Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Learning Agile Quadrotor Flight in the Real World

T0 review · 3 major / 5 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 self-adaptive quadrotor controller evolves a conservative 1.9 m/s policy to 7.3 m/s within about 100 seconds of real-world flight, without precise system identification or offline sim-to-real transfer.

desk verdict Impressive real-world adaptive flight with a genuinely new ATS mechanism, but the safety argument leans on an unvalidated inner-loop assumption and the evidence lacks repeated trials. read the letter →

arxiv 2602.10111 v2 pith:OTMBMP44 submitted 2026-02-10 cs.RO

Yunfan Ren , Zhiyuan Zhu , Jiaxu Xing , Davide Scaramuzza This is my paper
classification cs.RO
keywords agilequadrotorcontrolonlineadaptationresidualdynamicslearningdifferentiablesimulationbackpropagationthroughtimeadaptivetemporalscalingreal-worldreinforcementactuatorsaturation
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 argues that a quadrotor can safely learn agile flight directly in the real world, eliminating the need for precise system identification or offline sim-to-real transfer. It introduces a self-adaptive loop in which a neural residual network continually augments a simple nominal dynamics model, policy parameters are updated via short-horizon backpropagation through a differentiable version of that model anchored to the current real state, and an adaptive temporal-scaling mechanism adjusts how fast the reference trajectory is executed based on predicted tracking error. The result, demonstrated on a physical quadrotor, is that a conservative 1.9–2.2 m/s base policy evolves to 7.2–7.3 m/s within about 100 seconds of flight, with commanded body rates reaching the 6 rad/s actuator limit. This matters because it suggests that continual in-field learning can replace costly identification and simulate-then-deploy pipelines and can rapidly absorb hardware changes and external disturbances.

What carries the argument

The load-bearing mechanism is the combination of Adaptive Temporal Scaling (ATS) and Real-world Anchored Short-horizon Backpropagation Through Time (RASH-BPTT), both operating on an online-learned hybrid dynamics model. The hybrid model augments the nominal rigid-body dynamics with a residual neural network trained to minimize one-step prediction error, capturing unmodeled accelerations and body-rate discrepancies. RASH-BPTT optimizes the policy by differentiating through a short rollout of this differentiable model, initialized from the most recent real-world state to avoid compounding errors. ATS treats the reference trajectory's time dilation factor α as an optimizer variable, using the s

What would settle it

Conduct the same figure-eight learning with an artificially degraded inner-loop attitude-rate tracking (e.g., adding a 50 ms delay or lowering the attitude controller bandwidth). If the learned residual cannot capture the resulting closed-loop lag, the predicted tracking error will exceed the safety threshold before the system reaches high speeds, contradicting the claim that the framework safely explores near actuator saturation.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that real-world adaptation can serve as a practical mechanism for sustained performance improvement in aggressive flight regimes. The central evidence: starting from a base policy trained only on the nominal model, the system reaches near-actuator-saturation on a line and figure-eight tasks within roughly 100 seconds; after hardware alterations (added mass, clipped propellers, or both), the residual model compensates within one update and ATS pushes the compromised platform to its new limit; and in an inspection mission with fan-generated wind, total motion time drops 42% while tracking error stays bounded. The paper attributes this success to the syne

Load-bearing premise

The safety and performance claims rely on the assumption stated in Eq. (1c): that the low-level attitude-rate loop tracks commanded body rates fast enough for all angular uncertainty to be captured by a residual on the commanded rates; near the 6 rad/s saturation boundary this assumption is likely under stress.

Editorial extensions

If this is right

  • Quadrotors can operate near their physical limits in real time without offline sim-to-real transfer or system identification, starting from a conservative policy and adapting entirely in flight.
  • A simple nominal model augmented with online-learned residuals is sufficient for aggressive maneuvers, capturing unmodeled aerodynamics, delays, and hardware changes without expensive identification.
  • Adaptive temporal scaling provides a principled, automatic balance between speed and safety: the system accelerates when tracking is accurate and slows down when disturbances or model mismatch increase error.
  • Hardware degradation such as added mass or clipped propellers can be handled by rapid re-identification of the new dynamics, enabling operation near the degraded physical limit rather than conservatively.
  • The 42% mission-time reduction under unknown wind shows that online adaptation can yield practical task-level improvements in dynamic environments, not just lab demonstrations.

Reading between the lines

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

  • The 'temporal elasticity' idea—adjusting the execution rate of a fixed path based on a safety barrier—likely generalizes beyond quadrotors to any system with a differentiable dynamics surrogate and an online residual model, such as legged robots or manipulators.
  • The paper leaves joint spatial-temporal optimization as future work; if the same differentiable pipeline also optimizes path shape, it could enable simultaneous obstacle avoidance and time-optimal flight.
  • The safety claim depends on the fidelity of the hybrid model near the boundary; at actuator saturation the residual network may not capture inner-loop lag, so a stress test with artificially degraded attitude-rate tracking would reveal whether the ATS barrier still holds.
  • The framework currently assumes clean, low-latency state estimation; extending ATS to include perception-aware constraints (e.g., motion blur limits) could make it applicable to vision-only agile flight, which the paper lists as future work.
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

3 major / 5 minor

Summary. The paper proposes a self-adaptive framework for real-world quadrotor agile flight that removes the need for offline system identification or Sim2Real transfer. It combines three components: online residual learning that augments a nominal rigid-body model with a neural residual on acceleration and body-rate commands; RASH-BPTT, a short-horizon differentiable-model policy optimization anchored to the current real state; and Adaptive Temporal Scaling (ATS), which jointly optimizes the temporal scaling of a fixed reference path against a tracking-error safety barrier. Real-world experiments claim the system evolves from a conservative 2.0 m/s base policy to 7.3 m/s on a line shuttle and 7.2 m/s on a figure-eight within about 100 seconds of flight, reaches actuator saturation at 6 rad/s, adapts to propeller damage and added payload, and reduces an inspection time by 42% under wind. Simulation ablations isolate the contributions of residual learning and anchored rollouts.

Significance. If the central claims hold, this is a meaningful step toward safe, sample-efficient, continual learning of agile control directly on physical robots. The paper's strengths are its real-world experimentation across multiple tasks and hardware degradations, the clear modular design, and the ablation evidence that rotational residuals and anchored rollouts are both important. The approach is also technically coherent and, in principle, reproducible with the described pipeline. However, the headline claims of 'safety' and 'reliably reaching actuator limits' rest on a dynamics model whose key rotational assumption is asserted but not validated in the high-speed, near-saturation regime where it matters most. The experimental evidence also appears to consist of single-run demonstrations without statistical repetition, which weakens the reliability and safety statements. The paper's own Limitations section acknowledges that boundary regions may be under-modeled, which is precisely the regime at stake.

major comments (3)
  1. [§III-A1, Eq. (1c)] The rotational model assumes the low-level attitude-rate loop tracks commanded body rates fast enough that all mismatch can be captured by a static residual omega_res(zeta; theta). Near the claimed 6 rad/s actuator-saturation boundary this assumption is likely violated: the inner loop has finite bandwidth, and actual body rate will lag the command with state- and history-dependent dynamics. Since this same hybrid model is used both for RASH-BPTT policy gradients (§III-B, Eq. (5)) and for the ATS safety barrier (§III-C, Eq. (8)), any optimism in angular responsiveness directly inflates both the predicted performance gain and the claimed safety margin. The paper acknowledges the assumption in words but gives no bandwidth measurement, no predicted-vs-actual body-rate comparison during high-speed segments, and no held-out multi-step prediction error near saturation. Please add such validatio
  2. [§IV-A, Figs. 5–7] The central claims 'reliably executes agile maneuvers near actuator saturation limits' and 'safe exploration' are not supported statistically. Each hardware condition appears to be demonstrated with a single continuous run; there are no repeated flights, no distributions over runs, no failure counts, and no confidence intervals on the reported speeds, saturations, or tracking errors. For an experimental paper whose headline is safe real-world adaptation, at least a small number of repeated trials for the key conditions (line shuttle, figure-eight, and inspection) are needed, or the claims should be qualified as single-run demonstrations. This is especially load-bearing for the safety language in the abstract and Section V.
  3. [§IV-B, Fig. 8] The ablation study that supports the paper's main mechanistic conclusions ('residual learning is the primary enabler for agility', 'angular dynamics fidelity is critical') is performed entirely in simulation, with all baselines augmented by ATS. While this is a reasonable way to compare components, the paper's contribution is specifically real-world learning, and the simulation results do not automatically transfer to the real platform where the model mismatch is different. Please provide at least one real-world ablation (e.g., with and without the rotational residual, or with and without anchored rollouts) on the line-shuttle or figure-eight task, or clearly state that the component analysis is simulation-only and not validated in the real-world setting.
minor comments (5)
  1. [Abstract vs. §I] The initial conservative peak speed is inconsistent: the abstract says '1.9 m/s', while §I and Fig. 1 say '2.0 m/s'. Please align these numbers.
  2. [Fig. 1 caption] The caption says 'tripling its speed' for a change from 2.0 m/s to 7.3 m/s; the factor is about 3.65 and should be phrased as 'more than tripling' or 'quadrupling'.
  3. [§III-C1, Eq. (7)] The reference trajectory parameterization is defined only through the position polynomial; the use of differential flatness to obtain the full reference state is mentioned but not specified. Please clarify how yaw and higher-order derivatives are set, since the tracking error D(·,·) includes attitude.
  4. [Notation] The symbol D is used both as the discrepancy metric in Eq. (4) and in Eq. (9). This reuse is not harmful but could confuse readers; a distinct symbol (e.g., d) for the ATS error would improve clarity.
  5. [Fig. 3] The schematic in Fig. 3 uses 'Time Scale' and 'Tracking Error (RMSE)' as axes, but the heatmap labels 'Safe Zone' and 'Unsafe Zone' are not quantitatively defined. Clarify that this is illustrative, not a measured landscape.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the speed and time-reduction results are measured outcomes, and the learned-model loop is standard model-based RL rather than a definitional reduction.

full rationale

The paper's central empirical results (evolution from 2.0 m/s to 7.3 m/s, 42% mission-time reduction) are measured real-world outcomes, not quantities obtained by fitting constants and then renaming them as predictions. The hybrid residual model is trained on one-step prediction error (Eq. 4) and then used inside RASH-BPTT and ATS, which is a standard model-based RL loop; the terminal speed is not algebraically forced by the training loss or by the ATS objective alone, since it depends on real closed-loop tracking performance. Citations to prior group work (LOFT [16], Agilicious [36], differential flatness [34]) are contextual or used for implemented comparisons and standard components, not as unexamined load-bearing premises; the LOFT ablation is implemented in this paper rather than quoted as an external result. The rotational-model assumption in Eq. (1c) and the admitted limitation that passive residual learning may leave boundary regions under-modeled affect the reliability of the safety and near-saturation claims, but these are modeling-validity concerns, not circularity. No equation is defined in terms of the result it purports to establish, and no fitted parameter is presented as an independent prediction.

Assumptions & free parameters 7 free parameters · 7 assumptions · 0 invented entities

No new physical entities are introduced; the residual network and ATS are algorithmic components, not postulated ontology. The central claim rests on several domain assumptions about the inner loop, state estimation, and model fidelity. The fitted parameters are the residual network weights and hand-tuned thresholds/weights; the numerical speed and mission-time results are empirical outputs, not fitted constants.

free parameters (7)
  • Residual network weights theta = trained online on real transitions via Eq. (4)
    Central to the hybrid dynamics model; learned from data, not first principles.
  • Safety threshold E_th = 0.3 m (real flight), 0.35 m (simulation)
    Hand-set tracking-error threshold for the ATS barrier; directly controls how aggressive the trajectory scaling becomes.
  • ATS weights lambda_speed, lambda_safe, and softplus sharpness kappa = not reported
    Balance speed versus safety in Eq. (8); replication requires guessing these values.
  • Short horizon H and discount gamma = not reported
    RASH-BPTT rollout length and reward discount; critical for optimization stability and not specified.
  • Learning rates eta_pi and eta_alpha = not reported
    Policy and ATS update rates; omitted from the paper.
  • Regularization weight lambda_reg = not reported
    Spectral-norm penalty in the residual loss (Eq. 4).
  • Time-scale bounds alpha_min, alpha_max = not reported
    Projection bounds in Eq. (15) that constrain the speed envelope.
assumptions (7)
  • domain assumption Low-level attitude-rate loop tracks commanded body rates fast enough that uncertainty can be modeled as a residual on commanded rates.
    Invoked in Section III-A1, Eq. (1c); likely violated near the 6 rad/s saturation boundary where inner-loop lag is significant.
  • domain assumption The nominal rigid-body model plus neural residual accurately predicts real dynamics over the short-horizon rollout length H.
    Required for RASH-BPTT gradients and ATS safety barrier to be meaningful; one-step training does not guarantee multi-step accuracy.
  • domain assumption The differentiable hybrid model is a valid proxy for counterfactual inference in ATS.
    Eqs. (10)-(14) linearize the hybrid model along the real rollout; if this local sensitivity is wrong, alpha updates can move in the wrong direction.
  • domain assumption Model-predicted tracking error below E_th implies real-world safety.
    The ATS barrier (Eq. 8) treats predicted error as a safety certificate; no formal guarantee or robustness analysis is provided.
  • domain assumption Motion capture provides accurate, low-latency state estimates at 100 Hz.
    State estimation is used for residual learning and anchoring; degraded sensing is acknowledged as a limitation in Section V-C.
  • standard math Quadrotor dynamics are differentially flat for reference generation.
    Used to map position derivatives to the full reference state in Eq. (7), standard for quadrotors with rotor drag [34].
  • domain assumption The pre-trained nominal BPTT policy is safe to deploy in the real world.
    Initialization phase trains on the nominal model; no sim-to-real gap analysis is given for this initial transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Agile Quadrotor Flight in the Real World." pith.science (2026). https://pith.science/paper/OTMBMP44

@misc{pith2026260210111,
  author       = {Pith},
  title        = {Pith review of: Learning Agile Quadrotor Flight in the Real World},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTMBMP44}},
  note         = {Machine review of arXiv:2602.10111}
}
read the original abstract

Learning-based controllers have achieved impressive performance in agile quadrotor flight but typically rely on massive training in simulation, necessitating accurate system identification for effective Sim2Real transfer. However, even with precise modeling, fixed policies remain susceptible to out-of-distribution scenarios, ranging from external aerodynamic disturbances to internal hardware degradation. To ensure safety under these evolving uncertainties, such controllers are forced to operate with conservative safety margins, inherently constraining their agility outside of controlled settings. While online adaptation offers a potential remedy, safely exploring physical limits remains a critical bottleneck due to data scarcity and safety risks. To bridge this gap, we propose a self-adaptive framework that eliminates the need for precise system identification or offline Sim2Real transfer. We introduce Adaptive Temporal Scaling (ATS) to actively explore platform physical limits, and employ online residual learning to augment a simple nominal model. {Based on the learned hybrid model, we further propose Real-world Anchored Short-horizon Backpropagation Through Time (RASH-BPTT) to achieve efficient and robust in-flight policy updates. Extensive experiments demonstrate that our quadrotor reliably executes agile maneuvers near actuator saturation limits. The system evolves a conservative base policy with a peak speed of 1.9 m/s to 7.3 m/s within approximately 100 seconds of flight time. These findings underscore that real-world adaptation serves not merely to compensate for modeling errors, but as a practical mechanism for sustained performance improvement in aggressive flight regimes.

Figures

Figures reproduced from arXiv: 2602.10111 by the authors.

Figure 1
Figure 1. Autonomous online evolution of agile flight. The figure illustrates the continuous loop between Real-World Rollouts and Rapid Policy Learning, comparing the initial conservative flights (Iter 0, top) with the evolved agile behaviors (Iter 12, bottom). Trajectories are composited with frames sampled at uniform 0.5 s intervals. Consequently, trajectory density reflects flight speed, where sparser segments indicate hig… view at source ↗
Figure 2
Figure 2. Overview of the self-adaptive autonomous flight framework. The system operates as a continuous closed-loop cycle (bottom right) bridging physical execution and differentiable simulation: (A) Policy Learning: Leveraging a learned hybrid dynamics model in a differentiable simulator, we perform RASH-BPTT to optimize the control policy via massively parallelized rollouts. (B) Real-World Rollout: The agent executes the c… view at source ↗
Figure 3
Figure 3. The optimization landscape for ATS. The heatmap visualizes the composite potential JATS, balancing agility (low α) against safety. The landscape transitions from the Safe Zone (blue) to the Unsafe Zone (red) determined by the tracking error threshold Eth (dotted line) and a schematic representation of the system’s physical limits (dashed curve). The green dot marks the optimal equilibrium: the most aggressive time s… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Experimental platform with extreme modifications. To validate robustness, the nominal quadrotor (192 g) is subjected to drastic degradations: mechanically clipped propellers (inducing aerodynamic loss) and a 60 g payload. This 31% mass increase significantly alters the…
Figure 6
Figure 6. Figure 6: Adaptation to hardware variations. The framework is tested with i) [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: Ablation study of residual dynamics and anchored rollouts. Note: Since no existing methods support online learning at the physical limit in real-world settings, all baselines are augmented with our ATS framework to enable feasible deployment. We compare the Baseline (n…
Figure 9
Figure 9. Figure 9: Online adaptation under wind disturbance. Comparison between the Baseline and Ours. The Baseline fails to accelerate due to unmodeled wind drag, remaining stuck at lower speeds of 4.7 m/s. In contrast, Ours demonstrates a rapid adaptation cycle: the residual network fi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 4 linked inside Pith

  1. [1]

    Deep reinforcement learning for robotics: A survey of real-world successes,

    C. Tang, B. Abbatematteo, J. Hu, R. Chandra, R. Mart ´ın- Mart´ın, and P. Stone, “Deep reinforcement learning for robotics: A survey of real-world successes,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 8, no. 1, pp. 153–188, 2025

    2025

  2. [2]

    Learning high-speed flight in the wild,

    A. Loquercio, E. Kaufmann, R. Ranftl, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Learning high-speed flight in the wild,”Science Robotics, vol. 6, no. 59, p. eabg5810, 2021

    2021

  3. [3]

    Champion-level drone racing using deep reinforcement learning,

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Champion-level drone racing using deep reinforcement learning,”Nature, vol. 620, no. 7976, pp. 982–987, 2023

    2023

  4. [4]

    Learning quadrupedal locomotion over chal- lenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over chal- lenging terrain,”Science Robotics, vol. 5, no. 47, p. eabc5986, 2020

    2020

  5. [5]

    Learning robust perceptive locomotion for quadrupedal robots in the wild,

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning robust perceptive locomotion for quadrupedal robots in the wild,”Science Robotics, vol. 7, no. 62, p. eabk2822, 2022

    2022

  6. [6]

    Agile but safe: Learning collision-free high-speed legged locomotion,

    T. He, C. Zhang, W. Xiao, G. He, C. Liu, and G. Shi, “Agile but safe: Learning collision-free high-speed legged locomotion,”arXiv preprint arXiv:2401.17583, 2024

    arXiv 2024

  7. [7]

    Diffusion policy: Visuo- motor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burch- fiel, R. Tedrake, and S. Song, “Diffusion policy: Visuo- motor policy learning via action diffusion,”The Interna- tional Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  8. [8]

    Universal manipula- tion interface: In-the-wild robot teaching without in-the- wild robots,

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipula- tion interface: In-the-wild robot teaching without in-the- wild robots,”arXiv preprint arXiv:2402.10329, 2024

    arXiv 2024

Show all 38 references
  1. [9]

    Contrastive learning for enhancing robust scene transfer in vision-based agile flight,

    J. Xing, L. Bauersfeld, Y . Song, C. Xing, and D. Scara- muzza, “Contrastive learning for enhancing robust scene transfer in vision-based agile flight,” in2024 IEEE International Conference on Robotics and Automation (ICRA), pp. 5330–5337, IEEE, 2024

    2024

  2. [10]

    Domain randomization for transferring deep neural networks from simulation to the real world,

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel, “Domain randomization for transferring deep neural networks from simulation to the real world,” in2017 IEEE/RSJ international conference on intelligent robots and systems (IROS), pp. 23–30, IEEE, 2017

    2017

  3. [11]

    Autonomous drone racing: A survey,

    D. Hanover, A. Loquercio, L. Bauersfeld, A. Romero, R. Penicka, Y . Song, G. Cioffi, E. Kaufmann, and D. Scaramuzza, “Autonomous drone racing: A survey,” IEEE Transactions on Robotics, vol. 40, pp. 3044–3067, 2024

    2024

  4. [12]

    Neural-fly enables rapid learning for agile flight in strong winds,

    M. O’Connell, G. Shi, X. Shi, K. Azizzadenesheli, A. Anandkumar, Y . Yue, and S.-J. Chung, “Neural-fly enables rapid learning for agile flight in strong winds,” Science Robotics, vol. 7, no. 66, p. eabm6597, 2022

    2022

  5. [13]

    The reality gap in robotics: Challenges, solutions, and best practices,

    E. Aljalbout, J. Xing, A. Romero, I. Akinola, C. R. Garrett, E. Heiden, A. Gupta, T. Hermans, Y . Narang, D. Fox,et al., “The reality gap in robotics: Challenges, solutions, and best practices,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 9, 2025

    2025

  6. [14]

    Neurobem: Hybrid aerodynamic quadrotor model,

    L. Bauersfeld, E. Kaufmann, P. Foehn, S. Sun, and D. Scaramuzza, “Neurobem: Hybrid aerodynamic quadrotor model,” inProceedings of Robotics: Science and Systems, 2021

    2021

  7. [15]

    Sampling-based system identification with active explo- ration for legged sim2real learning,

    N. Sobanbabu, G. He, T. He, Y . Yang, and G. Shi, “Sampling-based system identification with active explo- ration for legged sim2real learning,” inConference on Robot Learning, pp. 578–598, PMLR, 2025

    2025

  8. [16]

    Learning on the fly: Rapid policy adaptation via differentiable simulation,

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

    2026

  9. [17]

    Con- trol of a quadrotor with reinforcement learning,

    J. Hwangbo, I. Sa, R. Siegwart, and M. Hutter, “Con- trol of a quadrotor with reinforcement learning,”IEEE Robotics and Automation Letters, vol. 2, no. 4, pp. 2096– 2103, 2017

  10. [18]

    Reaching the limit in autonomous racing: Op- timal control versus reinforcement learning,

    Y . Song, A. Romero, M. M¨uller, V . Koltun, and D. Scara- muzza, “Reaching the limit in autonomous racing: Op- timal control versus reinforcement learning,”Science Robotics, vol. 8, no. 82, p. eadg1462, 2023

    2023

  11. [19]

    Learning to fly in seconds,

    J. Eschmann, D. Albani, and G. Loianno, “Learning to fly in seconds,”IEEE Robotics and Automation Letters, vol. 9, no. 7, pp. 6336–6343, 2024

    2024

  12. [20]

    Bootstrapping reinforcement learning with imitation for vision-based agile flight,

    J. Xing, A. Romero, L. Bauersfeld, and D. Scaramuzza, “Bootstrapping reinforcement learning with imitation for vision-based agile flight,” inConference on Robot Learn- ing, pp. 2542–2556, PMLR, 2025

    2025

  13. [21]

    Learning to walk in minutes using massively parallel deep rein- forcement learning,

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter, “Learning to walk in minutes using massively parallel deep rein- forcement learning,” inConference on Robot Learning, pp. 91–100, PMLR, 2022

    2022

  14. [22]

    Learning quadro- tor control from visual features using differentiable simulation,

    J. Heeg, Y . Song, and D. Scaramuzza, “Learning quadro- tor control from visual features using differentiable simulation,” in2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 4033–4039, IEEE, 2025

    2025

  15. [23]

    Fast aquatic swimmer optimization with differentiable projective dynamics and neural network hydrodynamic models,

    E. Nava, J. Z. Zhang, M. Y . Michelis, T. Du, P. Ma, B. F. Grewe, W. Matusik, and R. K. Katzschmann, “Fast aquatic swimmer optimization with differentiable projective dynamics and neural network hydrodynamic models,” inInternational Conference on Machine Learn- ing, pp. 16413–...

    2022

  16. [24]

    Disect: A differentiable simulation en- gine for autonomous robotic cutting,

    E. Heiden, M. Macklin, Y . Narang, D. Fox, A. Garg, and F. Ramos, “Disect: A differentiable simulation en- gine for autonomous robotic cutting,”arXiv preprint arXiv:2105.12244, 2021

    2021 arXiv

  17. [25]

    Efficient tactile simulation with differentiability for robotic manipulation,

    J. Xu, S. Kim, T. Chen, A. R. Garcia, P. Agrawal, W. Ma- tusik, and S. Sueda, “Efficient tactile simulation with differentiability for robotic manipulation,” inConference on Robot Learning, pp. 1488–1498, PMLR, 2023

    2023

  18. [26]

    Accelerated policy learning with parallel differentiable simulation,

    J. Xu, V . Makoviychuk, Y . Narang, F. Ramos, W. Ma- tusik, A. Garg, and M. Macklin, “Accelerated policy learning with parallel differentiable simulation,” inInter- national Conference on Learning Representations, 2022

    2022

  19. [27]

    Adaptive horizon actor-critic for policy learning in contact-rich differentiable simulation,

    I. Georgiev, K. Srinivasan, J. Xu, E. Heiden, and A. Garg, “Adaptive horizon actor-critic for policy learning in contact-rich differentiable simulation,” inProceedings of the 41st International Conference on Machine Learning, pp. 15418–15437, 2024

    2024

  20. [28]

    Residual policy learning for perceptive quadruped control using differentiable simulation,

    J. Y . Luo, Y . Song, V . Klemm, F. Shi, D. Scaramuzza, and M. Hutter, “Residual policy learning for perceptive quadruped control using differentiable simulation,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 1–8, IEEE, 2025

    2025

  21. [29]

    Neural ordinary differential equations,

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duve- naud, “Neural ordinary differential equations,”Advances in neural information processing systems, vol. 31, 2018

    2018

  22. [30]

    Neural lander: Stable drone landing control using learned dynamics,

    G. Shi, X. Shi, M. O’Connell, R. Yu, K. Azizzadenesheli, A. Anandkumar, Y . Yue, and S.-J. Chung, “Neural lander: Stable drone landing control using learned dynamics,” in2019 IEEE International Conference on Robotics and Automation (ICRA), pp. 9784–9790, IEEE, 2019

    2019

  23. [31]

    When to trust your model: Model-based policy optimization,

    M. Janner, J. Fu, M. Zhang, and S. Levine, “When to trust your model: Model-based policy optimization,”Advances in neural information processing systems, vol. 32, 2019

    2019

  24. [32]

    JAX: com- posable transformations of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. Vander- Plas, S. Wanderman-Milne, and Q. Zhang, “JAX: com- posable transformations of Python+NumPy programs,” 2018

    2018

  25. [33]

    Efficient model-based reinforcement learning for robot control via online learning,

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

    2025 arXiv

  26. [34]

    Differen- tial flatness of quadrotor dynamics subject to rotor drag for accurate tracking of high-speed trajectories,

    M. Faessler, A. Franchi, and D. Scaramuzza, “Differen- tial flatness of quadrotor dynamics subject to rotor drag for accurate tracking of high-speed trajectories,”IEEE Robotics and Automation Letters, vol. 3, no. 2, pp. 620– 626, 2017

    2017

  27. [35]

    Woulda, coulda, shoulda: Counterfactually-guided policy search,

    L. Buesing, T. Weber, Y . Zwols, N. Heess, S. Racaniere, A. Guez, and J.-B. Lespiau, “Woulda, coulda, shoulda: Counterfactually-guided policy search,” inInternational Conference on Learning Representations, 2018

    2018

  28. [36]

    Agilicious: Open-source and open-hardware agile quadrotor for vision-based flight,

    P. Foehn, E. Kaufmann, A. Romero, R. Penicka, S. Sun, L. Bauersfeld, T. Laengle, G. Cioffi, Y . Song, A. Loquer- cio,et al., “Agilicious: Open-source and open-hardware agile quadrotor for vision-based flight,”Science Robotics, vol. 7, no. 67, p. eabl6259, 2022

    2022

  29. [37]

    Safety-assured high-speed nav- igation for mavs,

    Y . Ren, F. Zhu, G. Lu, Y . Cai, L. Yin, F. Kong, J. Lin, N. Chen, and F. Zhang, “Safety-assured high-speed nav- igation for mavs,”Science Robotics, vol. 10, no. 98, p. eado6187, 2025

    2025

  30. [38]

    Model predictive contouring control for time-optimal quadrotor flight,

    A. Romero, S. Sun, P. Foehn, and D. Scaramuzza, “Model predictive contouring control for time-optimal quadrotor flight,”IEEE Transactions on Robotics, vol. 38, no. 6, pp. 3340–3356, 2022

    2022

Pith tools

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