γ itself adaptive to observed switch frequency, which the paper leaves as a fixed hyperparameter.
  • Editorial inference: the hardware experiments include downwash disturbances on the target that are outside the modeled dynamics; the method's success there hints that the guarantees may tolerate unmodeled bounded disturbances, but the paper does not prove this.

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

3 major / 4 minor

Summary. The paper proposes a self-adaptive online learning + MPC method for tracking a target whose dynamics are unknown and may switch between structured, random, and adversarial behaviors. The ISO module trains several independent RFF-based predictors online via OGD, the AS module selects among them with Hedge, and the MPC module plans with the selected predictor. The central theoretical claim is Theorem 1: under Assumptions 1-4, the expected dynamic regret with respect to a non-causal target-aware MPC is O(T^{3/4}), with corollaries for nonzero prediction error and switching. The paper also reports simulations and Crazyflie hardware experiments comparing against PLOT, RFF-MPC, and MODL on lemniscate, sinusoidal, star, random-walk, adversarial, and switching trajectories, plus ablations isolating the ISO and AS modules.

Significance. The algorithmic architecture is plausible and the experimental component is a real strength: the Crazyflie hardware tests with switching targets and the ISO/AS ablations credibly demonstrate that a fixed predictor configuration is insufficient and that Hedge-based selection improves robustness. If the theoretical result were established, the paper would make a useful contribution toward online learning MPC with expert selection for non-stationary tracking. However, the advertised no-regret guarantee is not presently supported: the proof of Theorem 1 contains a trajectory-conflation error in the performance-difference lemma step, and the OGD regret bound is applied to a loss sequence that is not the one appearing in the MPC prediction error. In addition, Assumption 4 restricts the main theorem to exact finite-dimensional representability, which does not cover the random/adversarial targets emphasized in the experiments.

major comments (3)
  1. [Appendix 10.1, Eqs. (18)-(26) and Definition 1] The proof of Theorem 1 conflates the comparator's own trajectory with the one-step continuation from the algorithm's current state. In Eq. (20), the second term should be Q*_t(x_t, u*_t) = ||e_t||^2_Q + ||u*_t||^2_R + V*_{t+1}(f(x_t)+g(x_t)u*_t - y_{t+1}). Instead, the proof writes V*_{t+1}(e*_{t+1}) with e*_{t+1}=x*_{t+1}-y_{t+1}. Under Definition 1, x*_{t+1}=f(x*_t)+g(x*_t)u*_t, so Eq. (25), x_{t+1}-x*_{t+1}=g(x_t)(u_t-u*_t), is false unless x_t=x*_t. The PDL decomposition in Eq. (18) is therefore not justified as written. The authors need to define the comparator as a policy and use u*_t = pi*(x_t) consistently throughout, or provide a correct dynamic-regret decomposition for nonlinear dynamics. This is load-bearing for Theorem 1.
  2. [Appendix 10.1, Eq. (34) and Section 5.3 / Algorithm 1] The proof identifies the MPC prediction errors ||y_{t+1+j} - hat{y}^p_{t+1+j}||^2 with the OGD training losses l^{j,p}_t(alpha^{j,p}_t). But Algorithm 1 defines l^{j,p}_t = ||y_t - hat{h}^p_j(y_{t-1-tau-j:t-1-j}; alpha^{j,p}_t)||^2. The MPC prediction made at time t for y_{t+j+1} uses feature window y_{t-tau:t} and parameter alpha^{j,p}_t, whereas the same training loss with that feature and target is only observed at time t+j+1 and is evaluated with alpha^{j,p}_{t+j+1}. Proposition 2 bounds the cumulative online losses, not the delayed sequence appearing in the regret. A delayed-feedback regret analysis (or a redefinition of the training loss) is needed before Eq. (34) and the O(T^{3/4}) rate follow.
  3. [Assumption 4 vs. Sections 7-8] The no-regret Theorem 1 requires exact representability of the unknown target dynamics by a finite-dimensional RFF predictor with a fixed parameter alpha_* (Assumption 4). Random-walk and randomized adversarial targets in the experiments are not exactly representable in this class. Corollary 1 adds sqrt(T sum_t ||epsilon_t||^2), which can be linear in T if the approximation error is bounded away from zero, so the abstract's claim of no regret for adversarial/random target dynamics is not supported by the theory. The paper should either soften the claim or add a quantitative bound under a learnability assumption with explicit dependence on the number of random features M.
minor comments (4)
  1. [Algorithm 2, line 9] The denominator should be sum_p exp(lambda S^p_t), not sum_p exp(lambda S^p_{t-1}); the text in Section 6.2 has the correct expression.
  2. [Algorithm 1, lines 4-7] The loop observes y_{t+1} but the loss is defined with y_t. This indexing mismatch should be fixed for clarity, and it matters for the interpretation of Proposition 2.
  3. [Theorem 2 and Section 7] The statement 'with a high probability, the expectation of the regret' is imprecise; the standard Hedge regret bound for bounded losses is deterministic in expectation. Also, the loose sentence after Corollary 2 that the worst-case average regret becomes O-tilde(1) should be rechecked: for kappa=O(T), the kappa^{1/4}T^{3/4} term is O(T), which is O(1) average only after division by T.
  4. [Section 5.3 / Eq. (7)] The notation for the RFF approximation and the projection step is dense; a short worked example or a table of dimension/indices (t, j, p, M) would help readers follow the delay structure in the proof.

Circularity Check

0 steps flagged

No significant circularity; the no-regret claim is conditional on explicit assumptions and standard external learning bounds.

full rationale

I walked the derivation chain: Regret is decomposed via the performance difference lemma, then bounded through Assumptions 1-3, Hedge (Theorem 2), and OGD (Proposition 2). No step reduces a prediction to a fitted input or defines X in terms of Y. Assumption 4 assumes exact RFF representability of the target, but it is an explicit hypothesis, not a consequence smuggled into the proof; the theorem would be conditional even if the assumption is strong. The self-citation to Zhou and Tzoumas (2024) is used as an algorithmic subroutine and comparison baseline, but the regret bound for the online least-squares estimator is attributed to Hazan (2016), and the Hedge bound to Cesa-Bianchi and Lugosi (2006), so the self-citation is not load-bearing in the proof. I also examined the apparent proof gap in Appendix 10.1, eqs. (24)-(25): the proof replaces the Definition 2 continuation f(x_t)+g(x_t)u*_t-y_{t+1} with x*_{t+1}-y_{t+1}, which is not an identity unless x_t=x*_t. That is a correctness/rigor concern, not a circularity: it does not make the theorem an input by construction, nor does it rename a fitted parameter as a prediction. Assumption 1 similarly assumes bounded tracking errors rather than proving them, but boundedness alone does not force the O(T^{3/4}) regret rate, so the derivation retains independent content. Overall, the central claim is a conditional bound assembled from stated assumptions and standard external results, with no circular reduction identified.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The method relies on several hand-selected hyperparameters and four strong assumptions; the most fragile is Assumption 4, which assumes the target dynamics are exactly representable by a finite set of random Fourier features. The theoretical contributions are conditional on these. No new physical entities are introduced.

free parameters (6)
  • Kernel width σ for each predictor (P1-P4) = 0.1, 1.0, 0.1, 1.5
    Hand-chosen based on anticipated target behaviors; ablation (Fig. 10) shows optimal σ varies significantly with target type.
  • OGD learning rate η for each predictor = 0.25, 0.1, 0.15, 0.1
    Manual configuration; theory requires η=O(1/√T), but experiments use constants.
  • Feature history horizon τ = 10
    Manual; ablation shows larger τ helps smooth trajectories but has little effect for adversarial targets.
  • Hedge forgetting factor γ = 0.98
    Manual in experiments; theory (Theorem 2) assumes γ=1.
  • Hedge learning rate λ = 0.2
    Manual in experiments; theory requires λ=O(1/√T).
  • Reward shaping constant a = 4
    Chosen to normalize rewards to [0,1] and avoid the need for the upper bound ¯r.
axioms (6)
  • domain assumption Assumption 1: The MPC policies in eq. (3) and eq. (6) render the tracking errors e*_t and e_t uniformly bounded.
    This is a closed-loop stability-like assumption, close to what the theorem might hope to prove; it is stated but not derived.
  • domain assumption Assumption 2: The optimal value function V*_t(e) is Lipschitz in the tracking error e.
    Justified for linear MPC and some exponentially-stable nonlinear MPC, but not verified for the Crazyflie experiments.
  • domain assumption Assumption 3: A perturbation bound on the MPC policy holds, relating control-input differences to target-prediction errors with summable coefficients.
    Cites Lin et al. 2021, Shin et al. 2020, etc.; typical regularity conditions (small prediction errors, controllability) are not verified in experiments.
  • ad hoc to paper Assumption 4: There exists a predictor p whose RFF approximation error is zero: y_{t+j} = \hat{h}^p_j(\tilde{z}_j; α^{j,p}_*) for α^{j,p}_* ∈ D_j.
    This realizability assumption is essential for the no-regret result; it is implausible for unknown/adversarial target dynamics, and Corollary 1 shows the regret grows with the approximation error.
  • standard math Proposition 1 (RFF uniform approximation, from Boffi et al. 2022): functions in a bounded RKHS can be approximated uniformly by M random features with error O(1/√M).
    Standard result in random Fourier feature literature, used as a black box.
  • standard math Hedge switching regret bound (Matsuoka et al. 2021): in a switching environment, the regret of Hedge with a proper learning rate is Õ(√(κT)).
    Standard result; used to derive Corollary 2 without proof in the appendix.

pith-pipeline@v1.3.0-daily-deepseek · 30247 in / 17959 out tokens · 174804 ms · 2026-08-01T17:15:25.065097+00:00 · methodology

0 comments
read the original abstract

We propose a self-adaptive online learning for control method for tracking unknown target dynamics. The target dynamics can exhibit switching behavior, particularly, a mixture of structured, random, and/or adversarial motion. Such challenging target tracking scenarios arise in applications of dynamic mapping, traffic control, and pursuit evasion, where robots need to track, pursue, or avoid collision with moving landmarks, objects, humans, etc., whose dynamics are unknown. Our method simultaneously learns multiple predictors from scratch, via self-supervised, one-shot, and computationally efficient learning, and adaptively selects the best one to match the observed target behavior. The method enjoys finite-time near-optimality guarantees in expectation, characterized as a function of the learning error of the target dynamics and the frequency that the target dynamics switch. In the absence of both error and switching, the method asymptotically matches the optimal non-causal control policy that knows a priori the target dynamics, i.e., the method enjoys no regret in expectation. In the presence of learning errors and switching, the method degrades gracefully, \eg when there are errors and no switching, the average regret is proportional to the average learning error and switching times. To prove these guarantees, a novel technical approach is required compared to the existing works that employ RFF-based online learning. We validate our method in Crazyflie simulations and hardware experiments, across target trajectories that vary from structured to random to adversarial, in comparison to non-stochastic, kernel-based, and neural-network-based methods for online learning.

Figures

Figures reproduced from arXiv: 2607.26370 by Atharva Navsalkar, Hongyu Zhou, Vasileios Tzoumas.

Figure 1
Figure 1. Figure 1: Simplified Motivating Example of Self-Adaptive Learning and Model Predictive Control for Tracking Targets with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Self-Adaptive Learning and MPC for Tracking Unknown Dynamics. The method composes three interacting modules: (i) isolated (self-supervised) online learning of target dynamics (ISO), (ii) adaptive selection of predictors (AS), and (iii) model predictive control (MPC). In summary, ISO trains on-the-fly multiple predictors, AS evaluates the current best predictor, and MPC uses the selected best predictor by A… view at source ↗
Figure 3
Figure 3. Figure 3: Lemniscate target trajectory. (a) Trace of trajectories show that our algorithm achieves qualitatively better tracking performance. (b) The average tracking error plot shows an asymptotically decreasing curve that is converging towards the Optimal*. The inserted plot shows the mean tracking error (m) for the experiment. (c) The retrospective prediction error also shows similar or better performance to the … view at source ↗
Figure 4
Figure 4. Figure 4: Sinusoidal target trajectory. (a) Trace of trajectories. This plot shows the trace only for the final 25s of the total simulation duration of 90s for better clarity. (b) The average tracking error (m) shows superior performance by our algorithm. The inserted plot shows the mean tracking error (m) for the experiment. We consider high values of γ ∈ {0.95, 1.0} for PLOT since smaller values lead to a crash du… view at source ↗
Figure 5
Figure 5. Figure 5: Star-shaped target trajectory [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Random walk target. (a) Trace of trajectories. The target trajectories for each algorithm are different, hence we only plot the target trajectory corresponding to our algorithm. (b) Our algorithm has a lower tracking error (m) than others. The inserted plot shows the mean tracking error (m) for the experiment. (c) Retrospective prediction error shows a similar trend to that in (b). −5 0 5 10 x coordinate (… view at source ↗
Figure 7
Figure 7. Figure 7: Randomized adversarial target. (a) Trace of trajectories. In this case, the target trajectories for each algorithm are different, hence we only plot the target trajectory corresponding to our algorithm. (b) Our algorithm has substantially lower tracking error (m) than others. The inserted plot shows the mean tracking error (m) for the experiment. (c) Retrospective prediction error shows a similar trend to … view at source ↗
Figure 8
Figure 8. Figure 8: Setup for hardware experiments. Two physical Crazyflie drones are used, one as a pursuer and one as a target. To avoid collisions, a vertical offset between the pursuer and target drone is maintained. Due to the vertical offset, the target drone is also affected by the downwash effect of the pursuer drone, especially when the pursuer drone is able to closely track the target drone. Such downwash effects ad… view at source ↗
Figure 9
Figure 9. Figure 9: Hardware experiment with 2 Crazyflie drones. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of standard deviation σ and learning rate η. Mean tracking error vs. standard deviation of the Gaussian kernel σ for different target types and learning rates η. The red line is for the lemniscate trajectory target, while the blue line is for the randomized adversarial target. The optimal value of σ varies significantly for different target types. Increasing η generally improves the performance, bu… view at source ↗
Figure 12
Figure 12. Figure 12: Lemniscate target trajectory. (a) The evolution of the average tracking error and the mean tracking error (inserted graph) for our algorithm is similar to the performance of the best expert. (b) The probability assigned to each expert shows how the algorithm favors the agile expert initially, and then, as it learns, the accurate expert. This is also reflected in the individual performance of the isolated … view at source ↗
Figure 14
Figure 14. Figure 14: Trends for Regret, tracking and prediction errors. [PITH_FULL_IMAGE:figures/full_fig_p019_14.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

270 extracted references · 21 linked inside Pith

  1. [1]

    IEEE Communications Surveys & Tutorials , volume=

    Autonomous cars: Research results, issues, and future challenges , author=. IEEE Communications Surveys & Tutorials , volume=. 2018 , publisher=

    2018

  2. [2]

    IEEE Transactions on Robotics , volume=

    Cinempc: A fully autonomous drone cinematography system incorporating zoom, focus, pose, and scene composition , author=. IEEE Transactions on Robotics , volume=. 2024 , publisher=

    2024

  3. [3]

    IEEE Access , volume=

    Long Short-Term Memory-Based Neural Networks for Missile Maneuvers Trajectories Prediction , author=. IEEE Access , volume=. 2023 , publisher=

    2023

  4. [4]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Crazysim: A software-in-the-loop simulator for the crazyflie nano quadrotor , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  5. [5]

    Mathematical Programming Computation , Year =

    acados -- a modular open-source framework for fast embedded optimal control , Author =. Mathematical Programming Computation , Year =

  6. [6]

    Mathematical Programming Computation , volume =

    Joel A E Andersson and Joris Gillis and Greg Horn and James B Rawlings and Moritz Diehl , title =. Mathematical Programming Computation , volume =. 2019 , publisher =

    2019

  7. [7]

    2022 IEEE 61st Conference on Decision and Control (CDC) , pages=

    Online learning for predictive control with provable regret guarantees , author=. 2022 IEEE 61st Conference on Decision and Control (CDC) , pages=. 2022 , organization=

    2022

  8. [8]

    arXiv preprint arXiv:2404.02023 , year=

    On the Regret of Recursive Methods for Discrete-Time Adaptive Control with Matched Uncertainty , author=. arXiv preprint arXiv:2404.02023 , year=

    Pith/arXiv arXiv

  9. [9]

    2021 60th IEEE Conference on Decision and Control (CDC) , pages=

    Model learning predictive control in nonlinear dynamical systems , author=. 2021 60th IEEE Conference on Decision and Control (CDC) , pages=. 2021 , organization=

    2021

  10. [10]

    Foundations and Trends in Optimization , volume=

    Introduction to online convex optimization , author=. Foundations and Trends in Optimization , volume=. 2016 , publisher=

    2016

  11. [11]

    Foundations and Trends

    Online learning and online convex optimization , author=. Foundations and Trends. 2012 , publisher=

    2012

  12. [12]

    2006 , publisher=

    Prediction, learning, and games , author=. 2006 , publisher=

    2006

  13. [13]

    Online convex programming and generalized infinitesimal gradient ascent , author=. Interna. Conf. on Machine Learning (ICML) , pages=

  14. [14]

    SIAM Journal on Optimization , volume=

    Convergence analysis of a proximal-like minimization algorithm using Bregman functions , author=. SIAM Journal on Optimization , volume=. 1993 , publisher=

    1993

  15. [15]

    Conference on Learning Theory (COLT) , volume=

    Composite objective mirror descent , author=. Conference on Learning Theory (COLT) , volume=. 2010 , organization=

    2010

  16. [16]

    IEEE Transactions on Information Theory (TIT) , volume=

    On delayed prediction of individual sequences , author=. IEEE Transactions on Information Theory (TIT) , volume=. 2002 , publisher=

    2002

  17. [17]

    ACM Transactions on Intelligent Systems and Technology (TIST) , volume=

    Understand dynamic regret with switching cost for online decision making , author=. ACM Transactions on Intelligent Systems and Technology (TIST) , volume=. 2020 , publisher=

    2020

  18. [18]

    2004 , publisher=

    Convex optimization , author=. 2004 , publisher=

    2004

  19. [19]

    2021 , publisher=

    Algorithms for convex optimization , author=. 2021 , publisher=

    2021

  20. [20]

    2009 , publisher=

    Robust optimization , author=. 2009 , publisher=

    2009

  21. [21]

    Advanced Robotics , volume=

    Deep learning in robotics: a review of recent research , author=. Advanced Robotics , volume=. 2017 , publisher=

    2017

  22. [22]

    2018 , publisher=

    Foundations of machine learning , author=. 2018 , publisher=

    2018

  23. [23]

    2012 , publisher=

    Learning from data , author=. 2012 , publisher=

    2012

  24. [24]

    The International Journal of Robotics Research , volume=

    Autonomous helicopter aerobatics through apprenticeship learning , author=. The International Journal of Robotics Research , volume=. 2010 , publisher=

    2010

  25. [25]

    Journal of Guidance, Control, and Dynamics , volume=

    Real-time optimal control via deep neural networks: study on landing problems , author=. Journal of Guidance, Control, and Dynamics , volume=. 2018 , publisher=

    2018

  26. [26]

    IEEE Robotics and Automation Letters , volume=

    Data-driven model predictive control for trajectory tracking with a robotic arm , author=. IEEE Robotics and Automation Letters , volume=. 2019 , publisher=

    2019

  27. [27]

    arXiv preprint arXiv:1906.01728 , year=

    Bayessim: adaptive domain randomization via probabilistic inference for robotics simulators , author=. arXiv preprint arXiv:1906.01728 , year=

    Pith/arXiv arXiv 1906

  28. [28]

    2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages=

    Domain randomization for transferring deep neural networks from simulation to the real world , author=. 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages=. 2017 , organization=

    2017

  29. [29]

    Science robotics , volume=

    Learning quadrupedal locomotion over challenging terrain , author=. Science robotics , volume=. 2020 , publisher=

    2020

  30. [30]

    Information theoretic

    Williams, Grady and Wagener, Nolan and Goldfain, Brian and Drews, Paul and Rehg, James M and Boots, Byron and Theodorou, Evangelos A , booktitle=. Information theoretic. 2017 , organization=

    2017

  31. [31]

    IEEE Robotics and Automation Letters , volume=

    Model-based meta-reinforcement learning for flight with suspended payloads , author=. IEEE Robotics and Automation Letters , volume=. 2021 , publisher=

    2021

  32. [32]

    IEEE Transactions on Robotics , year=

    Active learning of discrete-time dynamics for uncertainty-aware model predictive control , author=. IEEE Transactions on Robotics , year=

  33. [33]

    1996 , publisher=

    Robust adaptive control , author=. 1996 , publisher=

    1996

  34. [34]

    PRENTICE-HALL google schola , volume=

    Applied Nonlinear Control , author=. PRENTICE-HALL google schola , volume=

  35. [35]

    1995 , publisher=

    Nonlinear and adaptive control design , author=. 1995 , publisher=

    1995

  36. [36]

    Science Robotics , volume=

    Neural-fly enables rapid learning for agile flight in strong winds , author=. Science Robotics , volume=. 2022 , publisher=

    2022

  37. [37]

    2019 International Conference on Robotics and Automation (ICRA) , pages=

    Neural lander: Stable drone landing control using learned dynamics , author=. 2019 International Conference on Robotics and Automation (ICRA) , pages=. 2019 , organization=

    2019

  38. [38]

    Learning for Dynamics and Control , pages=

    Regret bounds for adaptive nonlinear control , author=. Learning for Dynamics and Control , pages=. 2021 , organization=

    2021

  39. [39]

    Journal of Machine Learning Research , volume=

    Nonparametric adaptive control and prediction: Theory and randomized algorithms , author=. Journal of Machine Learning Research , volume=

  40. [40]

    L1 Quad:

    Wu, Zhuohuan and Cheng, Sheng and Zhao, Pan and Gahlawat, Aditya and Ackerman, Kasey A and Lakshmanan, Arun and Yang, Chengyu and Yu, Jiahao and Hovakimyan, Naira , journal=. L1 Quad:

  41. [41]

    2016 American Control Conference (ACC) , pages=

    A tutorial and overview of retrospective cost adaptive control , author=. 2016 American Control Conference (ACC) , pages=. 2016 , organization=

    2016

  42. [42]

    arXiv preprint arXiv:2401.01881 , year=

    Robust Control Barrier Functions using Uncertainty Estimation with Application to Mobile Robots , author=. arXiv preprint arXiv:2401.01881 , year=

    Pith/arXiv arXiv

  43. [43]

    IEEE Transactions on Robotics , year=

    EVOLVER: Online Learning and Prediction of Disturbances for Robot Control , author=. IEEE Transactions on Robotics , year=

  44. [44]

    2012 IEEE International Conference on Robotics and Automation , pages=

    Learning-based model predictive control on a quadrotor: Onboard implementation and experimental results , author=. 2012 IEEE International Conference on Robotics and Automation , pages=. 2012 , organization=

    2012

  45. [45]

    2012 American Control Conference (ACC) , pages=

    Extensions of learning-based model predictive control for real-time application to a quadrotor helicopter , author=. 2012 American Control Conference (ACC) , pages=. 2012 , organization=

    2012

  46. [46]

    arXiv preprint arXiv:2211.09619 , year=

    Introduction to online nonstochastic control , author=. arXiv preprint arXiv:2211.09619 , year=

    Pith/arXiv arXiv

  47. [47]

    International Conference on Machine Learning (ICML) , pages=

    Online control with adversarial disturbances , author=. International Conference on Machine Learning (ICML) , pages=

  48. [48]

    Algorithmic Learning Theory (ALT) , pages=

    The nonstochastic control problem , author=. Algorithmic Learning Theory (ALT) , pages=

  49. [49]

    Learning for Dynamics and Control Conference , pages=

    Adaptive regret for control of time-varying dynamics , author=. Learning for Dynamics and Control Conference , pages=. 2023 , organization=

    2023

  50. [50]

    Conference on Learning Theory (COLT) , pages=

    Improper learning for non-stochastic control , author=. Conference on Learning Theory (COLT) , pages=

  51. [51]

    International Conference on Machine Learning , pages=

    Logarithmic regret for adversarial online control , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  52. [52]

    AAAI Conference on Artificial Intelligence (AAAI) , volume=

    Online optimal control with affine constraints , author=. AAAI Conference on Artificial Intelligence (AAAI) , volume=

  53. [53]

    International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=

    Non-stationary online learning with memory and non-stochastic control , author=. International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=. 2022 , organization=

    2022

  54. [54]

    2023 62nd IEEE Conference on Decision and Control (CDC) , pages=

    Efficient online learning with memory via frank-wolfe optimization: Algorithms with bounded dynamic regret and applications to control , author=. 2023 62nd IEEE Conference on Decision and Control (CDC) , pages=. 2023 , organization=

    2023

  55. [55]

    2023 62nd IEEE Conference on Decision and Control (CDC) , pages=

    Safe non-stochastic control of linear dynamical systems , author=. 2023 62nd IEEE Conference on Decision and Control (CDC) , pages=. 2023 , organization=

    2023

  56. [56]

    IEEE Robotics and Automation Letters , year=

    Safe non-stochastic control of control-affine systems: An online convex optimization approach , author=. IEEE Robotics and Automation Letters , year=

  57. [57]

    arXiv preprint arXiv:2407.04143 , year=

    Simultaneous System Identification and Model Predictive Control with No Dynamic Regret , author=. arXiv preprint arXiv:2407.04143 , year=

    arXiv

  58. [58]

    arXiv preprint arXiv:2510.15626 , year=

    Adaptive Legged Locomotion via Online Learning for Model Predictive Control , author=. arXiv preprint arXiv:2510.15626 , year=

    arXiv

  59. [59]

    International Conference on Machine Learning , pages=

    Tracking adversarial targets , author=. International Conference on Machine Learning , pages=. 2014 , organization=

    2014

  60. [60]

    International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=

    Adversarial tracking control via strongly adaptive online learning with memory , author=. International Conference on Artificial Intelligence and Statistics (AISTATS) , pages=. 2022 , organization=

    2022

  61. [61]

    Transactions on Machine Learning Research , number=

    Online Optimal Tracking of Linear Systems with Adversarial Disturbances , author=. Transactions on Machine Learning Research , number=

  62. [62]

    IEEE Control Systems Letters , year=

    Online linear quadratic tracking with regret guarantees , author=. IEEE Control Systems Letters , year=

  63. [63]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Predictive linear online tracking for unknown targets , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  64. [64]

    2022 IEEE 61st Conference on Decision and Control (CDC) , pages=

    The power of linear controllers in LQR control , author=. 2022 IEEE 61st Conference on Decision and Control (CDC) , pages=. 2022 , organization=

    2022

  65. [65]

    arXiv preprint arXiv:2412.00922 , year=

    Online convex optimization for constrained control of nonlinear systems , author=. arXiv preprint arXiv:2412.00922 , year=

    Pith/arXiv arXiv

  66. [66]

    International Conference on Artificial Intelligence and Statistics , pages=

    Tracking regret bounds for online submodular optimization , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2021 , organization=

    2021

  67. [67]

    1999 , publisher=

    Differential games: a mathematical theory with applications to warfare and pursuit, control and optimization , author=. 1999 , publisher=

    1999

  68. [68]

    2020 American Control Conference (ACC) , pages=

    An introduction to pursuit-evasion differential games , author=. 2020 American Control Conference (ACC) , pages=. 2020 , organization=

    2020

  69. [69]

    Indiana University mathematics journal , volume=

    Differential games and representation formulas for solutions of Hamilton-Jacobi-Isaacs equations , author=. Indiana University mathematics journal , volume=. 1984 , publisher=

    1984

  70. [70]

    IEEE Transactions on Automatic Control , volume=

    Differential games and optimal pursuit-evasion strategies , author=. IEEE Transactions on Automatic Control , volume=. 1965 , publisher=

    1965

  71. [71]

    IEEE Transactions on automatic control , volume=

    Hamilton--Jacobi formulation for reach--avoid differential games , author=. IEEE Transactions on automatic control , volume=. 2011 , publisher=

    2011

  72. [72]

    SIAM Journal on Control and Optimization , volume=

    Reachability and minimal times for state constrained nonlinear problems without any controllability assumption , author=. SIAM Journal on Control and Optimization , volume=. 2010 , publisher=

    2010

  73. [73]

    IFAC Proceedings Volumes , volume=

    Minimal time problems with moving targets and obstacles , author=. IFAC Proceedings Volumes , volume=. 2011 , publisher=

    2011

  74. [74]

    Proceedings of the 18th international conference on hybrid systems: computation and control , pages=

    Reach-avoid problems with time-varying dynamics, targets and constraints , author=. Proceedings of the 18th international conference on hybrid systems: computation and control , pages=

  75. [75]

    2015 54th IEEE conference on decision and control (CDC) , pages=

    The pursuit-evasion-defense differential game in dynamic constrained environments , author=. 2015 54th IEEE conference on decision and control (CDC) , pages=. 2015 , organization=

    2015

  76. [76]

    IEEE Robotics and Automation Letters , volume=

    Autonomous flight for detection, localization, and tracking of moving targets with a small quadrotor , author=. IEEE Robotics and Automation Letters , volume=. 2017 , publisher=

    2017

  77. [77]

    Journal of Intelligent & Robotic Systems , volume=

    Vision based autonomous landing of multirotor UAV on moving platform , author=. Journal of Intelligent & Robotic Systems , volume=. 2017 , publisher=

    2017

  78. [78]

    IEEE Transactions on Control Systems Technology , volume=

    Robust visual servoing control for ground target tracking of quadrotors , author=. IEEE Transactions on Control Systems Technology , volume=. 2019 , publisher=

    2019

  79. [79]

    2014 IEEE international conference on robotics and automation (ICRA) , pages=

    Toward image based visual servoing for aerial grasping and perching , author=. 2014 IEEE international conference on robotics and automation (ICRA) , pages=. 2014 , organization=

    2014

  80. [80]

    Automatica , volume=

    Image-based tracking control of VTOL unmanned aerial vehicles , author=. Automatica , volume=. 2015 , publisher=

    2015

Showing first 80 references.