acceptodds
Withdrawn from ICLR 2027

DroneWAM: Efficient World Action Model for Drone Visual Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

World-action models give visual navigation agents a way to anticipate how candidate actions will change future observations and to act from the predicted consequences. For drones, this capability must operate under tight accuracy and efficiency constraints. We present DroneWAM, an efficient world-action model for drone visual navigation. DroneWAM adopts a JEPA-based architecture to model future states directly in representation space, avoiding the cost of explicit future image generation. A pretrained Resampler further compresses dense encoder features into fewer latent tokens, reducing the computation repeated at each imagined step. We also introduce adaptive rollout, where a preference-trained Gate adaptively allocates prediction depth according to the current scene. To support learning under richer aerial motion, we construct DroneNav-6D, a simulated visual navigation dataset with synchronized RGB observations, 6-DoF flight trajectories, control commands, and randomized wind disturbances. On DroneNav-6D, DroneWAM achieves the best trajectory accuracy among the compared methods. Adaptive rollout further reduces the average prediction depth from 8 to 4.58 while improving trajectory accuracy, demonstrating that predictive computation can be allocated more effectively across scenes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Model, UAV Navigation, Visual Navigation

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_f62UOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes an architecture for training behavioral cloning policies for aerial vehicles. The key contributions are:

  1. Adaptive depth of prediction. Traditional world models output a fixed number of trajectories, while this paper proposes a mechanism to stop early if the current action is deemed sufficient (by another module).
  2. A dataset built in Airsim to evaluate the proposed architecture.

The validation relies mainly on open-loop trajectory prediction, since the closed-loop prediction fail (both for the proposed method and the selected baselines). There are a lot of ablation of the approach, showing that the key technical contributions (resampling, and adaptive rollout) bring about relatively small improvements in open loop trajectory prediction.

Critical strengths

The main strength is the idea that the sampling should not be fixed but adaptive according to the scene and task complexity. This concept is well understood and used in model-based RL (see for example "Adaptive Complexity Model Predictive Control", Norby et al.). Applying this concept to a vision-based navigation policy is something I have not seen before. It is however very similar to work in manipulation ( see for instance "ELASTIC: Efficiently Learning to Adaptively Scale Test-Time Compute for Generative Control Policies"), but more specific to future state prediction.

Critical weaknesses

The most critical limitation is that the proposed approach is not capable of completing the tasks in closed loop. Since open-loop prediction is barely within the scope of this paper (the claims are specifically about control of resource-constrained platform), it seems that more work is necessary to accomplish this goal.

It is unclear why all the selected baselines fail: image-goal navigation is a well-studied topic and a lot of specifically designed baselines exist (see for instance "GOAT: Go to Anything"). Could it just be that the proposed dataset is not diverse enough, and all methods overfit? This potentially justifies why the prediction errors are so low (centimeter-level), but the closed-loop performance is poor. Overall, I did not understand why creating a new dataset, since so many large datasets already exist (see for example "OpenFly: A comprehensive platform for aerial vision-language navigation").

There are also a few parts in the method that are not fully justified. As far as I understood, several trajectories are predicted from the same starting state sts_t, and the predicted trajectories are then used to generate actions. The trajectories are not scored against each other (as typical in planning methods), but rather always used to predict the ground-truth action. Since the learning signal is the same for all trajectories (See eq. 5), what is stopping them to collapse into a single prediction (zt+k=zt+k+1∀kz_{t+k} = z_{t+k+1} \forall k)? There is no visualization or analysis that the trajectories are different. Looking at the error heatmap in Fig. 4, the actions are also barely different.

While I have previously alluded to it, another limitation worth mentioning is that there are no baselines representative of the state of the art in image based navigation, in particular for drones. This is a mature topic and therefore requires a better placement in the context of prior work.

Questions for authors
  1. There are very little technical details on the definition of states and control. Is the model predicting velocities, waypoints, or torques? What does it observe beyond the image (if anything)? At what frequency is each module updated?
  2. Similarly, there are very sparse details about the dataset: What is the average speed the drone moves? is it planar motion or 3D? How is the data generated? Why is this better than previous datasets?
  3. In table 3, why adding planning makes the model so much faster? Shouldn't it be faster if you only predict one set of actions?
  4. In table 4, the results show that stopping computation at random has a very similar performance than the proposed DPO-style approach (difference is on the second decimal digit). It should be clarified why that is the case.
Reviewer_uuWfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DroneWAM is a JEPA-style world-action model for drone image-goal navigation. It uses a Resampler to compress frozen ViT-L features from 512 to 128 tokens, a Predictor to predict future latents, a Planner to output 8-step actions, and a DPO-trained Gate to stop prediction early and return the best-scoring plan. The authors also build DroneNav-6D, an AirSim dataset of 3,837 clips over five worlds. Open-loop, DroneWAM beats five world-model baselines on trajectory error and latency. Closed-loop, no method completes the main benchmark, and DroneWAM reaches 88% success on 50 easier routes.

Critical strengths

The efficiency question is well posed and the ablation is set up to answer it: spatial and temporal compression are tested separately, error is reported together with latency, and the same mechanism is also tested on LIBERO.

Critical weaknesses

The accuracy gain comes from the Planner, not from the two efficiency components. In Table 3, the Planner takes Rel.ATE from 0.6621 to 0.0584, while the Resampler and adaptive rollout move it to 0.0603 and 0.0578. On DroneNav-6D adaptive rollout is compared only with fixed K=8K=8, not K=1K=1, 22 or 44; on LIBERO (Table 12) fixed D1 takes 129.7 ms and adaptive 411.9 ms, at N.RMSE 0.2735 vs. 0.2652. Since efficiency is the central claim, my assessment rests on this.

Questions for authors

Could you report fixed K=1K=1, 22 and 44 on DroneNav-6D, and the adaptive model when it returns the last plan rather than the best-scoring one?

Reviewer_8HqkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DroneWAM, an efficient world-action model for image-goal drone navigation. The method predicts future states in latent space using a JEPA-based architecture, compresses visual features from 512 to 128 tokens through a pretrained Resampler, and employs a preference-trained Gate to adaptively determine the rollout depth for each input. The authors also introduce DroneNav-6D, an AirSim-based dataset containing 3,837 trajectory clips, approximately 1.59 million RGB observations, and 63.03 hours of flight with synchronized 6-DoF motion, control commands, and randomized wind disturbances. Experiments on DroneNav-6D report improvements in trajectory accuracy and inference latency over several world-model and world-action baselines. Adaptive rollout reduces the average prediction depth from 8 to approximately 4.58 while improving trajectory metrics. Additional experiments on LIBERO investigate whether adaptive rollout transfers to robotic manipulation.

Critical strengths
  1. The decomposition into latent-space prediction, compact visual representations, and adaptive rollout provides a coherent approach to reducing the computational cost of repeated future prediction during navigation.
  2. In Table 4, Random Stop and the proposed method use almost identical average rollout depths: 4.581 and 4.578, respectively, but Random Stop performs substantially worse. This is stronger evidence than comparison with fixed-depth rollout alone, because it suggests that the learned Gate improves computation allocation rather than merely truncating prediction.
  3. The paper separately examines the Planner, Resampler, adaptive rollout, and number of latent tokens. The token-count study also identifies a reasonable accuracy–efficiency trade-off rather than reporting only the fastest configuration.
  4. Appendix D reports dataset composition, motion distributions across all six degrees of freedom, trajectory statistics, and motion over different prediction horizons. These analyses are useful for understanding the supervision provided by DroneNav-6D.
Critical weaknesses

\textbf{1. The distinction from RISE, the closest adaptive-imagination work, is insufficient.} RISE is cited, but the paper does not substantively explain how its adaptive-rollout mechanism differs. Both works address fixed imagination budgets through sequential Roll/Stop decisions, consider planning quality and computational cost, and evaluate Random Stop and Latent Margin as alternative stopping strategies. Consequently, the algorithmic novelty of the proposed Gate is currently unclear. Applying adaptive imagination to drones and using DPO-based preference training may be useful, but the paper must establish why these differences constitute a substantive methodological contribution. A controlled comparison with a RISE-style Scheduler using the same backbone, data, and computational budget is needed.

\textbf{2. The main evaluation does not adequately support the 6-DoF claims.} Tables 1–5 report trajectory errors based on translation, while Appendix C explicitly defines RPE using consecutive translation increments and evaluates rotational accuracy separately. However, rotational or attitude errors are absent from the main results. Demonstrating that the dataset contains six motion components does not establish that the model predicts them accurately. The main evaluation should report roll, pitch, and yaw errors or a geodesic rotation metric, together with experiments demonstrating the benefit of 6-DoF supervision over a 4-DoF alternative. In addition, Appendix C states that ego-frame translations are transformed into the global frame using only the current yaw. If actions are defined in a full body-frame 6-DoF convention, the paper should explain why roll and pitch can be omitted from this transformation or revise the integration accordingly.

\textbf{3. The efficiency results do not yet substantiate the onboard-deployment motivation.} The paper reports synchronized GPU latency but does not identify the inference GPU, CPU, precision, software environment, warm-up procedure, or number of timing repetitions. It is also unclear whether the reported latency includes the Encoder, Resampler, Gate, Planner, and trajectory integration. The proposed method’s 483.29 ms latency corresponds to only approximately 2.07 inference steps per second. Thus, the results support a relative latency improvement in an unspecified environment, but not necessarily efficient operation under constrained onboard computation. Hardware details, complete timing boundaries, memory consumption, and latency distributions are necessary; measurements on a resource-constrained platform would substantially strengthen the practical claim.

\textbf{4. The baseline coverage and experimental controls are insufficient for reliable attribution.} The closed-loop comparison includes only FastWAM and NWM. It lacks a same-backbone direct policy or no-imagination variant, recent navigation WAMs, and a RISE-style adaptive-rollout baseline. Without these controls, it is difficult to determine whether the gains arise from the Planner, architecture, training recipe, dataset, or the proposed adaptive mechanism. All baselines should use consistent training data, evaluation protocols, and timing conditions, with pretraining and tuning budgets clearly reported.

\textbf{5. The statistical evidence and data-splitting protocol are weak.} The experiments appear to use a single random seed, and no confidence intervals, variance estimates, or significance analyses are reported. This is particularly important for the easy-route closed-loop experiment, where only 50 trajectories are used and the reported success rates are 88%, 74%, and 72%. Episode-level bootstrap intervals or paired analyses would help establish whether these differences are reliable. Furthermore, the training and validation splits are described only as having distinct clip identifiers, with all five environments represented in both. The paper should clarify whether clips from the same original flight or nearby segments can cross splits. An independent test set, original-flight-level separation, and whole-model leave-one-environment-out evaluation would provide stronger evidence of generalization. The current cross-validation analysis primarily establishes generalization of the Gate, not of the entire model.

\textbf{6. Important details needed to reproduce the Gate and evaluation are missing.} The values or selection procedures for ϵA\epsilon_A, ϵR\epsilon_R, λc\lambda_c, β\beta and KminK_{min} are not fully specified. The threshold-calibration objective, ensemble aggregation rule, preference-data size, and use of uncertainty also require clarification. The Gate inputs described in Section 3.2 and Appendix B are not entirely consistent. Since these choices directly determine the accuracy–latency trade-off, they are central rather than incidental implementation details.

Questions for authors
  1. What are the precise methodological differences between the proposed Gate and RISE in terms of supervision, stopping decisions, plan selection, and inference computation? How do the two approaches compare using the same WAM backbone and training data?

  2. Why are attitude or rotational errors absent from the main DroneNav-6D results? What are the model’s roll, pitch, and yaw prediction errors, and how does performance change when trained or evaluated with 4-DoF rather than 6-DoF supervision?

  3. Why does the trajectory integration use only yaw when transforming ego-frame translations into the global frame? Is this consistent with the paper’s full 6-DoF action and coordinate conventions?

  4. On which inference hardware were Tables 1-6 measured? Does the reported latency include all model components and preprocessing required for deployment, and what are the corresponding memory and latency distributions?

  5. Can training and validation clips originate from the same underlying flight or from adjacent trajectory segments? Is there an independent test set, and how does the complete model perform when an entire environment is excluded from all training stages?

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

Related papers

Discussion (0)

Sign in to comment.