acceptodds
Under review as a conference paper at ICLR 2027

SUDA: A Simulation and Data Generation Platform for UAV Dynamics-Aware Policy Learning

Abstract

Language-guided UAV control is moving from route planning toward continuous action prediction from onboard observations and instructions. Existing datasets typically supervise policies with waypoint or 6-DoF reference trajectories. Yet a reference specifies intended motion, whereas closed-loop flight is governed by the response produced after commands pass through the flight controller and vehicle dynamics. The resulting state also determines the next onboard observation, making execution-aligned command–response–observation trajectories essential for policy learning. To provide this supervision at scale, we introduce SUDA, a simulation and data generation platform, benchmark, and learning framework for controller-executed UAV policies. SUDA-SIM combines OpenUSD/RTX sensing, PhysX dynamics, PX4 SITL execution, lockstep logging, and parallel rollouts under one simulation clock. Using SUDA-SIM, we construct SUDA-UAV with 13,678 controller-executed trajectories and 1.6 million synchronized observations, organized as Dynamics-aware Action Trajectories with 15D motion-response targets. We further build a 200-task PX4 benchmark spanning seven task groups and eight VLA, world-model, and video-action policy families under a common closed-loop protocol. We further use a QwenVLA-style reference policy to study text-to-action pretraining, multimodal Action Chunk fine-tuning, and PPO post-training within the same framework. Evaluations show an advantage for controller-executed supervision: all six chunked families improve both SR and nDTW, and nDTW improves across all eight families. These results provide consistent evidence that controller-executed supervision improves closed-loop UAV policy learning under a shared PX4 SITL evaluation protocol.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Unmanned Aerial Vehicles (UAVs), Vision-Language-Action (VLA), UAV Simulation, Autonomous Navigation, Vision-and-Language Navigation (VLN), Reinforcement Learning, Embodied AI, Robot Learning

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_uLTNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces SUDA, a platform that runs Isaac Sim and the PX4 autopilot on one clock step that makes it easier to synchronize aligned commands, executed motion and camera frames at 20Hz across 4 drones per scene. Because a real autopilot lags in real life and usually overshoots path at times, the construct SUDA-UAV, a controller-executed dataset for dynamics- aware UAV policy learning (13.7k flights, 1.6M frames), where each step carries a 15D motion-response vector. The third and final part of their paper is the execution adapter - where the policy predicts this 15D response but only 7 channels get sent to PX4 and the other 8 act as auxiliary training targets. On a 200-task closed loop benchmark, training on executed instead of planned targets lifts mean success from 41.4% to 51.3% across eight policy families.

Critical strengths
  1. This is a very well documented test bed for closed loop UAV testing and all policies run through through a common PX4 mapping
  2. A wide, matched comparison with a consistent result - 8 existing model families, plus the authors’ own Qwen-based model, are each trained twice with the same settings: once on the drone’s actual motion and once on the planned motion. Success goes up in all nine of them
Critical weaknesses
  1. The dynamics-awareness claim isn’t isolated - where the QwenVLA reference baseline is trained on reference-derived state (D.1) but fed executed state in flight (F.3). With no run that is separating the targets from inputs, its plausible that the gain could come from this mismatch or a better target adapter fit rather than dynamics awareness. Training and testing also use the same simulated PX4 setup, with no real flights or changed dynamics, so nothing shows the benefit carries beyond this one controller setup.
  2. The benchmark doesn’t show it measures real control - where tasks are short (median reference 1 m, B.4), so hovering alone would hit OR@1m on about half of them, yet there’s no hover baseline. An undescribed GPT-5.5 Policy (E.1) also matches the SUDA-trained family mean (51.5% vs 51.3% SR)
Questions for authors
  1. Could you add a hover baseline and the recorded PX4 flights of the 200 references under the same metrics - where it shows what staying put scores vs. the recorded flights themselves?
  2. The QwenVLA reference baseline is trained on reference-derived state (D.1) but gets executed state in flight (F.3). Can you add a run with reference targets and executed-state input?
  3. What is the GPT-5.5 Policy (E.1), and was it trained on SUDA-UAV?
  4. Does the gain still hold when the dynamics change - i.e that’s where dynamics awareness should show up ?
  5. Each checkpoint is trained once (D.3). Can you add 2 or 3 seeds, at least for the smaller gains?
  6. Tables A8 and A9 give different FDE for the same checkpoints and even flip whether FDE improves on average - which one is correct?
3 comments
AuthorsOpenReview ↗

Response to Reviewer uLTN (1/2)

Response to Reviewer uLTN

We thank you for recognizing the common PX4 testbed and broad comparisons, and for your specific suggestions on calibration, inputs, and result consistency.

Q1 Hover and recorded flights

Could you add a hover baseline and the recorded PX4 flights of the 200 references under the same metrics - where it shows what staying put scores vs. the recorded flights themselves?

A1. Thank you for suggesting this direct benchmark calibration. We compare the source paths for the 200 tasks, actual PX4 hover with the same starting conditions and time budgets, and both learned policies using common metrics.

ControlNSR↑nDTW↑Arc nDTW 0.1m↑Arc nDTW 0.2m↑ATE(m)↓FDE(m)↓
Recorded source paths155100.0%1.000——0.0000.000
Synthetic translation references45100.0%1.000——0.0000.000
Reference-supervised SFT20041.5%0.3360.1450.1561.6803.200
Actual PX4 simulated hover20041.0%0.5230.3270.3431.1081.766
SUDA-supervised SFT (Ours)20064.5%0.6280.5120.5230.7751.403

The sources contain 155 recorded executed paths and 45 synthetic translation references. The first two rows compare each source path with its own reference, calibrating the reference-path scores. The hover row uses measured PX4 states obtained while continuously sending hold targets. Its OR@1m and OR@3m are 59.0% and 82.0%.

To distinguish stationary sampling from spatial motion, arc-length nDTW resamples paths at fixed spatial intervals before applying the original DTW cost and exponential normalization. Ours improves over hover by 23.5 percentage points in SR and 0.186 in 0.1m arc-length nDTW. The policy ordering agrees at 0.1m and 0.2m spacing. SR follows the original reported evaluation protocol, while arc-length nDTW provides an additional spatial-path diagnostic.

Hover uses 50Hz physics updates and the nearest measured state at each 20Hz scoring boundary, with a maximum offset of 10ms. A separate matched 200-response replay yields position RMSE/ATE/FDE of 0.455/0.374/0.603m, quantifying re-execution through the seven-channel mapping. Supplementary Tables S2–S3 give settings and metric formulas.

Q2 Executed-state input held fixed

The QwenVLA reference baseline is trained on reference-derived state (D.1) but gets executed state in flight (F.3). Can you add a run with reference targets and executed-state input?

A2. Your question identifies the relationship between training-time motion states and online executed states. D.1 and F.3 specify the conditions below.

ConditionTraining state (D.1)Future-motion supervision (D.1)Online state (F.3)
Reference-supervised SFTReference-derived 15D motionReference-derived 15D motionCurrent executed 15D motion
SUDA-supervised SFT (Ours)Executed 15D motionExecuted 15D motionCurrent executed 15D motion

SUDA supervision constructs the current motion context from measured states and predicts their subsequent measured responses, aligning training data with the physical quantities observed during online execution. Both conditions share RGB, language, architecture, and training settings. Tables 3 and A11 report an SR improvement from 41.5% to 64.5%, an nDTW improvement from 0.336 to 0.628, and ATE/FDE reductions of 0.905/1.797m.

Q3 GPT-5.5 Policy

What is the GPT-5.5 Policy (E.1), and was it trained on SUDA-UAV?

A3. Thank you for requesting this clarification. GPT-5.5 Policy directly calls the official multimodal API and receives zero task-specific training on SUDA-UAV. It evaluates a general multimodal model used directly as a flight policy and validates the platform's unified closed-loop execution and evaluation workflow.

PolicySUDA-UAV task-specific trainingSR↑nDTW↑ATE(m)↓FDE(m)↓OR@1m
GPT-5.5 PolicyZero-shot51.5%0.5281.5002.92864.0%
SUDA-supervised SFT (Ours)Executed-data adaptation64.5%0.6280.7751.40368.0%
AuthorsOpenReview ↗

Response to Reviewer uLTN (2/2)

Both policies are evaluated on the same 200 tasks using the originally reported protocol. GPT-5.5 provides a zero-training reference for direct multimodal-model control and validates the platform’s closed-loop workflow. Matched comparisons within each trained policy family quantify the benefit of executed data, with a mean SR gain of 9.9 percentage points [7.2, 12.6] across eight families.

Q4 Changes in dynamics

Does the gain still hold when the dynamics change - i.e that’s where dynamics awareness should show up ?

A4. Thank you for suggesting a direct dynamics change. We hold the SFT weights fixed and increase vehicle mass by 20%, from 1.6186kg to 1.9423kg, on the supplementary backend. Both conditions complete the same 200 tasks with matched initialization, controller parameters, camera interfaces, and execution budgets. Using the original reported SR protocol, the mass change produces a +1.0 percentage-point SR difference, with close trajectory metrics. These results characterize the fixed policy’s sensitivity to the tested mass variation.

Fixed-policy testTasks per conditionSR changenDTW changeATE changeFDE change
20% mass increase200+1.0 pp+0.0010−0.0024m−0.0035m

Only mass changes. Inertia, rotors, motors, PX4 parameters, initialization, random sequences, and the camera interface remain paired. SR changes from 54.0% to 55.5%, with close trajectory metrics. Supplementary Table S4 gives the settings. The result supports similar aggregate performance for the fixed policy under the tested mass change.

Supplementary Table S1 verifies platform dynamics through 12 matched target-frequency, speed, and control-mode conditions, each repeated three times. The SUDA−AirSim tracking-RMSE difference ranges from −1.7 to +15.3mm.

Q5 Training randomness

Each checkpoint is trained once (D.3). Can you add 2 or 3 seeds, at least for the smaller gains? A5. Thank you for suggesting independent training repeats. We added three PPO seeds with a common SFT initialization and the same budget of 511 updates and 4085 trajectories. One run has completed training and evaluation on all 200 tasks. The remaining two runs are in progress.

Under the original reported SR protocol, the completed repeat gains five successful tasks and loses two relative to its matched SFT parent, giving a net SR improvement of 1.5 percentage points on the supplementary backend.

Completed additional repeatEvaluation tasksSR changeATE changeFDE change
PPO over matched SFT parent200+1.5 pp−0.0032m−0.0384m

The submitted paper reports the following paired-task statistics.

ComparisonSR change (pp)Paired-task 95% interval
Eight-family data adaptation+9.9[7.2, 12.6]
SFT over T2A+37.0[30.0, 44.0]
PPO over SFT+2.0[−1.5, 5.5]
DepthAux over SFT+0.5[−4.5, 5.5]

These intervals quantify task variation for fixed checkpoints. The additional independent runs provide the training-repeat measurements, with the three-run mean and sample standard deviation to follow completion of the remaining evaluations.

Q6 FDE discrepancy

Tables A8 and A9 give different FDE for the same checkpoints and even flip whether FDE improves on average - which one is correct?

A6. Thank you for checking the numerical consistency. A8 contains an FDE transcription error introduced when duplicate appendix columns were removed and the table was reformatted. A9 agrees with the task-level records for the same 200 tasks. The correction is

StatisticOriginal A8 FDE w/o / w/ (m)Correct FDE w/o / w/ (m)
VAMPO1.545 / 1.3021.777 / 1.715
Eight-family mean1.697 / 1.6021.837 / 1.890

FDE retains the Euclidean distance from each final position to its reference endpoint, averaged equally over 200 tasks. Unrounded VAMPO values are 1.776837/1.714779m, and the eight-family means are 1.836977/1.890251m. Supplementary Table S7 lists all corrected family values. QwenVLA-style SFT reduces FDE from 3.200m to 1.403m, alongside the reported improvements in SR, nDTW, and ATE. Supplementary Table S7 provides the corrected family-level endpoint statistics.

Thanks for the response, but the reference baseline does no better than hovering and the input-mismatch run is still missing and hence i'll keep the score. Few questions :-

  1. Since the hover gets a 41% SR almost same as the ref baseline and the 8 family ref mean, what does the w/ vs w/o gain actually show?
  2. Your A2 table confirms the reference baseline trains on reference derived state but flies on executed state - could you still add the run with reference targets and executed state input?
  3. The 20% mass change barely moved the paths. Could you perhaps change something that actually alters the flown response and test both supervision types?
Reviewer_yJs8OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

SUDA targets the mismatch between intended UAV trajectories and the motion and onboard observations actually produced by a flight controller. It integrates Isaac Sim, Pegasus-related multirotor components, PX4 SITL, synchronized logging, and parallel rollouts into SUDA-SIM. SUDA-UAV contains 13,678 trajectories and 1,601,000 synchronized frames, with 13,478 trajectories for training and 200 held out by trajectory ID. The 15D prediction target combines displacement, relative rotation, velocity, angular velocity, and acceleration. Seven channels are projected into PX4 setpoints; eight provide response-only supervision. A QwenVLA-style flow policy is trained through T2A and multimodal SFT, with PPO post-training and a separate DepthAux variant. Across eight adapted policy families, the reported mean SR increases from 41.4% to 51.3%. The matched QwenVLA-style comparison reports 41.5% versus 64.5% SR for reference-derived versus controller-executed supervision. The evidence concerns controlled, in-distribution simulation, rather than demonstrated real-flight transfer.

Critical strengths

1.A useful platform and supervision interface for a real control-learning mismatch. The paper distinguishes the task reference, PX4 setpoint, realized response, and next observation, rather than treating a desired pose trajectory as executed behavior. Synchronized data generation and a shared PX4 evaluation protocol could be valuable infrastructure for aerial VLA research. The contribution should be understood as an integrated data and benchmarking system building on existing simulators, rather than the introduction of PX4 lockstep itself. This positioning is consistent with the authors' Appendix A.1 and prior Pegasus and OmniDrones systems [R1, R2]. 2.The main data-adaptation result is broader than a single policy demonstration. Eight policy families improve in reported mean SR by 9.9 percentage points, with a paired task-bootstrap interval of [7.2, 12.6], and all eight improve in nDTW. The within-family design and shared execution protocol are meaningful controls. The QwenVLA-style supervision comparison is also substantial, at 41.5% versus 64.5% SR. The authors appropriately report that PPO's smaller gain has an interval overlapping zero and that DepthAux does not improve path scores. These qualifications help separate the main platform/data contribution from exploratory refinements (Tables 2, 3 and A10).

Critical weaknesses

1.The mechanism and robustness of the claimed benefit remain insufficiently isolated. The main comparison supports end-to-end adaptation within a fixed PX4 simulator. It does not isolate the value of all 15 targets from that of the seven execution-facing channels. Appendix D.1 also changes the training state input from reference-derived to executed motion across the pair, so the result should not be attributed to target provenance alone without a controlled input ablation. More fundamentally, Appendix C labels displacement and velocity in full attitude-dependent body frames, whereas Appendix F consumes them as yaw-aligned command inputs and treats body-z angular velocity as Euler yaw rate. These are approximations outside near-level flight, and predicting a realized response does not by itself guarantee that reusing it as a setpoint increment reproduces that response. A logged-response replay through the actual adapter would directly quantify this issue. Benchmark calibration also matters because the median reference path length is only 1.00 m and some success predicates permit partial-path matches; a hover baseline, category-level results, and threshold sensitivity would clarify absolute task difficulty. Each checkpoint is trained once, so task-bootstrap intervals do not establish training-run robustness [R10]. The paper acknowledges the in-distribution and single-run scope; my concern is the strength of the mechanism claim within that scope, not a requirement to demonstrate real flight. PPO's +2.0-point SR interval of [-1.5, 5.5] and DepthAux's mixed results should remain exploratory. 2.The experimental implementation and numerical provenance are not presently verifiable from the supplied materials. Appendix A.4 states that the submission package includes the code required for the pipeline, whereas key conversion, execution, environment, metric, and bootstrap functions in the package provided for this assessment explicitly omit their algorithmic bodies. There are also substantive discrepancies in visible code. The main loader and QwenVLA.forward/predict_action do not expose the claimed current-15D-state conditioning. The trajectory PPO entry imports a mean function that evaluates a single velocity prediction at zero actions and a middle timestep, followed by tanh, instead of the seeded ten-step flow integration in Appendix D.6.3. It computes Gaussian log-probabilities on already clipped samples, whereas the appendix specifies the unclipped latent sample. Separately, Table A8 reports VAMPO FDE as 1.545/1.302 m, while Table A9 reports 1.777/1.715 m for the same named conditions; the discrepancy extends to other families. These observations do not prove that the original experiments used the visible implementation, but prevent establishing which implementation and records generated the results. For a platform and benchmark contribution, an exact versioned artifact and reconciled per-task records are central to the assessment (Appendices A.4, D.1, D.6.3 and E).

Questions for authors

1.Which exact artifact generated the reported checkpoints? Please identify the commit and entry points implementing current-state conditioning, next-step label alignment, the ten-step flow mean, and the unclipped-sample PPO likelihood. Can you provide an executable minimal pipeline and the fixed train, validation, and 200-task ID manifests 2.What drives the matched-supervision improvement? Can you compare 7D-only with 15D supervision while keeping the architecture, state input, data, budget, and execution interface fixed, and clarify the state supplied to both conditions at online evaluation? For PPO, why should reward-independent response-only samples impose physical consistency without a response-matching loss, and how does an execution-channel-only likelihood compare? These controls would materially affect interpretation of the 15D and PPO mechanisms. 3.How should reference-guided improvement be distinguished from path optimization, and can SUDA support reference-free navigation? The PPO reward includes reference-tracking terms, but the optimality of the reference trajectories has not been established. Could these terms penalize alternative routes that satisfy the task semantics while being safer or more efficient? In an unseen environment without a precise prior map or reference trajectory, could SUDA support navigation based on onboard images, UAV state, and a specified task goal? What additional components and safety mechanisms would this require? Please clarify how safety would be enforced and evaluated, and how local relative optimality would be defined and assessed against feasible alternatives under explicit task and safety constraints, using objectives such as path length, completion time, or energy use. Please distinguish capabilities demonstrated in this paper from potential future extensions.

5 comments
AuthorsOpenReview ↗

Response to Reviewer yJs8 (1/3)

Response to Reviewer yJs8

We thank you for recognizing the control/learning gap, synchronized data, and cross-model comparison, and for examining representation, execution, and reproducibility in detail.

Q1/W2 Executable materials and FDE provenance

Which exact artifact generated the reported checkpoints? Please identify the commit and entry points implementing current-state conditioning, next-step label alignment, the ten-step flow mean, and the unclipped-sample PPO likelihood. Can you provide an executable minimal pipeline and the fixed train, validation, and 200-task ID manifests

A1. Thank you for examining the implementation details. The supplementary code provides an executable workflow from data conversion and T2A/SFT through PPO and PX4 closed-loop evaluation, including dependencies, minimal run instructions, checkpoint recovery, and the fixed 200-task manifest. The four computations you identify are implemented as follows.

ComponentImplemented computation
Current-state conditioningThe current measured 15D motion state enters a state projection and conditions the policy together with RGB and language
Next-step labelsObservations at the current boundary predict the following 16 motion responses, aligned using adjacent measured timestamps
Conditional flow meanTen flow-integration steps from fixed initial noise produce the 16-step prediction
PPO likelihoodOld/new joint log-likelihoods are evaluated on the Gaussian sample before clipping

The supplementary code provides an executable reconstruction of the documented pipeline. Training and execution share the 15D channel ordering and P01/P99 scaling statistics, with inverse scaling before PX4 setpoint construction. The implementation includes forward, backward, and checkpoint-reload checks for current-state conditioning, together with next-step label alignment, ten-step flow integration, and unclipped-sample PPO likelihoods.

The A8/A9 discrepancy you identify is an FDE transcription error introduced during appendix column reduction and reformatting. A9 agrees with the task-level statistics. The correction is

StatisticOriginal A8 FDE w/o / w/ (m)Correct FDE w/o / w/ (m)
VAMPO1.545 / 1.3021.777 / 1.715
Eight-family mean1.697 / 1.6021.837 / 1.890

FDE is the Euclidean distance from each final position to its reference endpoint, averaged equally over the same 200 tasks. Unrounded VAMPO values are 1.776837/1.714779m, and the eight-family means are 1.836977/1.890251m. Complete family results appear in Supplementary Table S7.

Q2/W1 State, supervision, and execution

What drives the matched-supervision improvement? Can you compare 7D-only with 15D supervision while keeping the architecture, state input, data, budget, and execution interface fixed, and clarify the state supplied to both conditions at online evaluation? For PPO, why should reward-independent response-only samples impose physical consistency without a response-matching loss, and how does an execution-channel-only likelihood compare? These controls would materially affect interpretation of the 15D and PPO mechanisms.

A2. Thank you for distinguishing state inputs, response supervision, and execution channels. The 15D representation contains displacement, relative rotation, velocity, angular velocity, and acceleration, with three channels each. Current 15D motion provides the policy's motion context, future measured 15D responses supply per-channel supervision, and three displacement, three velocity, and one body-z angular-rate channel form the seven-channel execution interface. Relative rotation, the other angular-rate channels, and acceleration remain response-supervision targets sharing the decoder with the execution channels.

Data adaptation and online inputs

D.1 and F.3 specify the paired conditions below.

ConditionTraining state (D.1)Future-motion supervision (D.1)Online state (F.3)
Reference-supervised SFTReference-derived 15D motionReference-derived 15D motionCurrent executed 15D motion
SUDA-supervised SFT (Ours)Executed 15D motionExecuted 15D motionCurrent executed 15D motion

Both conditions share RGB, language, architecture, and training settings. Executed data jointly aligns training-time motion context and future responses with the states produced by the flight controller. The reported matched comparison on the same 200 tasks gives

AuthorsOpenReview ↗

Response to Reviewer yJs8 (2/3)

ConditionSR↑nDTW↑ATE(m)↓FDE(m)↓
Reference-supervised SFT41.5%0.3361.6803.200
SUDA-supervised SFT (Ours)64.5%0.6280.7751.403
Ours−reference+23.0 pp+0.292−0.905−1.797

The matched comparison improves SR by 23.0 percentage points, together with substantial improvements in path similarity and tracking accuracy. Both conditions share RGB, language, architecture, and training settings. D.1 specifies their training-state and supervision sources, and F.3 specifies the current measured state used during online execution.

Seven-channel execution and measured responses

F.1–F.3 rotate predicted displacement and velocity into ENU using measured yaw. Position and yaw targets accumulate across substeps, while velocity and yaw rate provide feedforward.

ΔrkE=Rz(ψkmeas)d^k,vksp,E=Rz(ψkmeas)v^k,\Delta r^E_k=R_z(\psi^{\mathrm{meas}}_k)\hat d_k,\qquad v^{\mathrm{sp},E}_k=R_z(\psi^{\mathrm{meas}}_k)\hat v_k,

rk+1sp,E=rksp,E+ΔrkE,ψk+1sp,E=ψksp,E+ω^z,k/fc,ψ˙ksp,E=ω^z,k.r^{\mathrm{sp},E}_{k+1}=r^{\mathrm{sp},E}_k+\Delta r^E_k,\qquad \psi^{\mathrm{sp},E}_{k+1}=\psi^{\mathrm{sp},E}_k+\hat\omega_{z,k}/f_c,\qquad \dot\psi^{\mathrm{sp},E}_k=\hat\omega_{z,k}.

Here fcf_c is the control frequency. This mapping expresses body-frame motion as PX4 setpoints. Across 13,324 frames from the original 200 tasks, the P95 absolute roll/pitch values are 7.76°/7.95°. The P95 differences between full-attitude and yaw-only rotation are 14.93mm for displacement and 0.300m/s for velocity, quantifying the coordinate approximation over the measured flight range.

We directly quantify response re-execution by collecting measured 15D responses for 200 curves and replaying them through the seven-channel interface with matched Iris vehicle, dynamics, actuators, and PX4 settings.

PairsPosition RMSE(m)ATE(m)FDE(m)Velocity RMSE(m/s)Yaw MAEAttitude MAE
2000.4550.3740.6030.3770.194°2.119°

Both stages use 20Hz target updates, with a mean initial-position difference of 2.86mm. These errors include coordinate mapping and PX4 re-execution dynamics. Supplementary Table S2 gives the complete settings.

Three control repetitions with the same source position reference and yaw-rate condition further examine velocity feedforward.

Control conditionPosition RMSE(m), mean±SDReduction from 4D
Position + yaw rate, 4D2.3765 ± 0.0003—
Position + velocity + yaw rate, 7D0.7809 ± 0.000667.1%
7D + derived acceleration, PVA extension0.3793 ± 0.000184.0%

Velocity feedforward reduces tracking RMSE by 67.1%, providing a control-level basis for retaining velocity in the seven-channel interface.

PPO learning signals

SFT's per-channel measured labels directly constrain the 15D physical response. PPO optimizes the shared policy using executed-stride rewards, with a prediction horizon of 16 and execution stride of 8. Old and new policies use the same initial noise ξ\xi, and joint log-likelihood sums the 15 channels of the execution prefix on the unclipped sample.

\widetilde{Y}\sim\mathcal{N}(\mu_\theta,\,0.2^2I),\qquad Y=\operatorname{clip}(\widetilde{Y},-1,1).$$ $$\ell_{\theta,i}=\sum_{h=1}^{8}\sum_{d=1}^{15}\log\mathcal N(\widetilde Y_{i,h,d};\mu_{\theta,i,h,d},0.2^2).$$ $$r_i=\exp(\ell_{\theta,i}-\ell_{\mathrm{old},i}),\qquad \hat S_i=\frac{R_i-\mu_B}{\sigma_B+\epsilon_{\mathrm{std}}},$$ $$J(\theta)=\mathbb E_i\left[\min\left(r_i\hat S_i,\mathrm{clip}(r_i,0.8,1.2)\hat S_i\right)\right].$$ Here $R_i$ is the executed-stride reward and $\epsilon_{\mathrm{std}}>0$ is a numerical stabilizer. Old/new likelihood ratios and the clipped surrogate objective optimize closed-loop performance, a class of policy-optimization mechanisms also used by Qwen-VLA and SimpleVLA-RL. [Qwen-VLA §4.2](https://arxiv.org/html/2605.30280v2#S4.SS2), [SimpleVLA-RL §3.4](https://arxiv.org/html/2509.09674v1#S3.SS4)
AuthorsOpenReview ↗

Response to Reviewer yJs8 (3/3)

Additional response to benchmark difficulty in W1

Actual PX4 hover over the same 200 starting conditions and budgets scores 41.0% SR, compared with 64.5% for Ours under the original reported evaluation protocol. At 0.1m arc-length spacing, scores are 0.327 for hover, 0.512 for Ours, and 0.145 for reference supervision. At 0.2m spacing, the corresponding scores are 0.343, 0.523, and 0.156. Ours retains higher spatial-path scores at both spacings. Supplementary Tables S3 and S6 provide the metric definitions and category results.

Q3 Reference guidance and reference-free navigation

How should reference-guided improvement be distinguished from path optimization, and can SUDA support reference-free navigation? The PPO reward includes reference-tracking terms, but the optimality of the reference trajectories has not been established. Could these terms penalize alternative routes that satisfy the task semantics while being safer or more efficient? In an unseen environment without a precise prior map or reference trajectory, could SUDA support navigation based on onboard images, UAV state, and a specified task goal? What additional components and safety mechanisms would this require? Please clarify how safety would be enforced and evaluated, and how local relative optimality would be defined and assessed against feasible alternatives under explicit task and safety constraints, using objectives such as path length, completion time, or energy use. Please distinguish capabilities demonstrated in this paper from potential future extensions.

A3. Thank you for this detailed suggestion on platform applications. The paper concerns a simulation platform, executed-data generation, and unified policy validation. References provide common conditions for data collection, training rewards, and evaluation. Online, the policy generates controls from RGB, language instructions, and the current measured 15D state. Task SR evaluates instruction requirements, while trajectory metrics and tracking rewards assess execution of the specified motion.

Tracking terms score distance to the demonstration, so two routes satisfying the same instruction can receive different rewards. The paper optimizes completion and control under a specified task and reference. Local optimality in route length, time, or energy requires comparison against feasible alternatives under common goals and safety constraints.

SUDA's onboard-image, state-feedback, and flight-control interfaces provide integration points for navigation without a prior map. Environment memory, local-route generation, and setpoint safety checks can connect to the existing closed-loop workflow as subsequent components. Your suggestion extends the platform's application directions, and we appreciate it for future study.

Reviewer_CTBYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents SUDA, a simulation, data-generation, and benchmarking framework for dynamics-aware UAV policy learning, aiming to address the gap that existing UAV datasets provide reference trajectories but not actual controller- and dynamics-produced motion responses. Extensive experiments demonstrate the potential value of incorporating flight-controller execution into training data.

Critical strengths
  1. The paper identifies a meaningful discrepancy between geometric reference trajectories and actual UAV execution.
  2. SUDA-SIM performs comprehensive simulations, and its unified infrastructure could be useful for UAV learning research and reproducible benchmarking.
  3. Using matched model architectures and training configurations, the authors compare reference-based and controller-executed supervision, reporting multiple closed-loop metrics and bootstrap confidence intervals.
Critical weaknesses
  1. The proposed method is the integration of existing components. Although the system integration is practically valuable, the paper introduces limited new learning algorithms or theoretical insights.
  2. All experiments are conducted in simulation, without real UAV flight validation. Simulated dynamics cannot fully capture real-world disturbances, actuator uncertainty, sensor noise, and hardware latency.
  3. While the average success rate and nDTW improve, not all methods benefit consistently across all metrics.
Questions for authors
  1. Since all experiments use the same PX4 SITL framework, how would it perform under different UAV dynamics, controller parameters, payloads, wind disturbances, or sensor noise?
  2. How do the authors explain the substantial decrease in OR@1m for the QwenVLA-style policy after SUDA-UAV adaptation?
  3. Can the authors provide evidence that the learned policies transfer to real UAV hardware or to an independently configured simulator?
4 comments
AuthorsOpenReview ↗

Response to Reviewer CTBY (1/2)

Response to Reviewer CTBY

We thank you for recognizing the execution gap, unified infrastructure, and matched comparisons.

W1 Contribution

SUDA synchronizes reference motion, controller setpoints, realized responses, and next observations into executed data, and evaluates policies through a shared interface. All eight policy families improve in reported SR and nDTW under the shared evaluation protocol, supporting the cross-model value of controller-executed supervision. Attribution of the underlying simulation infrastructure follows A.1.

Policy familyReference SRSUDA SRSR change (pp)nDTW change
VAMPO49.0%63.0%+14.0+0.103
FastWAM37.0%41.0%+4.0+0.009
ABot-M044.0%63.0%+19.0+0.100
OpenVLA-OFT42.0%51.5%+9.5+0.025
Pi0.534.5%49.0%+14.5+0.036
OpenVLA-7B47.0%52.5%+5.5+0.005
UAV-Flow-OpenVLA45.5%51.5%+6.0+0.009
LingBot-VLA32.0%38.5%+6.5+0.010
Eight-family mean41.4%51.3%+9.9+0.037

Reference and SUDA supervision within each family use matched architectures and training settings. Mean OR improves by 5.1 percentage points [3.3,7.1], and mean nDTW by 0.037 [0.030,0.044].

Q1 Changes in dynamics and control conditions

Since all experiments use the same PX4 SITL framework, how would it perform under different UAV dynamics, controller parameters, payloads, wind disturbances, or sensor noise?

A1. Thank you for suggesting evaluation under changed dynamics. We hold the SFT weights fixed and complete the same 200 tasks at nominal mass and with a 20% mass increase. Initialization, random seed, controller, camera interface, execution budget, and A4 predicates are matched.

Mass condition, fixed SFT weightsNSR↑nDTW↑ATE(m)↓FDE(m)↓
Nominal mass20054.0%0.56331.21552.2906
Mass increased by 20%20055.5%0.56431.21312.2871

Aggregate completion and trajectory metrics remain close under the tested mass change. Supplementary Table S4 gives the physical parameters and OR results.

We further examine the execution model through 12 matched control conditions combining 20/50Hz target updates, two speeds, and P4/PV/PVA, with three repetitions each. P4 uses three-dimensional position and yaw targets, PV adds velocity feedforward, and PVA adds acceleration feedforward. The SUDA−AirSim position-tracking RMSE difference ranges from −1.7 to +15.3mm. Supplementary Table S1 gives the complete matrix. Mass sensitivity evaluates a fixed policy under changed conditions, while standard-trajectory tests quantify backend control responses.

Q2/W3 OR@1m and category-dependent outcomes

How do the authors explain the substantial decrease in OR@1m for the QwenVLA-style policy after SUDA-UAV adaptation?

A2. Thank you for highlighting this metric difference. OR@1m measures whether any point on the executed path enters the demonstration endpoint’s 1m neighborhood, including the starting point. SR evaluates instruction completion through task-specific predicates. In the reported QwenVLA-style comparison, OR@1m changes from 78.0% to 68.0%, while SR increases from 41.5% to 64.5%, ATE decreases from 1.680m to 0.775m, and FDE from 3.200m to 1.403m. The endpoint-occupancy and category analyses below explain how these metrics capture different execution outcomes.

ConditionOR@1mInitially within 1mFinally within 1m
Reference-supervised SFT156/200114/20027/200
SUDA-supervised SFT (Ours)136/200114/200127/200

Final endpoint occupancy improves by 50.0 percentage points. Among the 86 initially outside tasks, final hits increase from 4 to 19. OR measures visitation of the selected demonstration endpoint, while category SR evaluates the required spatial relation or motion event. Different valid terminal locations can produce different reference-endpoint distances. For example, reaching the left side of a vehicle can admit locations beside its center or rear, with different distances to one demonstration endpoint. Category SR and path nDTW jointly assess semantic completion and demonstration-path agreement.

AuthorsOpenReview ↗

Response to Reviewer CTBY (2/2)

Task familyNReference OR@1mOurs OR@1mReference SROurs SR
3D translation5646.4%48.2%16.1%23.2%
Object-distance control4885.4%83.3%66.7%87.5%
Object-goal and side relations3794.6%70.3%78.4%73.0%
Pass-through1770.6%5.9%76.5%35.3%
Turning32100.0%100.0%0.0%96.9%
Orbiting10100.0%100.0%0.0%100.0%
Overall20078.0%68.0%41.5%64.5%

3D translation combines horizontal, vertical, and combined vertical/side motion. Object-distance control combines approach and retreat, while object relations combine goal-position and side-position tasks. Grouped percentages are calculated from the total success counts and task counts, retaining all 200 tasks and the original reported SR protocol.

Fine-grained UAV language control already combines semantic SR and path nDTW. We accordingly report task completion together with spatial-path quality and endpoint distance. UAV-Flow §3.2

Q3 Independent simulation and real flight

Can the authors provide evidence that the learned policies transfer to real UAV hardware or to an independently configured simulator?

A3. Thank you for suggesting practical transfer evidence. The same SFT weights complete 200 tasks on the independently configured supplementary backend, scoring 54.0% SR and 0.5633 nDTW. With a 20% mass increase, SR is 55.5% and nDTW is 0.5643, providing closed-loop results for a fixed policy under changed execution conditions. Supplementary Table S4 gives the settings.

To examine the dynamics underlying execution, we use AirSim as a physical reference because its motion equations are public and its square/circle experiments compare simulation and real flight under the same Offboard commands. SUDA uses PX4 control messages and state feedback, and examines free-flight responses using the same rigid-body, rotor thrust/torque, and drag formulas. AirSim paper, PX4 interface, AirSim physics implementation

This comparison tests whether the platform reproduces the control response of a traceable reference backend. Across 12 matched conditions combining 20/50Hz target updates, two speeds, and P4/PV/PVA control, each repeated three times, the SUDA−AirSim tracking-RMSE difference ranges from −1.7 to +15.3mm. On the same hardware, the physical update kernel at 65,536 parallel states achieves 16.02× throughput relative to the serial AirSim-source driver. Supplementary Table S1 gives the conditions.

We also directly track public Real position-time curves in SUDA, deriving velocity and acceleration feedforward and running three repetitions. H denotes horizontal symmetric Hausdorff distance.

TrajectoryPublic AirSim curve, H to Real (m)SUDA tracking Real target, H (m)SUDA tracking Real target, time RMSE (m)
Square0.64610.3074 ± 0.00040.2185 ± 0.0005
Circle1.46980.4664 ± 0.00040.2917 ± 0.0001

The AirSim values come from the original same-Offboard-command sim/real experiment. SUDA values come from PVA tracking of the public Real output curves. The column labels identify these input roles, and the public Real trajectories provide verifiable spatial and temporal references. A separate experiment reconstructs a high-level orbit target from the official real-flight log and examines velocity feedforward, as described in Supplementary Table S5.

The shared protocol, traceable motion model, and matched trajectory responses provide a basis for connecting SUDA to AirSim studies using compatible control assumptions. Real-flight validation is subject to current flight-approval and experimental constraints. The paper contributes a simulation platform, executed-data generation, and unified policy evaluation.

Dear Reviewer CTBY, It seems we have not received your feedback on our response yet. As the discussion period will end soon, we were wondering if there are anything else we could address to help further clarify? We sincerely look forward to your feedback. If these additions have addressed your concerns, would you consider updating your rating? Thanks again for your time!

Thank you for the detailed rebuttal and additional experimental results. I appreciate the authors' efforts to address my concerns, particularly through the matched comparisons across eight policy families, robustness evaluation under increased UAV mass, and validation against an independently configured simulation backend. I believe several concerns remain only partially addressed, such as the novelty and real-world transfer. I would prefer to keep the score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.