acceptodds
Withdrawn from ICLR 2027

Minutes of Data, Milliseconds of Control: Shared Physical Structure for Quadrotor World Modeling and Convex MPC

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quadrotor world models must learn from limited noisy observations while supporting optimization at high control rates. We introduce ABS-PWM, a probabilistic world model based on an analytically constructed bilinear lifting of quadrotor dynamics. Its analytical lifting and reconstruction maps maintain physically meaningful predictions, while a structured stochastic transition captures aerodynamic effects, unmodeled dynamics, and observation uncertainty. We train ABS-PWM directly from noisy trajectories using generalized expectation conditional maximization, combining latent-state smoothing with closed-form covariance updates. The same bilinear structure enables BUC-MPC, a convex uncertainty-aware model predictive control (MPC) scheme. Experiments in high-fidelity simulation show that ABS-PWM achieves competitive short-horizon prediction accuracy from limited flight data while providing faster rollouts than the evaluated state-of-the-art baselines. In closed loop, BUC-MPC achieves accurate tracking, high completion rates under shifts in vehicle and environmental dynamics, and computation compatible with high-frequency control. Additional experiments with a separately trained model on a second quadrotor platform demonstrate the framework's applicability across vehicle configurations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Probabilistic world models, bilinear SSM, data-efficient learning, generalized expectation-maximization, model predictive control, quadrotors, aerial robotics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_x2q5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies quadrotor trajectory tracking and proposes a probabilistic system identification framework. The method consists of two stages. First, the authors analytically lift standard nonlinear quadrotor dynamics into a pre-defined 63-dimensional Koopman representation, obtaining approximate bilinear dynamics by truncating higher-order terms. Instead of learning the representation or unrestricted dynamics operators, they estimate physical parameters (mass, inertia, drag, and damping) and Gaussian noise covariances from noisy flight trajectories using generalized expectation conditional maximization. Second, for trajectory tracking, the authors approximate the bilinear dynamics by fixing the state-dependent control matrices along a guide trajectory, which yields a convex quadratic program for MPC. A covariance-weighted disturbance observer further compensates for model mismatch.

Experiments in MuJoCo evaluate the proposed method on the Skydio X2 quadrotor using 90 seconds of noisy training data. The method achieves competitive open-loop prediction and closed-loop tracking performance compared with the evaluated baselines, while maintaining low control latency. The appendix provides qualitative tracking results on the Crazyflie platform.

Critical strengths
  • The overall method is technically sound. Its data efficiency is largely enabled by incorporating strong domain-specific knowledge into the model: the method identifies four groups of physical parameters (mass, inertia, drag, and damping), along with relatively low-dimensional Gaussian noise covariances, within a fixed, analytically defined Koopman representation. This physics-informed approach is well motivated, and the experimental results demonstrate its effectiveness with limited training data.

  • Although qualitative, the additional experiments on Crazyflie 2 demonstrate the applicability of the proposed framework to different quadrotor dynamics.

Critical weaknesses
  • The presentation is hard to follow and makes the core novelty difficult to identify. Several components have been explored in prior work, including EM-based bilinear Koopman identification [1], data-driven Koopman modeling with convex quadrotor MPC [2], and analytical Koopman lifting with truncation bounds and linear MPC [3]. These works should be discussed more thoroughly in related work and preliminaries (or appendix), so the method section can better emphasize its novel contributions: whether physics-constrained probabilistic identification improves control under model mismatch.

    • Appendix G.5 shows that the method remains relatively insensitive to modest increases in zero-mean observation noise. While encouraging, it is unclear whether this robustness comes from covariance learning or the relatively simple isotropic noise structure. Evaluating anisotropic or correlated observation noise would provide stronger evidence for the benefits of probabilistic modeling.

    • An ablation disabling the covariance-weighted wrench observer would help isolate its contribution to closed-loop performance.

    • The authors compare their convex MPC against nonlinear rigid-body MPC in Appendix G.6, but do not discuss prior work on accelerating nonlinear MPC through imitation learning [4].

  • The data-efficiency claim could be strengthened. Since the method restricts dynamics learning to four groups of physical parameters, its advantage should be evident with limited training data. A data-scaling experiment comparing prediction performance against learned Koopman representations would help demonstrate whether the proposed method saturates with less data more directly.

[1] Otto, Samuel, Sebastian Peitz, and Clarence Rowley. "Learning bilinear models of actuated Koopman generators from partially observed trajectories." SIAM Journal on Applied Dynamical Systems 23.1 (2024): 885-923.

[2] Narayanan, Sriram SKS, et al. "Se (3) koopman-mpc: Data-driven learning and control of quadrotor uavs." IFAC-PapersOnLine 56.3 (2023): 607-612.

[3] Rajkumar, Santosh Mohan, et al. "Real-time linear MPC for quadrotors on SE (3): An analytical koopman-based realization." IEEE Robotics and Automation Letters (2025).

[4] Mao, Katherine, et al. "Sequence Modeling for Time-Optimal Quadrotor Trajectory Optimization with Sampling-based Robustness Analysis." CoRL (2025).

Questions for authors
  • Is there a specific reason for considering only nine reference trajectory families? Since the method performs system identification, it should be able to handle a broader range of trajectories and control behaviors.

  • Could the authors discuss the suggested ablations in the weaknesses section and explain why they may be unnecessary or infeasible?

Reviewer_9BwnOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ABS-PWM, a probabilistic world model for quadrotors based on an analytical bilinear representation of the rigid-body dynamics. The main idea is to use physical knowledge to construct a model that can be learned from a small amount of noisy data while remaining efficient enough for real-time control.

The authors introduce a learning procedure based on generalized expectation conditional maximization to estimate physical parameters and uncertainty from flight observations. They also propose BUC-MPC, which uses the same model structure to formulate a convex tracking problem with disturbance compensation.

The method is evaluated in MuJoCo using a Skydio X2 model, with 90 seconds of training data. The results show competitive prediction accuracy and relatively low computational cost compared with the evaluated baselines. In closed-loop control, the method achieves 100% success in the in-distribution setting and 86% under out-of-distribution conditions. The authors also include a smaller experiment with a Crazyflie model to demonstrate applicability to another quadrotor platform.

Critical strengths

1. The connection between the learned dynamics model and MPC is well designed.

I think one of the strongest aspects of this paper is that the authors use the same physical structure for both model learning and control. The analytical lifting allows the system to maintain physically meaningful states, while the proposed identification method handles noisy observations and estimates uncertainty. The resulting MPC can also be formulated as a convex optimization problem by fixing the guide trajectory.

I especially appreciate the computational efficiency. The paper reports approximately 1.27 ms for a one-second model rollout and 5.24 ms per control step, which is promising for high-frequency quadrotor control. I think this is a useful contribution from an engineering perspective.

2. The method achieves reasonable results with limited training data.

Another positive aspect is that the model is trained using only 90 seconds of noisy flight data. The authors evaluate several disturbances, including wind, drag, changes in vehicle parameters, and other unmodeled dynamics. I also appreciate the ablation studies comparing the analytical structure with learned alternatives. These results provide some evidence that incorporating physical knowledge can improve learning efficiency when the available data is limited.

Critical weaknesses

1. The experimental comparisons do not clearly show where the improvements come from.

My main concern is that the paper does not sufficiently separate the benefits of the proposed bilinear representation, the physics-based model, and the disturbance observer. Although the authors present the method as a probabilistic world model, much of the learning involves identifying physical parameters and uncertainty within a predefined dynamics structure. I therefore see the approach as relatively close to classical model-based system identification.

I think a more relevant baseline would be to use the same identified rigid-body dynamics and drag model within an efficient nonlinear MPC solver, with the same disturbance observer. Appendix G.6 includes a comparison in this direction, but the nonlinear MPC uses IPOPT and takes substantially longer to solve. I am not convinced that this provides a fair assessment of the computational advantage, since efficient real-time NMPC implementations are available.

I also could not find a direct ablation that removes the disturbance observer or replaces the learned covariance weighting with a simpler alternative. These comparisons seem particularly important because uncertainty-aware control is presented as one of the main contributions. Without them, it is difficult to understand whether the improvements are mainly due to the proposed model structure or the additional compensation mechanism.

More generally, some of the learning-based baselines may be disadvantaged by the limited amount of training data, and it is not clear whether the control parameters were sufficiently tuned for the evaluated platform. For example, the reported 86% versus 84% out-of-distribution success rates for BUC-MPC and FNN-MPC are very close. I would be cautious about drawing strong conclusions from this difference without additional statistical evidence.

This matters to my assessment because the central claim is that the shared physical structure benefits both model learning and control. I do not think the current experiments establish this clearly enough.

2. The evaluation is limited to simulation, and the theoretical guarantees do not fully cover the tested conditions.

My second concern is the gap between the theoretical formulation and the experimental evaluation. The paper emphasizes data efficiency and real-time quadrotor control, but all experiments are conducted in simulation. While the MuJoCo setup includes several disturbances, it is still based on rigid-body dynamics that are relatively close to the physical model used by the proposed method.

I am also concerned about the assumptions behind Proposition 1. The convergence result requires a normalized angular-velocity bound below one, while the experimental settings allow much larger body rates. Since the experiments use a low-order truncation, I am not sure how informative this theoretical result is for the operating conditions being evaluated.

In addition, the out-of-distribution tracking results appear less convincing than the success rates alone suggest. The reported tracking errors remain relatively large in some cases, and some of the tested parameter ranges, such as negative drag coefficients, are difficult to interpret physically.

Finally, the additional Crazyflie experiment is largely qualitative. I do not think it provides enough evidence to support a broad claim about generalization across quadrotor platforms.

Overall, I think the paper presents an interesting and efficient control formulation, but I am not fully convinced that the experiments demonstrate a sufficiently strong advantage over existing methods, especially for practical deployment. These concerns make me lean toward rejection.

Questions for authors
  1. Could the authors clarify how the baseline models and controllers were trained and tuned? Did all methods use exactly the same 90 seconds of training data? Were the MPC horizons, cost weights, and other control parameters tuned specifically for the Skydio X2 platform?

  2. Why was IPOPT used for the nonlinear MPC baseline in Appendix G.6? Would the computational advantage remain as large compared with an efficient real-time NMPC implementation, such as acados with real-time iteration?

  3. How much of the closed-loop improvement comes from the proposed bilinear model itself, and how much comes from the disturbance observer? Could the authors provide results without the observer or without covariance-based weighting?

  4. Proposition 1 gives a convergence condition that appears more restrictive than the body-rate range used in the experiments. How should the theoretical result be interpreted under the actual operating conditions, especially when using a low-order truncation?

  5. The paper reports strong out-of-distribution success rates, but some tracking errors remain relatively large. Could the authors provide more details about the failure cases and clarify how successful completion is defined?

Reviewer_VN2vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a probabilistic world model (ABS-PWM) based on bilinear lifting of quadrotor dynamics together with a probabilistic identification process and uncertainty-aware MPC. The approach is evaluated in simulation and reported prediction and tracking performance are competitive whilst offering favourable efficiency. The problem is well motivated.

Critical strengths
  • This is an interesting and relevant problem, where the authors address a challenging combination of physics informed dynamics learning, probabilistic identification and efficient predictive control for quadrotors.

  • There is a potentially useful contribution in the analytical bilinear representation.

  • Empirical performance in benchmarking does appear to show competitive predictive and tracking performance with favorable efficiency under the evaluated conditions.

  • The authors provide detailed derivations, algorithms, ablations and robustness experiments, in addition to code in support of reproducibility. The limitations appendix is useful and candid.

Critical weaknesses
  • The authors do not clearly differentiate the claimed principal novelty. It is difficult to identify the genuinely new methodological contributions from established identification and control techniques, or to determine whether the main advance is the analytical representation or the integration of the overall framework.

  • The benchmarking establishes the relative performance of the integrated approach, but the ablations do not appear sufficient to consistently isolate and substantiate the contributions of the sub-components (novel or otherwise). The limited data claim is also difficult to assess given that the rationale for the 90 s dataset appears not to be provided nor analysis of sensitivity to training data quantity.

  • The training algorithm, experimental details, important ablations and limitations appear in appendices. As such, the main paper does not provide a sufficiently clear and self-contained argument for the novelty and supporting evidence. This makes the paper unnecessarily difficult to assess.

  • The scope of the experimental claims needs better qualification insofar as the evaluation is entirely simulation-based, noisy full-state measurements are assumed, and the computational hardware used is fairly substantial. The reported computational efficiency is encouraging, but the 100 Hz result does not establish feasibility for onboard operation on either platform. Evaluation on a second simulated platform is useful, but seeing as the models were trained independently, cross-platform transfer without retraining is not demonstrated.

Questions for authors
  1. Could the authors clarify precisely what is novel about the lifting construction in equations (4)–(6), particularly relative to established approaches?

  2. What is the rationale behind the 90 seconds of training data? How do prediction and tracking performance change as the amount of data is systematically reduced?

  3. The ablations in Appendix G, particularly G2, do not appear to sufficiently isolate the respective components' effects. To what extent can the improvements be related to the analytical representation, identification and uncertainty-aware controller in isolation?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.