DroneWAM: Efficient World Action Model for Drone Visual Navigation
Liang Yao, Fan Liu, Hongbo Lu, Wei Xu, Jianyu Jiang, Yijun Shen, Chuanyi Zhang, Pai Peng
Hohai University · Shanghai Jiao Tong University · Nanjing University · Southeast University · East China Normal University · CowaRobot
Published September 27, 2026 · on the feed September 29, 2026 · ▲ 5 · arXiv 2609.33148
The 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. https://github.com/1e12Leon/DroneWAM{Codes and data} will be released.
The Brief
2:15A short podcast summary of the paper. Written and voiced by AI.
Transcript
DroneWAM. Efficient World Action Model for Drone Visual Navigation. By Liang Yao and seven others, from Hohai University, Shanghai Jiao Tong University, and four other institutions. Published on the twenty-seventh of September. Its code repository is a placeholder: a readme, and nothing to run yet.
A world-action model lets a visual navigation agent anticipate how each candidate action would change what it sees next, and then act on those predicted consequences. It is imagination put to work for control. A drone needs that ability under tight constraints on both accuracy and efficiency. Imagining the future is only useful if it is precise enough to steer by and cheap enough to repeat, step after step.
DroneWAM is built as an efficient world-action model for exactly that setting. Its central choice is to predict the future in representation space rather than in pixels. It adopts a JEPA-based architecture that models future states directly as representations, so the expense of explicitly generating future images is avoided altogether.
Two further savings follow. A pretrained Resampler compresses the dense features from the encoder into a smaller number of latent tokens, which cuts the computation that is repeated at every imagined step. And adaptive rollout lets the model decide how far ahead to imagine. A Gate, trained on preferences, allocates prediction depth according to the current scene, so some scenes get a longer look ahead than others. To train under the richer motion of flight, the authors also build DroneNav-6D, a simulated visual navigation dataset. It pairs synchronised RGB observations with flight trajectories in six degrees of freedom, control commands, and randomised wind disturbances.
On DroneNav-6D, DroneWAM achieves the best trajectory accuracy among the methods compared. Adaptive rollout brings the average prediction depth down from eight steps to four point five eight, and trajectory accuracy improves at the same time. The authors take this as evidence that predictive computation can be spread more effectively across scenes. Code and data are to be released.
Every design decision here trims the cost of imagining. Skip the pixels, shrink the tokens, and let each scene decide how deep to look. The notable result is the last one: cutting average depth nearly in half made trajectories more accurate rather than less.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is a drone that must imagine the future on a tight budget.
You are building a drone that navigates by imagining how each move will change what its camera sees next. It must re-imagine the future at every step, on a small onboard computer.
The road ahead
- 1 · The problem ~7 min A drone that steers by imagining its future views must repeat that imagining at every step, so each imagined step has to be cheap.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Predict the future as compact learned features, not images, shrink those features to fewer tokens, and let a learned Gate choose how far to look ahead.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Encode each view, compress it to few tokens, imagine future states as features step by step, and let a trained Gate say when to stop.
- 6 · The experiment ~6 min You’ll guess: When DroneWAM's Gate let the drone imagine fewer steps ahead on average, what happened to its trajectory accuracy?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.