PaperlayerView exact revision on arXiv

Source-linked AI summary

A Decade of Learning-Based Control: From Theoretic Foundations to Practical Deployment

Angela P. Schoellig, Matthias A. Müller, Sebastian Trimpe, Melanie N. Zeilinger, Mahathi Anand, Andrea Carron, Oliver Hausdörfer, Victor G. Lopez, Friedrich Solowjow, Haocheng Zhao, Siqi Zhou

arXiv:2610.29187v1eess.SYcs.LGcs.SY

TL;DR

Learning-based control must reconcile expressive, data-driven methods with the safety and reliability demands of physical systems. This tutorial reviews approaches that combine machine learning and control structure, concluding that this combination is central to reliable high-performance behavior in complex, uncertain, safety-critical settings.

  • Problem

    Purely data-driven models can struggle with strict safety and reliability requirements and may require large amounts of costly or infeasible real-system data.

    Source: §I · p. 1

  • Method

    The tutorial synthesizes learning-based control foundations and methods, including dynamics learning, direct data-driven control, reinforcement learning, imitation learning, and Bayesian optimization.

    Source: §I · p. 2, D. Controller Tuning with Bayesian… · p. 20

  • Results

    Across the surveyed approaches, machine learning expands modeling, optimization, and adaptation, while control theory supplies structure for stability, safety, robustness, and reliable closed-loop behavior.

    Source: §VIII · p. 22

  • Takeaways & Limitations

    The authors conclude that data-based methods must be incorporated directly into the closed loop for reliable high-performance behavior in complex, uncertain, safety-critical settings.

    Source: §VIII · p. 22

    Reliable scaling remains challenging because systems must learn continually while avoiding catastrophic forgetting, meeting real-time resource constraints, and preserving closed-loop guarantees.

    Source: §VIII · p. 22

Abstract

from arXiv · show

Motivated by ten years of Learning-Based Control Invited Sessions at the IEEE Conference on Decision and Control (CDC), this tutorial paper reviews key advances at the intersection of machine learning and control. We focus on methods that combine the flexibility of machine learning techniques with the structure and theoretical foundations of control to design autonomous systems that are capable, reliable, and safe. We first introduce core notions for establishing stability and safety guarantees, then review approaches that learn dynamics models to improve control under uncertainty, and finally discuss data-driven methods that directly synthesize or learn policies, including reinforcement learning, imitation learning, and Bayesian optimization. We further discuss representative applications and open challenges. The goal is to provide an accessible entry point to the field and to support continued progress toward reliable learning-based control in real-world systems. To support learning and practical exploration, we share implementations and illustrative examples along with our discussion, together with additional resources on our website: https://learning-based-control.github.io/tutorial/.

I. INTRODUCTION

The introduction argues that learning’s expressive power must be paired with control structure to address safety, reliability, and data demands in physical systems. The paper surveys learning-based control and formulates a general problem in which empirical data and available prior knowledge support policy synthesis under objectives and constraints.

I. INTRODUCTION

  • The tutorial connects machine-learning flexibility with control-theoretic structure to address safety, reliability, and costly data requirements in physical systems.It highlights developments spanning foundations, dynamics learning, direct data-driven control, and policy optimization, with representative applications and future directions.Source: §I · pp. 1–2
  • The CDC invited sessions have hosted 30 dedicated sessions and 182 accepted papers from over 100 institutions since 2016.Recurring themes combine neural networks, reinforcement learning, and Gaussian processes with control tools such as MPC, reachability, CLFs, and CBFs.Source: §I · p. 2

II. PROBLEM FORMULATION

  • The formulation treats system evolution and sensor measurements separately, with process disturbances and measurement noise represented explicitly.State feedback, stochastic noise, bounded uncertainty, and discrete-time nonlinear dynamics are common specializations.Source: §II · p. 3
  • A policy maps the current state or measured output, and possibly a reference, to a control input; the objective and constraints encode the desired task and admissible behavior.Objectives may represent reference tracking, energy minimization, or disturbance attenuation.Source: §II · pp. 3–4
  • Learning-based control uses empirical data, possibly with partial prior knowledge, to synthesize policies when models, objectives, or constraints are partly or wholly unknown.The tutorial focuses mainly on state-feedback and offline learning, while discussing output measurements, state estimation, and online learning at selected points.Source: §II · p. 4

III. STABILITY AND SAFETY

The section introduces stability and safety foundations, moving from nominal dynamics to robust and stochastic formulations that account for uncertainty. In the nominal setting, Lyapunov functions certify stability through decrease along closed-loop trajectories.

III. STABILITY AND SAFETY

  • The section frames stability and safety guarantees first for known nominal dynamics, then extends the discussion to uncertainty-aware robust and stochastic settings.For learned models, approximation error motivates treating uncertainty as a bounded disturbance or probabilistic description.Source: §III · p. 4
  • Class K, K∞, and KL functions provide comparison-function classes used to state the stability conditions.K∞ functions additionally grow without bound, while KL functions decay in their second argument.Source: §III · p. 4

A. Nominal Stability

  • Nominal stability concerns local asymptotic behavior around an equilibrium, and Lyapunov’s direct method certifies it using a scalar function’s decrease along closed-loop trajectories.Nonincrease establishes stability; a suitable strict decrease condition establishes asymptotic stability.Source: A. Nominal Stability · p. 4

B. Robust Stability

Robust stability accounts for model error and external disturbances, while stochastic stability handles probabilistic disturbances. These formulations bound residual effects rather than requiring convergence exactly to the equilibrium.

B. Robust Stability

  • The figure contrasts robust and stochastic convergence to a disturbance-dependent neighborhood with safety enforced by an invariant set inside the constraint set.It also depicts safety filters modifying a desired input that would violate a constraint.Source: Figure 5 · p. 5
  • Under bounded disturbances, input-to-state stability separates the decaying effect of the initial condition from residual error induced by uncertainty.An ISS-Lyapunov condition provides a way to establish this property; when disturbances vanish, the nominal asymptotic stability is recovered.Source: B. Robust Stability · p. 5
  • Incremental stability and δ-ISS compare pairs of trajectories, allowing nominal and perturbed trajectories to be compared under model mismatch.This differs from stability notions centered on convergence toward one equilibrium.Source: B. Robust Stability · p. 5

C. Stochastic Stability

  • With stochastic disturbances, the state may remain bounded in mean square within a neighborhood of the equilibrium, with long-term average performance bounded by c.When c = 0, the equilibrium is asymptotically stable in expectation.Source: C. Stochastic Stability · p. 6

D. Invariance and Safety Filters

Safety is an invariance problem: a safe set must keep trajectories within constraints over time. HJ reachability, CBFs, and PSFs certify safety through different safe-set representations and computational trade-offs.

D. Invariance and Safety Filters

  • Safety requires identifying states from which constraints remain satisfied for all future times, not merely checking constraints at one instant.Forward invariance keeps trajectories inside a set; safety follows when that set lies within the constraints.Source: D. Invariance and Safety Filters · p. 6
  • HJ reachability computes an approximate maximal safe set but scales exponentially with state dimension, while CBFs are cheaper to obtain but generally provide an inner approximation.Both filters are cheap to evaluate online once their certificates are available.Source: D. Invariance and Safety Filters · p. 6
  • A PSF certifies safety by finding a feasible backup trajectory to a local invariant terminal set, rather than computing a global safe set explicitly.If the desired input has a feasible backup trajectory, it passes unchanged; otherwise, the filter minimally adjusts it, with guarantees provided the terminal set is control invariant.Source: D. Invariance and Safety Filters · p. 7
  • Model predictive shielding simulates a fixed backup policy and applies it when the learned policy's proposed action cannot be verified as safe.It is a simulation-based variant of a PSF rather than an online optimal-control problem.Source: D. Invariance and Safety Filters · p. 7
  • HJ reachability represents the safe set explicitly, CBFs characterize it with a barrier function, and PSFs define it through feasible backup trajectories.Their online problems range from evaluating a precomputed value function to a one-step barrier constraint or finite-horizon optimization.Source: D. Invariance and Safety Filters · p. 7

E. Robust and Stochastic Invariance

Robust and stochastic invariance extend safety guarantees to disturbances and uncertainty. Relaxed safety formulations trade strict constraint satisfaction for budgeted or soft violations, while dynamics learning uses data to refine models for control under uncertainty.

E. Robust and Stochastic Invariance

  • Robust invariance must hold for all admissible disturbances, whereas stochastic invariance is expressed in probability or expectation.Reachability, barrier methods, and PSFs account for uncertainty through worst-case sets, uncertainty-aware conditions, or constraint tightening in robust settings, and probabilistic or risk-based formulations in stochastic settings.Source: E. Robust and Stochastic Invariance · p. 7

F. Relaxed Safety Formulations

  • Constrained Markov decision processes optimize the control objective subject to a budget on expected safety costs, trading strict invariance for long-term probabilistic or average-performance guarantees.Soft constraints permit temporary violations when strict satisfaction may be infeasible or excessively conservative under uncertainty and disturbances.Source: F. Relaxed Safety Formulations · p. 7
  • Soft-constrained optimization introduces nonnegative slack variables for permitted constraint violations and penalizes them to trade performance against violations.In convex settings, a sufficiently large linear penalty makes the penalty exact: if hard constraints are feasible, the soft-constrained problem recovers their solution.Source: F. Relaxed Safety Formulations · pp. 7–8

IV. DYNAMICS LEARNING FOR CONTROL

  • Learned dynamics models can complement incomplete or unavailable prior models, improving control performance while control-theoretic analysis supports desired closed-loop properties.Performance and guarantees depend on model accuracy and how uncertainty is accounted for.Source: §IV · p. 8
  • Indirect data-driven control first builds a model from real-system data and then uses it for controller design, unlike direct methods that map observations to actions without an explicit model.The tutorial's dynamics-learning discussion spans linear, deterministic nonlinear, and probabilistic nonlinear models.Source: §IV · p. 8

A. Dynamics Learning Problem

Dynamics learning estimates models from measured inputs and states, potentially combining prior knowledge with data. The tutorial reviews linear and deterministic nonlinear model formulations and how model uncertainty and structure affect control synthesis and guarantees.

A. Dynamics Learning Problem

  • The dynamics-learning problem estimates a model from a finite sequence of measured inputs and states, combining available prior knowledge with a learned component.When the dynamics are fully unknown, the nominal component can be set to zero; the augmented state-input vector is fed to both model components.Source: A. Dynamics Learning Problem · p. 8

B. Learning Linear System Models

  • Linear models can approximate nonlinear dynamics locally when approximation error is sufficiently small, while norm-bounded or confidence-set error estimates support robust and probabilistic control methods.Safety filters for unknown linear systems with sub-Gaussian noise can ensure constraint satisfaction with high probability.Source: B. Learning Linear System Models · pp. 8–9
  • Linear model parameters are commonly estimated by least squares, and the resulting uncertainty bounds can inform LQR or MPC synthesis and robust constraint handling.Linear regression has computational complexity linear in sample count, but convergence to true parameters may require substantial data; worst-case sample complexity is exponential in certain cases.Source: B. Learning Linear System Models · pp. 8–9

C. Learning Deterministic Nonlinear Models

  • For strongly nonlinear systems, learned nonlinear models can capture behavior that local linear approximations may miss, and prior structure such as flatness or control-affine form can make learning more efficient.Deterministic approaches include residual models linear in parameters and neural networks; neural ODEs, physics-informed networks, and autoregressive models can incorporate richer structure or temporal dependence.Source: C. Learning Deterministic Nonlinear… · pp. 9–10
  • Neural-network black-box structure makes closed-loop guarantees more difficult, but Lipschitz constraints and temporal stability properties can be used in control-theoretic analyses.Lipschitz properties support Lyapunov and small-gain analyses, while ISS or δ-ISS properties of dynamical models can inform MPC stability and recursive-feasibility guarantees.Source: C. Learning Deterministic Nonlinear… · p. 10

D. Learning Probabilistic Nonlinear Models

Probabilistic nonlinear models represent dynamics uncertainty so controllers can trade performance against confidence in constraint satisfaction. The section reviews Bayesian models, uncertainty propagation, and robust and stability tools, alongside computational and conservatism limitations.

D. Learning Probabilistic Nonlinear Models

  • Chance constraints and risk-sensitive formulations use probabilistic model uncertainty to trade control performance against confidence in constraint satisfaction.Predicted state distributions can be propagated approximately, then constraints tightened according to uncertainty and the selected confidence level.Source: D. Learning Probabilistic Nonlinear… · p. 11
  • GPs model nonlinear dynamics through residuals around nominal predictions, while Bayesian linear regression provides a closed-form Gaussian parameter posterior under Gaussian priors and noise assumptions.For GPs, the kernel specifies function-value covariance; posterior predictions combine the nominal model with the residual posterior mean.Source: D. Learning Probabilistic Nonlinear… · p. 11
  • BLR fitting scales linearly in data size while GP fitting scales cubically, trading computational efficiency against representational flexibility, especially relevant for online adaptation.Reliable uncertainty estimates for both depend on prior structure, including BLR basis functions and GP kernels and hyperparameters.Source: D. Learning Probabilistic Nonlinear… · p. 11
  • Learned probabilistic models make one-step rather than trajectory predictions, complicating closed-loop methods such as MPC that must account for dynamics uncertainty over a horizon.Source: D. Learning Probabilistic Nonlinear… · p. 11
  • GP-MPC can propagate predicted means and covariances and tighten constraints, while robust set-based propagation supports tube-based MPC for constraint satisfaction.Covariance propagation increases optimization complexity; tailored SQP Jacobians and ancillary feedback can reduce computational effort and uncertainty growth, respectively.Source: D. Learning Probabilistic Nonlinear… · pp. 11–12
  • Robust confidence regions can be difficult to obtain or too conservative, and nonlinear uncertainty bounds may also impose conservatism while reliable neural-network uncertainty estimates remain challenging.GP-based MPC stability requires assumptions such as robust positively invariant sets and careful terminal-constraint selection, which can be difficult to construct in practice.Source: D. Learning Probabilistic Nonlinear… · pp. 10, 12

E. Data Collection for Offline Versus Online Learning

Offline dynamics learning relies on informative fixed datasets, while active and online approaches seek informative data during learning or control. Direct data-driven control instead uses measured trajectories to design controllers without an intermediate system-identification step, with Willems’ lemma giving a data-based trajectory representation under controllability and excitation conditions.

E. Data Collection for Offline Versus Online Learning

  • Offline dynamics-learning performance depends on whether a fixed input-output dataset is informative enough; active learning selects informative data using uncertainty, entropy, or Fisher-information criteria.Examples include mutual information, entropy, and D- or A-optimality criteria.Source: E. Data Collection for Offline Versus… · p. 12
  • Online learning adapts models during closed-loop control, and dual control balances task performance with information gathering through probing inputs.Data dictionaries with sparse GPs, Kalman-filter updates, spatio-temporal GPs, and event-triggered learning can address the cost of updating at every control step.Source: E. Data Collection for Offline Versus… · pp. 12–13
  • Without ancillary feedback, predicted uncertainty grows more rapidly and requires more restrictive constraint tightening in the car-terrain learning-based MPC example.Figure 7 also plots closed-loop cost against constraint violation for different β choices.Source: E. Data Collection for Offline Versus… · p. 13, Figure 7 · p. 13

V. DIRECT DATA-DRIVEN CONTROL

  • Direct data-driven control designs a controller directly from data, avoiding the intermediate system-identification step used by model-based control.Source: §V · p. 13

A. Willems’ Fundamental Lemma

  • Willems’ fundamental lemma represents every trajectory of a controllable discrete-time LTI system using one sufficiently persistently exciting measured input-output trajectory.The input must be persistently exciting of order L + n_x for the stated representation.Source: A. Willems’ Fundamental Lemma · p. 14

B. Extensions to the Fundamental Lemma

Extensions to Willems’ fundamental lemma address restrictions on excitation, controllability, and trajectory length, and broaden data-based representations to other system classes. These extensions relax or adapt conditions needed to use measured data for system analysis and control.

B. Extensions to the Fundamental Lemma

  • Generalized persistence-of-excitation conditions provide a data-based representation without requiring the original lemma’s persistently exciting input on a controllable system.In, the representation holds if and only if the stated rank condition holds for L greater than the system lag or observability index.Source: B. Extensions to the Fundamental Lemma · p. 14
  • Collective persistence of excitation combines multiple input trajectories, whose concatenated Hankel matrices meet the rank condition even when no individual input is persistently exciting.This extension is useful when collecting one sufficiently long trajectory is inconvenient, such as for an open-loop unstable system or intermittent data loss.Source: B. Extensions to the Fundamental Lemma · p. 14
  • Extensions cover time-varying, parameter-varying, descriptor, multiagent, stochastic, noisy-data, and irregularly measured settings, while nonlinear representations often use known nonlinear terms or basis dictionaries.Source: B. Extensions to the Fundamental Lemma · p. 15

C. Data-Based Control Design

Data-based control uses collected input and state or output data to construct stabilizing controllers and predictive control schemes without requiring a known system model. DDPC replaces the model in MPC with a data-driven predictor, with regularization addressing noisy data and ongoing work extending guarantees and methods to nonlinear systems.

C. Data-Based Control Design

  • When the data satisfy the rank condition, the LMI parameterization yields a stabilizing state-feedback gain, and every stabilizing gain can be represented this way.The controller can be designed without knowledge of the system model; persistently exciting input of order n_x + 1 can guarantee the condition.Source: C. Data-Based Control Design · p. 15
  • For general controllable linear systems, an input-output-data-derived state enables dynamic output-feedback control using strategies developed for state feedback.The passage contrasts this approach with a previous construction limited to systems where n_y^l = n_x.Source: C. Data-Based Control Design · p. 15

D. Data-Driven Predictive Control

  • DDPC predicts future input and output sequences from data, initializes them with recent measurements, enforces constraints, and applies the first optimized input.Its data-driven prediction constraint replaces the system model used in classical MPC.Source: D. Data-Driven Predictive Control · pp. 15–16
  • For noisy measurements, DDPC adds slack variables and regularization to preserve feasibility and robustify predictions against overfitting.Data and computation demands also motivate variants that reduce required offline samples or optimization variables; the standard minimum sample count scales linearly with prediction horizon L.Source: D. Data-Driven Predictive Control · p. 16
  • DDPC closed-loop stability, safety, and robustness can be studied with MPC tools such as terminal constraints or longer prediction horizons, but noisy data complicate the analysis.Extensions for nonlinear systems remain an active challenge, with existing work using system-class-specific extensions of Willems’ lemma, basis functions, or online local affine approximations.Source: D. Data-Driven Predictive Control · p. 16

VI. REINFORCEMENT LEARNING, IMITATION LEARNING, AND BAYESIAN OPTIMIZATION

This section reviews direct policy learning and data-based predictive control, then describes model-free reinforcement learning approaches that learn through environment interaction. In the flat-terrain DeePC example, reaching the prescribed data-length bound matches exact-model MPC behavior, while actor-critic methods are data hungry and lack safety guarantees.

VI. REINFORCEMENT LEARNING, IMITATION LEARNING, AND BAYESIAN OPTIMIZATION

  • The section covers direct policy learning from data, often through interaction, and illustrates DeePC using random offline inputs and position measurements for a goal-reaching task.The terrain example uses L = 20 and T_ini = 4.Source: §VI · pp. 16–17
  • N ≥51 is required for the PE condition in the flat-terrain LTI example; at that bound, DeePC matches exact-model MPC behavior, with no further performance gain from more data.Below the bound, the optimization problem is infeasible.Source: §VI · p. 17, Figure 8 · p. 17
  • With bumpy terrain and measurement noise, more data are needed to stabilize DeePC, a nonzero error remains, and suitable regularization brings performance close to MPC with the imperfect flat-terrain model.Regularization balances fitting measured history against biasing the predictor.Source: A. Model-Free Reinforcement Learning · p. 17, Figure 9 · p. 17

A. Model-Free Reinforcement Learning

  • Model-free RL learns by trying actions and observing transitions and rewards, with policy-based, value-based, and actor-critic methods differing in what they learn and how they search.Policy gradients estimate the return gradient from trajectories using log-policy gradients, without requiring the unknown system dynamics.Source: A. Model-Free Reinforcement Learning · p. 18
  • Actor-critic methods combine a policy and value estimate, but are typically data hungry and provide no safety guarantees.Their data-efficiency limitations motivate model-based RL; standard online RL requires active environment sampling, whereas offline RL learns from fixed data and faces distribution shift.Source: A. Model-Free Reinforcement Learning · p. 18

B. Model-Based Reinforcement Learning

Model-based RL learns a dynamics model to support planning and policy improvement, while safety filters can constrain exploration. Imitation learning instead learns from expert behavior, including policies that approximate computationally expensive MPC controllers.

B. Model-Based Reinforcement Learning

  • A learned dynamics model predicts future state transitions and enables planning and policy improvement in model-based RL.Dyna-style methods use the model to generate additional training data, while decision-time planning uses it for online look-ahead action selection.Source: B. Model-Based Reinforcement Learning · pp. 18–19
  • In the cart-pole example, DQN learns to swing up and balance the pole, but training performance can be stochastic and requires many environment interactions.The epsilon-greedy strategy sometimes selects random actions, which can lead to poor trajectories even as exploration helps discover better actions.Source: B. Model-Based Reinforcement Learning · p. 19
  • A CBF-based safety filter can modify potentially unsafe DQN actions; the filtered agent satisfies the cart-position constraint and reaches high returns with fewer interactions.The example imposes p ∈ [−1, 2.4].Source: B. Model-Based Reinforcement Learning · p. 19, Figure 11 · p. 19

C. Imitation Learning

  • Imitation learning uses expert behavior rather than reward-driven trial and error; behavior cloning fits a policy to state-action pairs, while inverse RL learns an underlying reward.The expert may be human or a machine, including an expensive-to-evaluate policy.Source: C. Imitation Learning · p. 20
  • Imitating MPC with a neural policy can reduce deployment evaluation time to constant runtime, but hardware applications remain few.Open research questions include constraint satisfaction under approximation errors and adapting policies to parameter changes.Source: C. Imitation Learning · p. 20

D. Controller Tuning with Bayesian Optimization

Bayesian optimization tunes low-dimensional controller parameters from costly closed-loop evaluations using a probabilistic surrogate and acquisition function. Across applications and outlook, the tutorial emphasizes combining learning’s flexibility with control structure for reliable autonomy while identifying scaling, benchmarking, and reproducibility challenges.

D. Controller Tuning with Bayesian Optimization

  • BO is attractive for tuning controllers directly on physical systems because its sample efficiency and black-box formulation suit costly evaluations and unavailable accurate simulators.Controller tuning typically involves only a few to several tens of parameters, enabling specialized methods that may require less data than general-purpose deep RL.Source: D. Controller Tuning with Bayesian… · p. 20
  • A probabilistic surrogate models the unknown closed-loop cost, while an acquisition function selects the next parameter by balancing exploration and exploitation.The surrogate is usually a Gaussian process, though other probabilistic models such as neural networks can be used.Source: D. Controller Tuning with Bayesian… · p. 20

VII. APPLICATIONS

  • By exploiting prior control structure and tuning one parameter, BO found a near-optimal value after five iterations, illustrating efficient optimization with few system experiments.Figure 12 depicts the GP surrogate of cost over controller parameters and the expected-improvement acquisition function.Source: §VII · p. 21, Figure 12 · p. 21
  • Learning-based control has been applied across healthcare, robotics, energy, process control, and networked systems to improve prediction, adaptation, and performance while preserving desired closed-loop properties.Representative examples include medication tapering, autonomous driving, quadrotor control, manipulation, multiagent coordination, power electronics, and wireless sensor networks.Source: §VII · p. 21
  • Robotics exposes challenges including uncertain nonlinear dynamics, limited data, real-time computation, perception and actuation uncertainty, and strict safety requirements.Learning methods offer flexible policy optimization and dynamics modeling, while control methods exploit system structure to improve data efficiency and support closed-loop analysis.Source: §VII · p. 21
  • Recent directions extend control-guided learning to humanoids, multimodal MPC policies, and safety specifications beyond geometric constraints.Examples include safety filtering during quadrotor RL, CLF-shaped rewards for humanoid locomotion, and diffusion models representing multimodal MPC solutions.Source: §VII · pp. 21–22

VIII. CONCLUSIONS AND OUTLOOK

  • The tutorial’s central synthesis is that learning expands modeling, optimization, and adaptation, while control supplies structure for stability, safety, robustness, and reliable closed-loop behavior.The conclusion argues that incorporating data-based methods into the closed loop will be critical as systems enter more complex, uncertain, and safety-critical real-world settings.Source: §VIII · p. 22
  • 38 of 182 invited-session papers (20.9%) released open-source code, indicating substantial room to improve reproducibility and cross-community progress.The paper calls for deployment-relevant benchmarks comparing task performance, sample complexity, prior knowledge, inference time, robustness, safety, and generalization.Source: Figure 13 · p. 22, §VIII · pp. 22–23

APPENDIX

The appendix describes how the tutorial assembled its paper dataset and analyzed concept coverage, code availability, and related-venue information.

APPENDIX

  • The dataset includes 182 CDC Learning-Based Control invited-session papers from 2016–2025, with paper lists and affiliations from ETH Zurich’s session index.Abstracts came from OpenAlex and citation counts from Google Scholar.Source: APPENDIX · p. 30
  • For Figure 2, the authors searched titles and abstracts for 25 LLM-selected domain phrases, then manually checked and consolidated semantically related phrases.Reinforcement learning, deep reinforcement learning, and safe reinforcement learning were counted jointly.Source: APPENDIX · p. 30
  • For Figure 13, code availability was assessed by finding repository links and GitHub repositories referencing each paper’s arXiv ID, followed by a manual repository check.Related-venue information for Figure 4 came from DBLP, and required information was obtained in May 2026.Source: APPENDIX · p. 30
Loading 2610.29187v1…