Pith. sign in

REVIEW 4 major objections 6 minor 28 references

Sequence Modeling for Time-Optimal Quadrotor Trajectory Optimization with Sampling-based Robustness Analysis

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A learned LSTM imitates a full-dynamics time-optimal quadrotor planner by predicting only squared-speed and yaw profiles, cutting per-path planning from about 10.7 seconds to 0.078 seconds.

desk verdict A solid imitation-learning approach for time-optimal quadrotor path parameterization with a real speedup and honest hardware validation; the main weakness is an underspecified segment-wise generalization scheme and a few overstated feasibility claims. read the letter →

arxiv 2506.13915 v1 pith:RVBP7RVT submitted 2025-06-16 cs.RO

Katherine Mao , Hongzhan Yu , Ruipeng Zhang , Igor Spasojevic , M Ani Hsieh , Sicun Gao , Vijay Kumar This is my paper · ORCID
classification cs.RO
keywords quadrotortrajectoryplanningtime-optimalpathparametrizationimitationlearningLSTMencoder-decoderdifferentialflatnessbackwardreachabletubedataaugmentationrobustnessanalysis
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

Time-optimal quadrotor trajectories are usually produced by solving a non-convex optimization problem, which takes seconds per path and rules out real-time replanning. This paper tries to show that a sequence model can learn the pattern behind those trajectories and reproduce them almost instantly. The input is the geometric path and its first two derivatives; the output is the squared-speed profile and a cosine-encoded yaw profile, the minimal information needed to reconstruct the full time-optimal state trajectory. In simulation the LSTM encoder-decoder reduces average compute time from 10.656 s to 0.078 s per path while staying close to the optimization-based planner in tracking accuracy, and on a physical CrazyFlie it tracks reference paths with deviations comparable to the classical planner. The paper also introduces a sampling-based robustness analysis that measures whether the predicted trajectory stays inside the backward reachable tube of the low-level controller, and shows that training on randomly perturbed paths improves that robustness.

What carries the argument

The load-bearing object is the squared-speed profile h(·), which encodes how fast the quadrotor moves along the geometric path and therefore determines traversal time and higher-order states through differential flatness. The model predicts h(·) together with cosθz(·), the cosine-encoded yaw, and the full state is reconstructed by finite differences, by aligning the body z-axis with the net thrust direction, and by inverting quaternion kinematics to recover angular velocity. Robustness is analyzed through the finite-time Backward Reachable Tube ξ_U, the set of states from which a target state can be reached within a given time under the low-level controller; the paper estimates membership in it by sampling and simulation, and augments training with randomly perturbed copies of the input paths, governed by the assumption that the original optimal control remains valid for small perturbations.

What would settle it

Take the trained network and run it on straight-line paths of increasing length, since the paper's appendix already reports that one augmented model crashed on a simple straight-line trajectory, and record failure rate and in-BRT probability; if straight paths systematically fail or position deviation exceeds 1 m, the claimed generalization to unseen path geometries does not hold.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the difficult high-dimensional time-optimal path parametrization problem for a quadrotor can be reduced to a sequence-to-sequence imitation task over two scalar profiles. Because a quadrotor is differentially flat, the position path determines most of the state and the planner's output is effectively summarized by the squared speed h(·) along the path and the yaw angle θz(·). An LSTM encoder-decoder trained on 10,000 TOPPQuad solutions learns to map [γ(·), γ′(·), γ″(·)] to [h(·), cosθz(·)] and, after unrolling the trajectory and applying a geometric controller, produces flights that are near-time-optimal with negligible thrust violations and a planning time around 0.078 s per path. The same model runs on a CrazyFlie 2.0 with maximum position deviation 0.355 m versus 0.347 m for TOPPQuad, and the paper reports generalization to path lengths longer than those in the training set via segment-wise prediction.

Load-bearing premise

The load-bearing premise is that a model trained on relatively short, polynomial reference paths keeps producing flyable, near-time-optimal speed and yaw profiles for unseen and much longer paths, and this claim rests in part on a segment-wise prediction scheme whose segment length, overlap, and conditioning state features are not fully specified in the text; it also assumes the perturbation scale in data augmentation stays small enough that the original optimal control sequence remains valid for the perturbed path.

Editorial extensions

If this is right

  • Trajectory replanning for agile quadrotors moves from seconds to under a tenth of a second per path, making online re-computation feasible during flight.
  • Because the learned mapping is the minimal (squared speed, yaw) parametrization, the same design can be retrained for new hardware settings and dynamic limits whenever the vehicle configuration changes.
  • The sampling-based BRT metric gives a principled, measurable notion of dynamic feasibility for a learned trajectory, beyond position error and failure counts.
  • Training with perturbed paths increases output stability and in-BRT probability, at some cost in time-optimality, indicating a tunable robustness-optimality trade-off.
  • On hardware, the learned planner tracks unseen paths with position deviation comparable to the optimization-based planner across eight test paths.

Reading between the lines

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

  • The same 'predict only speed and yaw' reduction would likely transfer to other differentially flat vehicles, since the reconstruction formulas depend on flatness rather than on quadrotor-specific code; testing on fixed-wing or underwater vehicles would show how general the reduction is.
  • The in-BRT membership estimator could be turned from an evaluation metric into a training loss or a safety filter, rejecting predicted profiles that fall outside the backward reachable tube before they are sent to the controller.
  • The unquantified perturbation scale epsilon could be characterized by measuring how the TOPPQuad solution changes under path perturbations; a Lipschitz or sensitivity bound would turn epsilon from a hyperparameter into a certified robustness guarantee.
  • Segment-wise prediction, if specified and released, would likely let the same model handle arbitrarily long paths at bounded compute, which is the key to using this planner in large-scale missions.
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 an imitation-learning framework that approximates TOPPQuad, a model-based time-optimal path parametrization for quadrotors with full dynamics and per-motor actuation constraints. An LSTM encoder-decoder maps a discretized geometric path together with its first and second derivatives to a squared-speed profile and a cosine-encoded yaw profile, from which the full robot state trajectory is reconstructed and executed by a low-level geometric controller. The authors introduce a sampling-based Backward Reachable Tube (BRT) robustness metric, propose a perturbation-based data augmentation scheme, and evaluate the approach in RotorPy simulation and on a CrazyFlie 2.0 hardware platform. They report large compute speedups (from about 10.656 s to 0.078 s per path in simulation), near-zero failure rates, comparable tracking to TOPPQuad, and qualitative generalization to longer hardware trajectories.

Significance. If the claims hold, this is a useful step toward real-time time-optimal quadrotor planning with full dynamic and actuation constraints: the learned planner achieves a two-orders-of-magnitude compute speedup over TOPPQuad while keeping trajectory deviation and thrust violations small, and the hardware experiments add practical credibility. The paper's strengths include public code, an explicit architecture ablation, a concrete robustness evaluation that goes beyond mean tracking error, and a candid limitations section. The BRT-inspired robustness analysis and the noise-injection augmentation are reasonable and potentially transferable to other learning-based planners. However, the central claim of generalization to previously unseen path lengths is supported only by an underspecified segment-wise inference scheme, and several internal inconsistencies affect the strength of the dynamic-feasibility claims; these issues must be resolved before the results can be fully trusted.

major comments (4)
  1. [§5.2, Table 4, Fig. 3] The abstract and contributions claim generalization to previously unseen path lengths, but the only description of the mechanism is one sentence in §5.2 stating that the model predicts a segment of the overall trajectory at a time while conditioning on the robot's state at the start of each segment. The manuscript does not report the segment length, whether segments overlap, which state features are used as conditioning inputs, how the initial state for each segment is obtained, or how this inference-time protocol relates to training on 100-point sequences. Table 4 reports only aggregate max deviation and travel time, and Figure 3 gives no path-length numbers for the extended trajectories, so the reader cannot verify that the test paths actually exceed the training lengths or that the result is not an artifact of the segmentation procedure. Please specify the protocol completely and add quantitative support (path lengths, number of segments, and per-trajectory results) for the generalization claim.
  2. [§4.1, Table 1 and caption] The text in §4.1 states that recovering u(·) from the predicted orientation and accelerations 'ensuring that the per-motor actuation constraints are satisfied,' but Table 1 reports nonzero thrust violations for every learned model (e.g., 0.009 N for the LSTM on the test set) and the caption explicitly says that predicted motor thrusts are 'not bounded by construction.' These statements are in direct conflict. Please clarify whether the trajectory-recovery step enforces the per-motor bounds or whether feasibility is only conferred by the low-level controller, and adjust the wording and the reported violations accordingly.
  3. [§5.2 vs Appendix A.1.2] The hardware experiment section states that the maximum acceleration is 10 m/s² and the maximum angular velocity is 10 rad/s, while Appendix A.1.2 states amax = 5 m/s² and ωmax = 8 rad/s. Because these limits define the dataset and the safety envelope for the hardware flights, the discrepancy must be resolved and a single set of values used.
  4. [Eq. (2)] Equation (2) defines the simulated trajectory update as an integral of the time-derivative dynamics ˙f with respect to the path parameter s over [0, Δs]. Since Δs is a spatial discretization step and ˙f is a time derivative, the expression is dimensionally inconsistent; the integral should be over the traversal time interval t_i introduced in Eq. (4), or the notation must be changed to make the time reparameterization explicit. This is the formal basis for the sampling-based BRT evaluation, so it should be corrected.
minor comments (6)
  1. [Appendix C.1] The appendix reuses 'Table 1' for the input ablation, colliding with the main text's Table 1; renumber the appendix table.
  2. [Various] Typos in Section 1 ('Methodolgy'), Section 5.2 ('expeirments'), Appendix A.1.2 ('expeirments'), and Appendix A.1.1 ('approximatley') should be corrected.
  3. [Table 2] Table 2's layout is ambiguous: the top-level model names (LSTM, LSTM-0.01, LSTM-0.1) are placed alongside a row labeled 'ϵ (perturbation scale)' with values 0.001/0.01/0.1; clarify which columns correspond to training perturbation and which to evaluation perturbation.
  4. [§4.2, Eq. (5)] The BRT probability estimate in Eq. (5) is described as computed by sampling r(s_i) and r̂(s_i) from model predictions; the number of samples and the simulation horizon used for ψ should be stated (Table 2 reports only 10 perturbation sets per trajectory).
  5. [§1, Related Work] The claim in the introduction that this is 'the first work' to develop a learning-based algorithm for this problem is strong; consider softening or supporting it with a more explicit comparison to [12] and [15], which also learn time allocations for quadrotor trajectories.
  6. [Appendix C.3] The statement that the LSTM-0.01 model 'would be more likely to struggle generalizing to a straight line path' is speculative; if this observation is important, report the actual straight-line performance separately.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned outputs are evaluated against their teacher TOPPQuad in a transparent supervised-imitation setup.

full rationale

The paper's central claim is imitation of TOPPQuad, an external optimization-based planner (published IROS 2024) that generates the training target. Training on teacher outputs and evaluating against them is standard supervised learning; the near-match to TOPPQuad is a measure of fit quality, not a derivation that presupposes the learned model's own parameters. The output variables [h, cos(theta_z)] are defined as TOPPQuad's decision variables, not in terms of the LSTM's parameters, and the trajectory recovery uses standard differential-flatness and quaternion kinematics, not a loop back through the learned model. The robustness framework (Propositions 4.1 and 4.3) is a sufficient-condition definition evaluated with an independent simulator and low-level controller, so the in-BRT probability is an empirical measurement rather than a circular construct. The data augmentation scheme under Assumption 4.4 is an explicitly stated modeling assumption: it reuses unperturbed ground-truth labels for perturbed paths, but the paper does not conceal this or claim that the assumption is proven. The hardware segment-wise prediction in Section 5.2 is underspecified, which is a reproducibility gap and a genuine limitation, but no passage exhibits the target result being equivalent to the model's own fitted values by construction. The self-citation to TOPPQuad is load-bearing for the definition of the expert target, but it is an externally published, parameter-free optimizer, and the paper is transparent that it is imitating that planner rather than re-deriving time-optimality from the learned model. The Limitations section explicitly acknowledges retraining requirements, approximate infeasibility, and open perturbation bounds; these are honest limitations and do not convert the supervised fit into circular reasoning.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the simulator model, the differential flatness representation, an unverified robustness assumption for data augmentation, and an empirical sampling approximation of reachability. No new physical entities are introduced. The free parameters are the augmentation scale, the angular acceleration penalty weight, model hyperparameters not reported, and the speed limit.

free parameters (4)
  • Angular acceleration penalty weight lambda = 1e-4
    Added to the TOPPQuad cost function in Appendix A.1 to regularize yaw profile consistency; chosen by hand and affects the ground-truth labels used for training.
  • Perturbation scale epsilon for data augmentation = 0.01 and 0.1 in simulation, 0.001 for hardware
    Used in Section 5.1.2 and 5.2 to generate perturbed training paths. The paper treats epsilon as a tunable hyperparameter and states that no precise bound is derived (Section 4.2.2).
  • Model hyperparameters of the LSTM = Not reported
    Hidden size, number of layers, learning rate, attention mechanism details, and training schedule are not specified in the text, yet the central results depend on them. These are free parameters of the method.
  • Maximum speed constraint v_max = 5 m/s in simulation, 2 m/s on hardware
    Chosen for the task and safety on the CrazyFlie platform; affects the time-optimal profiles and is set by the experimenter.
assumptions (6)
  • standard math Quadrotor differential flatness with position and yaw as flat outputs
    Used throughout Section 4.1 and Appendix B to reconstruct the full state from the learned h and cos(theta_z) profiles. This is a standard result from Mellinger and Kumar and is accepted as background.
  • domain assumption RotorPy simulator with CrazyFlie 2.0 parameters accurately models the real quadrotor dynamics
    All simulation results and training data rely on this model. Hardware experiments show some sim-to-real gap (e.g., settling at the goal), so the assumption is only approximate.
  • domain assumption The sampling-based Backward Reachable Tube approximation provides a reliable estimate of true reachability
    Section 4.2 replaces the true BRT with xi_c obtained by sampling and simulating the controller. No convergence or coverage guarantees are given, so the in-BRT probability is an empirical estimate.
  • ad hoc to paper Assumption 4.4: the control sequence optimal for the unperturbed path gamma remains epsilon-robust for perturbed paths gamma_hat
    Introduced in Section 4.2.2 to justify training on perturbed paths with the same ground-truth outputs. The paper acknowledges it is not verified and limits the permissible perturbation scale.
  • domain assumption The modified TOPPQuad with an angular acceleration penalty yields consistent yaw profiles across small path perturbations
    Appendix A.1 argues the penalty clusters yaw differences near zero based on 100 trajectory pairs. This underpins the data augmentation scheme and the yaw recovery assumption in Appendix B.
  • domain assumption The low-level geometric controller tracks the planned trajectory closely enough that simulated states r_hat approximate the planned r
    The error-tracking controller U is used to integrate r_hat in Eq. (2). If tracking is poor, the BRT condition in Proposition 4.1 may not reflect actual feasibility, and the reported position deviations indicate this is only approximately true.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sequence Modeling for Time-Optimal Quadrotor Trajectory Optimization with Sampling-based Robustness Analysis." pith.science (2026). https://pith.science/paper/RVBP7RVT

@misc{pith2026250613915,
  author       = {Pith},
  title        = {Pith review of: Sequence Modeling for Time-Optimal Quadrotor Trajectory Optimization with Sampling-based Robustness Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RVBP7RVT}},
  note         = {Machine review of arXiv:2506.13915}
}
read the original abstract

Time-optimal trajectories drive quadrotors to their dynamic limits, but computing such trajectories involves solving non-convex problems via iterative nonlinear optimization, making them prohibitively costly for real-time applications. In this work, we investigate learning-based models that imitate a model-based time-optimal trajectory planner to accelerate trajectory generation. Given a dataset of collision-free geometric paths, we show that modeling architectures can effectively learn the patterns underlying time-optimal trajectories. We introduce a quantitative framework to analyze local analytic properties of the learned models, and link them to the Backward Reachable Tube of the geometric tracking controller. To enhance robustness, we propose a data augmentation scheme that applies random perturbations to the input paths. Compared to classical planners, our method achieves substantial speedups, and we validate its real-time feasibility on a hardware quadrotor platform. Experiments demonstrate that the learned models generalize to previously unseen path lengths. The code for our approach can be found here: https://github.com/maokat12/lbTOPPQuad

Figures

Figures reproduced from arXiv: 2506.13915 by the authors.

Figure 1
Figure 1. We uniformly discretize a geometric path γ(·) from a sequence of waypoints using the minimum snap planner [1]. Based on this discretized trajectory, our LSTM encoder-decoder model (which performed the best in our ablation study) predicts squared-speed h(·) and yaw θz profiles, imitating the model-based time￾optimal planner TOPPQuad [2]. From these predicted profiles, we unroll the full robot state trajectory. Finall… view at source ↗
Figure 2
Figure 2. Experiment visualization, plotted from the collected motion capture flight data. The dashed black line represents the reference geometric path, the dashed blue line shows the tracked trajectory generated by TOPPQuad, and the solid purple lines show the tracked trajectories output by our model across different runs. 6 Conclusion In this work, we propose an imitation-learning framework for time-optimal quadrotor traje… view at source ↗
Figure 3
Figure 3. a) Partial depiction of Traj 9. b) Experiment Visualization, showcasing good tracking of geometric paths whose lengths exceede what is seen in the dataset 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 1
Figure 1. Figure 1: Maximum difference of h, time, and yaw between 100 pairs of perturbed and unperturbed [PITH_FULL_IMAGE:figures/full_fig_p012_1.png]
Figure 2
Figure 2. Figure 2: Sample paths generated by our planner (left), MFBO (middle) and AllocNet (right). We also show the bounding boxes given to the AllocNet planner. 5 [PITH_FULL_IMAGE:figures/full_fig_p015_2.png]
Figure 3
Figure 3. Figure 3: Experiment visualization of trajectories predicted from the LSTM-0.1 model, plotted from the col￾lected motion capture flight data. The dashed black line represents the reference geometric path, the dashed blue line shows the tracked trajectory generated by TOPPQuad, a…
Figure 4
Figure 4. Figure 4: Experiment visualization of trajectories predicted from the LSTM-0.01 model, plotted from the collected motion capture flight data. The dashed black line represents the reference geometric path, the dashed blue line shows the tracked trajectory generated by TOPPQuad, a…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 15 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Mellinger and V

    D. Mellinger and V. Kumar. Minimum snap trajectory generation and control for quadrotors. In 2011 IEEE International Conference on Robotics and Automation, pages 2520--2525, 2011. doi:10.1109/ICRA.2011.5980409

    arXiv 2011

  3. [3]

    K. Mao, I. Spasojevic, M. A. Hsieh, and V. Kumar. Toppquad: Dynamically-feasible time-optimal path parametrization for quadrotors. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13136--13143, 2024. doi:10.1109/IROS58592.2024.10801611

    work page arXiv 2024

  4. [4]

    Z. Wang, X. Zhou, C. Xu, and F. Gao. Geometrically constrained trajectory optimization for multicopters. IEEE Transactions on Robotics, 38 0 (5): 0 3259--3278, 2022. doi:10.1109/TRO.2022.3160022

    arXiv 2022

  5. [5]

    W. A. De Vries, M. Li, Q. Song, and Z. Sun. An alternating peak-optimization method for optimal trajectory generation of quadrotor drones. In 2024 European Control Conference (ECC), pages 3267--3272, 2024. doi:10.23919/ECC64448.2024.10591215

    arXiv 2024

  6. [6]

    S. Liu, M. Watterson, S. Tang, and V. Kumar. High speed navigation for quadrotors with limited onboard sensing. In 2016 IEEE International Conference on Robotics and Automation (ICRA), pages 1484--1491, 2016. doi:10.1109/ICRA.2016.7487284

    arXiv 2016

  7. [7]

    S. Liu, K. Mohta, N. Atanasov, and V. Kumar. Search-based motion planning for aggressive flight in se(3). IEEE Robotics and Automation Letters, 3 0 (3): 0 2439--2446, 2018. doi:10.1109/LRA.2018.2795654

    arXiv 2018

  8. [8]

    Foehn, A

    P. Foehn, A. Romero, and D. Scaramuzza. Time-optimal planning for quadrotor waypoint flight. Science Robotics, 6 0 (56): 0 eabh1221, 2021. doi:10.1126/scirobotics.abh1221. URL https://www.science.org/doi/abs/10.1126/scirobotics.abh1221

    doi:10.1126/scirobotics.abh1221 2021

Show all 28 references
  1. [9]

    Romero, S

    A. Romero, S. Sun, P. Foehn, and D. Scaramuzza. Model predictive contouring control for time-optimal quadrotor flight. IEEE Transactions on Robotics, 38 0 (6): 0 3340--3356, 2022. doi:10.1109/TRO.2022.3173711

    2022

  2. [10]

    Teissing, M

    K. Teissing, M. Novosad, R. Penicka, and M. Saska. Real-time planning of minimum-time trajectories for agile uav flight. IEEE Robotics and Automation Letters, 9 0 (11): 0 10351--10358, 2024. doi:10.1109/LRA.2024.3471388

    2024

  3. [11]

    Meyer, K

    F. Meyer, K. Glock, and D. Sayah. Top-uav: Open-source time-optimal trajectory planner for point-masses under acceleration and velocity constraints. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2838--2845, 2023. doi:10.1109/IROS5555...

    2023

  4. [12]

    G. Ryou, E. Tal, and S. Karaman. Multi-fidelity black-box optimization for time-optimal quadrotor maneuvers. The International Journal of Robotics Research, 40 0 (12-14): 0 1352--1369, 2021

    2021

  5. [13]

    G. Ryou, E. Tal, and S. Karaman. Real-time generation of time-optimal quadrotor trajectories with semi-supervised seq2seq learning. In Conference on Robot Learning, pages 1860--1870. PMLR, 2023

    2023

  6. [14]

    Zhang, C

    R. Zhang, C. Yu, J. Chen, C. Fan, and S. Gao. Learning-based motion planning in dynamic environments using gnns and temporal encoding. Advances in Neural Information Processing Systems, 35: 0 30003--30015, 2022

    2022

  7. [15]

    Tankasala and M

    S. Tankasala and M. Pryor. Accelerating trajectory generation for quadrotors using transformers. In N. Matni, M. Morari, and G. J. Pappas, editors, Proceedings of The 5th Annual Learning for Dynamics and Control Conference, volume 211 of Proceedings of Machine Learning Researc...

    2023

  8. [16]

    Y. Wu, X. Sun, I. Spasojevic, and V. Kumar. Deep learning for optimization of trajectories for quadrotors. IEEE Robotics and Automation Letters, 9 0 (3): 0 2479--2486, 2024

    2024

  9. [17]

    Y. Song, M. Steinweg, E. Kaufmann, and D. Scaramuzza. Autonomous drone racing with deep reinforcement learning. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1205--1212, 2021. doi:10.1109/IROS51168.2021.9636053

    2021

  10. [18]

    Kaufmann, L

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. M \"u ller, V. Koltun, and D. Scaramuzza. Champion-level drone racing using deep reinforcement learning. Nature, 620 0 (7976): 0 982--987, 2023

    2023

  11. [19]

    T. Lee, M. Leok, and N. H. McClamroch. Geometric tracking control of a quadrotor uav on se(3). In 49th IEEE Conference on Decision and Control (CDC), pages 5420--5425, 2010. doi:10.1109/CDC.2010.5717652

    2010

  12. [20]

    Watterson and V

    M. Watterson and V. Kumar. Control of quadrotors using the hopf fibration on so (3). In Robotics Research: The 18th International Symposium ISRR, pages 199--215. Springer, 2019

    2019

  13. [21]

    Bansal, M

    S. Bansal, M. Chen, S. Herbert, and C. J. Tomlin. Hamilton-jacobi reachability: A brief overview and recent advances. In 2017 IEEE 56th Annual Conference on Decision and Control (CDC), pages 2242--2253. IEEE, 2017

    2017

  14. [22]

    S. Folk, J. Paulos, and V. Kumar. RotorPy : A python-based multirotor simulator with aerodynamics for education and research. arXiv preprint arXiv:2306.04485, 2023

    2023 arXiv

  15. [23]

    Giernacki, M

    W. Giernacki, M. Skwierczyński, W. Witwicki, P. Wroński, and P. Kozierski. Crazyflie 2.0 quadrotor as a platform for research and education in robotics and control engineering. In 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), pages...

    2017

  16. [24]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

    1997

  17. [25]

    Sutskever, O

    I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27, 2014

    2014

  18. [26]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need, 2023. URL https://arxiv.org/abs/1706.03762

    2023 arXiv

  19. [27]

    Bengio, O

    S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015

    2015

  20. [28]

    B. Graf. Quaternions and dynamics, 2008. URL https://arxiv.org/abs/0811.2889

    2008 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.