Pith. sign in

REVIEW 4 major objections 6 minor 20 references

Vision-based Goal-Reaching Control for Mobile Robots Using a Hierarchical Learning Framework

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A hierarchical framework combining visual SLAM, reinforcement-learning motion planning, and robust adaptive actuator control lets a 6,000 kg skid-steered robot reach goals with roughly 3–4 cm final-position error, even on loose soil.

desk verdict A genuine heavy-robot integration whose stability guarantee doesn't survive the algebra in Eqs. (23)–(25), and whose cm-level accuracy is measured against the same SLAM used for control. read the letter →

arxiv 2601.00610 v2 pith:OTC6N7LM submitted 2026-01-02 cs.RO

Mehdi Heydari Shahna , Pauli Mustalahti , Jouni Mattila This is my paper
classification cs.RO
keywords hierarchicalcontrolgoal-reachingreinforcementlearningrobustadaptivevisualSLAMskid-steeredrobotlogarithmicbarrierin-wheelactuator
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a hierarchical decomposition can make reinforcement learning safe and accurate for large-scale mobile robots. The upper layer is a tabular Q-learning/SARSA planner that takes visual pose estimates and outputs smooth acceleration commands; the lower layer is a scalar conjugate-gradient-trained neural network approximating each in-wheel actuator's inverse dynamics, combined with a logarithmic-barrier-based robust adaptive controller. The authors prove that the wheel-tracking error is uniformly ultimately bounded with exponential convergence to a residual set, and that the barrier supervisor guarantees safety by stopping the robot and returning it to a safe area on fault detection. Experiments on a 6,000 kg vehicle on asphalt and loose soil report final-position RMSE near 3–4 cm and successful recovery after injected SLAM faults. The significance is that centimeter-level goal-reaching on a heavy, slip-prone platform would otherwise be difficult for exploration-based RL alone.

What carries the argument

The key mechanisms are the layered control stack and its two learning components. The RL motion planner discretizes distance and heading into bins, acts on linear and angular accelerations, and uses reward shaping to enforce smooth, non-oscillatory approach. At the actuator level, a fully connected deep network trained with scaled conjugate gradient maps wheel speed to nominal control input, approximating inverse actuator dynamics; this feedforward map is combined with a robust adaptive control term containing a logarithmic barrier function of the robot's pose error. The barrier's singularity at the safety boundary triggers the supervisor's safe-return mode, while the Lyapunov argument for t

What would settle it

Re-analyze the logged RTK-INS data from the asphalt and loose-soil runs and compute each goal's final-position error against those ground-truth values; if the resulting RMSE exceeds the reported 3–4 cm by a substantial margin, the accuracy claim would be an artifact of using SLAM as both feedback and reference. A second decisive test: rerun the same controller with RTK-INS pose as the only feedback, and check whether the final-position RMSE stays near 3–4 cm.

Watch

Extended reading notes

Core claim

The central discovery is that separating planning from actuation allows RL to operate on a simple kinematic model while a stability-guaranteed controller handles the complex actuator dynamics and slip. Concretely, the paper establishes exponential convergence of wheel-speed tracking errors to a disturbance-dependent residual set, and reports final-position RMSE of approximately 0.0317 m on asphalt and 0.0382 m on loose soil using a 6,000 kg skid-steered robot, with the safety supervisor returning the robot to a safe area after a fault is injected into the SLAM output.

Load-bearing premise

Everything in the reported accuracy and safety rests on the visual SLAM pose being close to true position, yet the paper never verifies the SLAM estimates against the logged RTK-INS, so if SLAM drifts by a few centimeters the headline error could be estimator bias rather than control performance.

Editorial extensions

If this is right

  • Centimeter-level goal-reaching can be achieved on large skid-steered robots in off-road conditions without GNSS, relying only on stereo vision for pose.
  • Exploration-based RL need not directly command the physical actuators; a kinematic RL planner plus a robust actuator controller can preserve safety and stability.
  • The wheel-tracking error converges exponentially to a residual set whose size scales with the disturbance bound, meaning predictable performance degradation under unmodeled slip.
  • The logarithmic safety supervisor can detect injected localization faults and autonomously return the robot to a safe inspection area, reducing the need for human intervention.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the reported RMSE is computed from the same SLAM pose estimates used as feedback, a fair reader should treat the 3–4 cm accuracy as unvalidated against independent ground truth until compared with the logged RTK-INS data.
  • The modular design implies that either layer can be swapped independently: a different planner or a different actuator model could be retrained while retaining the stability proof, provided the assumptions on boundedness and Lipschitz continuity hold.
  • A direct extension would be online adaptation of the safety barrier to arbitrary obstacle layouts, since the current logarithmic barrier is defined by a single circle around the initial and goal poses.
  • The stability analysis covers wheel-speed tracking, not the full pose-tracking or safety of the overall navigation; the latter is argued through the supervisor rather than proved.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a hierarchical goal-reaching control framework for a 6,000 kg skid-steered mobile robot. The architecture combines ORB-SLAM3 stereo visual localization, a tabular RL motion planner (Q-learning/SARSA with reward shaping and motion constraints), an SCG-trained DNN that maps wheel speed to a feedforward actuator command, a logarithmic-barrier robust adaptive controller (RAC), and a safety supervisor that initiates a safe-return mode. The authors claim uniform exponential stability/UUB of the actuator-level wheel-tracking system and safety of the whole operation. Experiments on asphalt and loose-soil terrains report roughly 3--4 cm final-position RMSE and successful recovery after fault injection.

Significance. If correct, the paper would be a valuable integration of learning-based planning with stability-certified low-level control on a realistic heavy platform. The experimental platform is substantial: real experiments on a 6-ton vehicle over two terrains, with detailed DNN and RL configuration tables. The adaptive control law is not circular in the sense that the adaptive term is not fitted to the experimental outcome; the formal analysis is intended to show robustness to bounded uncertainty. These are real strengths. However, the central formal guarantee is compromised by an algebraic error in the error dynamics, the DNN as described does not implement the feedforward term used in the proof, and the headline experimental accuracy is not compared to the logged RTK ground truth. The significance of the claimed results cannot be assessed until these load-bearing issues are resolved.

major comments (4)
  1. [§IV-C, Eqs. (23)–(25)] The error dynamics are derived with an incorrect sign. Substituting u_IDM_i = F_i(v_di,t) - A_i \dot v_di (Eq. 23) into A_i[\dot v_di + \dot e_i] = u_i + F_i + d_i (Eq. 24) gives A_i \dot e_i = u_i + u_IDM_i + d_i, not u_i - u_IDM_i + d_i. Consequently, Eq. (31)'s cancellation e_i(u_IDM_i - u_IDM_i) is spurious. With u_i = u_IDM_i + u_f, the closed-loop error dynamics contain the un-bounded term 2u_IDM_i, which the Lyapunov analysis in Eqs. (32)–(34) never controls. The claimed UUB/exponential stability of the wheel-tracking subsystem is therefore not established as written. This is load-bearing because the abstract and Section IV-C rest their formal guarantee on this derivation.
  2. [§IV-A, Eq. (18) vs. Eq. (23)] The DNN is a static scalar map f_β(v_i) taking only the current wheel speed as input, but the ideal inverse model in Eq. (23) requires the term -A_i \dot v_di, which depends on the desired acceleration. A static map from v_i alone cannot realize this term. If the trained DNN is only an approximation, then the approximation error should be absorbed into the uncertainty d_i and treated explicitly in the Lyapunov analysis; the current proof assumes that u_IDM_i is exactly available as defined in Eq. (23). This mismatch makes the connection between the implemented learning component and the formal stability guarantee unclear.
  3. [§V, Tables III–IV and Fig. 5] The reported final-position RMSE is computed from ORB-SLAM3 pose estimates, which are the same estimates used as control feedback. The platform logs a Trimble BD992 INS-RTK, but no comparison to this independent ground truth is reported. If SLAM drift or localization bias is on the order of the claimed 3–4 cm, the headline accuracy reflects estimator error rather than control accuracy. In addition, each goal is traversed only once, with no repeated trials, error bars, or statistical measures. The comparison between asphalt and loose-soil performance is therefore currently anecdotal. Please report repeated runs and the RTK-based final-position errors.
  4. [§V, fault-injection and safety claim] The safety supervisor is validated only by injecting an external signal into the SLAM output; the manuscript does not specify the fault-detection logic, the supervisor state machine, or the braking/return algorithm. The logarithmic barrier in Eqs. (27)–(28) only imposes E(t) < O during nominal operation; it does not by itself detect localization inconsistencies or guarantee safe return. The claim that the framework 'guarantees safety of the whole operation' is therefore stronger than the provided evidence.
minor comments (6)
  1. [Eq. (35)] The constant ℓ is defined using only d_R^* and d_L^*, but the Lyapunov sum runs over i = 1,...,4. It should be a sum over all wheels with the appropriate κ_i.
  2. [§IV-B] Typos: 'unknwon' and 'contants' should be 'unknown' and 'constants'.
  3. [Eq. (28)] E(t) is defined as distance to the midpoint between the start and goal, not a 'robot pose error'. This terminology is confusing and should be clarified.
  4. [References] References [2] and [20] appear to be the same paper; unify them.
  5. [Table I] The distance binning in Eq. (3) uses N_d bins over [0, d_max], but d_max and N_d are not specified in Table I. The grid resolution [1.0, 1.0] m also seems coarse relative to the cited 3–4 cm final accuracy; this deserves a comment.
  6. [Conclusion] The claim of being 'the first comprehensive control framework' should either be supported by a detailed comparison with prior integrated hierarchical systems or toned down.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: Lyapunov/adaptive design is a standard construction, not a fitted prediction; sign inconsistency and SLAM-only evaluation are correctness/validation issues, not circular reductions.

full rationale

The main claimed derivation (Sections IV-B and IV-C) is not circular. Equations (23), (24), (26), and (27) form a standard feedforward-plus-adaptive Lyapunov design: the log-barrier adaptive term is selected so that its product with e_i^2 log(...) cancels in dot(V), and boundedness/locally-Lipschitz assumptions yield a UUB inequality. This is a design that produces the stability property, not a property assumed as an input. The DNN is trained on recorded (v_i,u_i) data and used as u_IDM; the robust/adaptive feedback is designed to handle the residual, so the comparison in Table V is not a renamed fit. Self-citations to [19] and [20] are used for the log-barrier construction and for Definition 1 of the stability notion; both are explicitly written in this paper or are standard definitions, so they are not load-bearing circular support. Equation (13) is a direct algebraic identity for potential-based shaping and is explicitly labeled as well-known, not a renaming. There are, however, serious non-circular flaws: substituting (23) into (24) gives A_i * dot(e_i) = u_i + u_IDM + d_i, not the u_i - u_IDM + d_i in (25); since u_i in (26) includes u_IDM as a positive feedforward, the claimed cancellation does not occur as written. In addition, final-position RMSE (Tables III-IV) is computed from ORB-SLAM3 estimates, the same pose signal used for control feedback, while the logged Trimble RTK-INS ground truth is never compared; this weakens the experimental claim but is a validation limitation rather than a by-construction equivalence. These issues lower confidence in the proof and experiments but do not make the derivation circular; the circularity score is low.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The central demonstration rests on many hand-picked weights, discretization choices, and control gains, plus strong assumptions about the actuator model, the learned policy's transfer, and the accuracy of SLAM as ground truth. The formal proof covers the actuator loop; the experimental headline depends on unvalidated SLAM-as-truth.

free parameters (10)
  • Reward shaping weights (14 gains) = k_d=5.0, k_θ=6.0, k_v=0.08, k_lat=0.9, k_ω=0.28, k_av=0.10, k_aω=0.10, k_step=0.04, k_timeout=3.0, k_ωflip=0.85, k_head,
    Chosen by hand in Section V; RL behavior (smoothness, oscillation suppression, approach angle) depends critically on these weights, and no sensitivity study is reported.
  • RL state discretization and limits = goalTol=0.10 m, grid 1.0 m, nθ=24, Nv=4, Nω=5, v∈[0,0.25], ω∈[−0.15,0.15], a_v∈[−0.10,0.10], a_ω∈[−0.02,0.02]
    Hand-selected bins and limits in Table I; the Q-table's resolution and coverage depend on them.
  • RL exploration schedule = ε_0=1.0, ε_final=1e-3, episodes=30000
    Training schedule chosen manually; no ablation is provided.
  • Learning rate and discount factor = α=0.10, γ=0.95
    Default values; no sensitivity analysis.
  • RAC gains = ε_i=1, γ_i=0.01, δ_i=0.2
    Chosen in Section V; stability only requires ε_i>κ_i, but κ_i is unspecified.
  • Safety offset and barrier constants = ζ∈R+ (value not given), O defined as safety circle
    The safety supervisor's trigger radius depends on an unspecified safety offset ζ.
  • DNN architecture and hyperparameters = hidden sizes [320,210,105], train/validation/test 0.34/0.33/0.33, goal 1e-6, min grad 1e-10, max epochs 500
    Design choices; no architecture search or cross-validation beyond early stopping.
  • ORB-SLAM3 tuning = ~2000 features/frame, FAST threshold 12 px, stereo depth limit 40 m
    Hand-tuned offline parameters; no evaluation of localization accuracy against RTK is reported.
  • Reward deadbands and hysteresis windows = e_db=0.01 rad, ω_db=0.001 rad/s, e_lock=0.03 rad, d_lock=0.30 m, k_ws=1.2
    Small shaping windows that strongly affect final-approach behavior; values chosen by hand.
  • Lyapunov constants κ_i = unspecified, must satisfy ε_i>κ_i
    Used in the Cauchy–Schwarz bound; no values or selection procedure are given.
assumptions (6)
  • domain assumption Assumption IV.1: control gain A_i and mappings F_i and d_i are bounded and locally Lipschitz continuous.
    Needed for the tracking-error dynamics and the Lyapunov bound; not verified from actuator data.
  • domain assumption Ideal inverse-dynamics model (Eq. 23): u_IDMi = F_i(v_di,t) - A_i ẍ_di(t) represents the actuator under no disturbance.
    The DNN is trained to approximate this map, and the stability proof relies on the residual being absorbed into d_i.
  • domain assumption The deterministic MDP transition (Eq. 5) with binning and saturation accurately models robot kinematics.
    RL is trained in this discretized simulation; if the kinematic model or binning loses information, the learned Q-table may not transfer to the real robot.
  • domain assumption The non-potential reward terms (rewards 4–12) lead to a policy that is still desirable, and tabular Q-learning/SARSA converges to a good policy under ε-greedy exploration.
    The paper explicitly notes that these terms change the underlying objective, so no optimality-preservation guarantee applies to the full shaped reward.
  • domain assumption ORB-SLAM3 stereo pose estimates are accurate enough to serve as both controller feedback and evaluation truth.
    No comparison to the logged Trimble RTK-INS is presented; the reported RMSE is SLAM-based.
  • domain assumption The logarithmic-barrier supervisor detects all unsafe conditions that matter and can always return the robot to the safe area.
    The supervisor is a state machine with braking profiles; its completeness and robustness are not formally analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision-based Goal-Reaching Control for Mobile Robots Using a Hierarchical Learning Framework." pith.science (2026). https://pith.science/paper/OTC6N7LM

@misc{pith2026260100610,
  author       = {Pith},
  title        = {Pith review of: Vision-based Goal-Reaching Control for Mobile Robots Using a Hierarchical Learning Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTC6N7LM}},
  note         = {Machine review of arXiv:2601.00610}
}
read the original abstract

Reinforcement learning (RL) has strong potential in robotics, but exploration-based training complicates safe deployment on large-scale robots. For such applications, this paper proposes a novel hierarchical goal-reaching framework that integrates stereo visual pose estimation, constrained RL-based motion planning, actuator-level robust adaptive control (RAC), and supervisory safe-return logic. Stereo visual localization is used as the real-time pose-estimation interface with loop closing, map fusion, and relocalization. The RL planner generates smooth, feasible goal-reaching references using a problem-specific reward structure and motion constraints that promote goal progress, reduce oscillations, preserve vision-consistent smoothness, and respect the mechanical limits of a heavy skid-steered robot. At the actuation layer, a scaled conjugate-gradient (SCG)-trained deep neural network (DNN) approximates a quasi-static actuator feedforward map from wheel-speed data to nominal control input. This feedforward map is combined with a logarithmic-barrier-based RAC to compensate for residual modeling errors, slip-induced disturbances, and bounded mismatch between the nominal map and real actuator response. For the actuator-level wheel-tracking subsystem, uniformly ultimately bounded tracking with exponential convergence to a disturbance-dependent residual set is established under bounded uncertainty. A logarithmic safety supervisor monitors execution, detects unsafe operating conditions, including faults and localization inconsistencies, and switches the robot to safe-return mode. Experiments on a 6000 kg robot over asphalt and loose-soil terrain demonstrate approximately 3--4 cm final-position root mean square error (RMSE), accurate tracking of RL-generated commands, improved actuator-level performance over two RAC baselines, and successful autonomous recovery after fault injection.

Figures

Figures reproduced from arXiv: 2601.00610 by the authors.

Figure 1
Figure 1. Architecture of the RL motion planning [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. DNN training an inverse dynamics model of an actuator system used in many large-scale robots. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The proposed framework in real time. Taking the derivative of (29) and using (25), we have ¯˙V = X 4 i=1 ei [ui − uIDMi + di ] + ˆχi ˙χˆi (30) Substituting (26), we get ¯˙V = X 4 i=1 − 1 2 ϵie 2 i − γie 2 i log2 ( O O − E ) ˆχi + eidi + ei(uIDMi − uIDMi ) + ˆχi ˙χˆi (31) In view of Assumption IV.1, introduce an unknown nonnegative constant d ∗ i satisfying |di | ≤ d ∗ i . By substituting (27), ¯˙V ≤ X 4 i=1 − 1 2 ϵi… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Goal reaching performance (meter). In both terrains, the goal-reaching accuracy is very similar, with a position Root mean square error (RMSE) of approx￾imately 3 cm. Achieving nearly the same accuracy on low￾slip asphalt and highly slippery loose soil shows that the c…
Figure 6
Figure 6. Figure 6: Robot tracking motion performance. TABLE V: Comparison of different RAC approaches. Control method Peak time (s) Maximum overshoot (m/s) Settling time (s) Steady-state error (m/s) DNN-based RAC 4.100 0.028 4.850 0.009 Backstepping RAC [21] 5.200 0.051 4.800 0.019 Model…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 3 linked inside Pith

  1. [1]

    Adaptive heading correction for an industrial heavy-duty omnidirectional robot,

    R. Galati, G. Mantriota, and G. Reina, “Adaptive heading correction for an industrial heavy-duty omnidirectional robot,”Scientific Reports, vol. 12, no. 1, p. 19608, 2022

    2022

  2. [3]

    A multi- plicative value function for safe and efficient reinforcement learning,

    N. B ¨uhrer, Z. Zhang, A. Liniger, F. Yu, and L. Van Gool, “A multi- plicative value function for safe and efficient reinforcement learning,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 5582–5589

    2023

  3. [4]

    Lyapunov-based safe policy optimization for continuous control,

    Y . Chow, O. Nachum, A. Faust, E. Duenez-Guzman, and M. Ghavamzadeh, “Lyapunov-based safe policy optimization for continuous control,”arXiv preprint arXiv:1901.10031, 2019

    arXiv 1901

  4. [5]

    A novel agent with formal goal-reaching guarantees: an experimental study with a mobile robot,

    G. Yaremenko, D. Dobriborsci, R. Zashchitin, R. C. Maestre, N. Q. H. Hoang, and P. Osinenko, “A novel agent with formal goal-reaching guarantees: an experimental study with a mobile robot,”arXiv preprint arXiv:2409.14867, 2024

    arXiv 2024

  5. [6]

    Safe reinforcement learning for probabilistic reachability and safety specifications: A lyapunov-based approach,

    S. Huh and I. Yang, “Safe reinforcement learning for probabilistic reachability and safety specifications: A lyapunov-based approach,” arXiv preprint arXiv:2002.10126, 2020

    arXiv 2002

  6. [7]

    Actor-critic reinforcement learning for control with stability guarantee,

    M. Han, L. Zhang, J. Wang, and W. Pan, “Actor-critic reinforcement learning for control with stability guarantee,”IEEE Robotics and Au- tomation Letters, vol. 5, no. 4, pp. 6217–6224, 2020

    2020

  7. [8]

    Safe reinforcement learning using robust control barrier functions,

    Y . Emam, G. Notomista, P. Glotfelter, Z. Kira, and M. Egerstedt, “Safe reinforcement learning using robust control barrier functions,”IEEE Robotics and Automation Letters, 2022

    2022

  8. [9]

    Whole-body locomo- tion and posture control on a torque-controlled hydraulic rover,

    S.-H. Hyon, Y . Ida, J. Ishikawa, and M. Hiraoka, “Whole-body locomo- tion and posture control on a torque-controlled hydraulic rover,”IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 4587–4594, 2019

    2019

Show all 20 references
  1. [10]

    Integrating DeepRL with robust low-level control in robotic manipulators for non-repetitive reaching tasks,

    M. H. Shahna, S. A. A. Kolagar, and J. Mattila, “Integrating DeepRL with robust low-level control in robotic manipulators for non-repetitive reaching tasks,” in2024 IEEE International Conference on Mechatronics and Automation (ICMA). IEEE, 2024, pp. 329–336

    2024

  2. [11]

    A sequential approach for accurate parameters identification of heavy-duty hydraulic manipulators ensuring physical feasibility,

    W. Huang, Z. Chen, F. Zhang, M. Cheng, R. Ding, J. Zhang, and B. Xu, “A sequential approach for accurate parameters identification of heavy-duty hydraulic manipulators ensuring physical feasibility,”IEEE Robotics and Automation Letters, 2025

    2025

  3. [12]

    Fault-tolerant control for system availability and continuous operation in heavy-duty wheeled mo- bile robots,

    M. H. Shahna, P. Mustalahti, and J. Mattila, “Fault-tolerant control for system availability and continuous operation in heavy-duty wheeled mo- bile robots,” in2025 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), 2025, pp. 1–8

    2025

  4. [13]

    Aivio: Closed-loop, object-relative navigation of uavs with ai- aided visual inertial odometry,

    T. Jantos, M. Scheiber, C. Brommer, E. Allak, S. Weiss, and J. Stein- brener, “Aivio: Closed-loop, object-relative navigation of uavs with ai- aided visual inertial odometry,”IEEE Robotics and Automation Letters, 2024

    2024

  5. [14]

    Dynavins: A visual-inertial slam for dynamic environments,

    S. Song, H. Lim, A. J. Lee, and H. Myung, “Dynavins: A visual-inertial slam for dynamic environments,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11 523–11 530, 2022

    2022

  6. [15]

    An accurate open-source library for visual, visual–inertial, and multimap slam., 2021, 37,

    C. Campos, R. Elvira, J. G. Rodr ´ıguez, and J. Montiel, “An accurate open-source library for visual, visual–inertial, and multimap slam., 2021, 37,” pp. 1874–1890, 2021

    2021

  7. [16]

    The tum vi benchmark for evaluating visual-inertial odometry,

    D. Schubert, T. Goll, N. Demmel, V . Usenko, J. St ¨uckler, and D. Cre- mers, “The tum vi benchmark for evaluating visual-inertial odometry,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 1680–1687

    2018

  8. [17]

    The euroc micro aerial vehicle datasets,

    M. Burri, J. Nikolic, P. Gohl, T. Schneider, J. Rehder, S. Omari, M. W. Achtelik, and R. Siegwart, “The euroc micro aerial vehicle datasets,” The International Journal of Robotics Research, vol. 35, no. 10, pp. 1157–1163, 2016

    2016

  9. [18]

    Fixed-time con- current learning-based robust approximate optimal control,

    J. Tan, S. Xue, T. Niu, K. Qu, H. Cao, and B. Chen, “Fixed-time con- current learning-based robust approximate optimal control,”Nonlinear Dyn., pp. 1–21, 2025

    2025

  10. [19]

    Model reference-based control with guaranteed predefined performance for uncertain strict- feedback systems,

    M. H. Shahna, J.-P. Humaloja, and J. Mattila, “Model reference-based control with guaranteed predefined performance for uncertain strict- feedback systems,”Control Engineering Practice, vol. 164, p. 106516, 2025

    2025

  11. [20]

    Robust torque-observed control with safe input–output constraints for hydraulic in-wheel drive systems in mobile robots,

    M. H. Shahna, P. Mustalahti, and J. Mattila, “Robust torque-observed control with safe input–output constraints for hydraulic in-wheel drive systems in mobile robots,”Control Engineering Practice, vol. 164, p. 106459, 2025

    2025

  12. [21]

    Adaptive prescribed finite time control for strict- feedback systems,

    G. Zuo and Y . Wang, “Adaptive prescribed finite time control for strict- feedback systems,”IEEE Transactions on Automatic Control, vol. 68, no. 9, pp. 5729–5736, 2022

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.