pith. sign in

arxiv: 2606.04534 · v1 · pith:T2UIB455new · submitted 2026-06-03 · 💻 cs.RO

MAD: Mapping-Aware World Models for Agile Quadrotor Flight

Xinhong Zhang , Runqing Wang , Yunfan Ren , Ding Yu , Boyu Zhou , Jian Sun , Fang Deng , Jie Chen
show 1 more author
Gang Wang
This is my paper

Pith reviewed 2026-06-28 06:21 UTC · model grok-4.3

classification 💻 cs.RO
keywords quadrotor flightworld modelsoccupancy mappingvision-based navigationagile flightcollision avoidancelatent dynamicsreinforcement learning
0
0 comments X

The pith

A world model that reconstructs robocentric occupancy and visibility maps from depth images produces better collision-avoidance policies for agile quadrotor flight than image-reconstruction baselines.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Agile quadrotor flight requires remembering observed space, inferring nearby obstacles, and acting under partial visibility. MAD trains recurrent latent dynamics to reconstruct robocentric occupancy and visibility grid maps plus proprioceptive states instead of raw images. This design makes the latent representation encode local geometry and ego-motion directly useful for avoidance. Policies derived from the model show higher success rates, faster speeds, and stronger transfer across navigation and racing tasks. Real-world tests on a physical quadrotor confirm safe flight at up to 5.05 m/s in forests.

Core claim

MAD learns recurrent latent dynamics that reconstruct robocentric occupancy and visibility grid maps together with proprioceptive states. This forces the latent state to encode local geometry, visibility history, and ego-motion in a form directly relevant to collision avoidance. The resulting representation supports imagination-based and feature-extractor policies that outperform vision-only baselines in success rate, speed, and cross-task transfer while also yielding interpretable map predictions and accurate ego-motion estimates.

What carries the argument

Mapping-Aware Dreamer (MAD), whose self-supervised objective reconstructs robocentric occupancy and visibility grid maps rather than raw images.

If this is right

  • MAD-based agents reach higher success rates and faster flight speeds than vision-only baselines in navigation and racing tasks.
  • The learned representation transfers better across tasks than image-reconstruction baselines.
  • The model generates interpretable occupancy and visibility predictions along with accurate ego-motion estimates from depth.
  • The same policy deploys successfully on physical hardware, attaining 9.66 m/s in simulation and 5.05 m/s in real forest flights.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The map-reconstruction objective could be combined with explicit SLAM modules to improve robustness when depth sensing is noisy or sparse.
  • Extending the grid reconstruction to include predicted future occupancy might allow longer-horizon planning without additional modules.
  • The GPU-parallel map supervision used in training suggests the approach scales to larger environments if map resolution or field of view is increased.

Load-bearing premise

Reconstructing occupancy and visibility grid maps forces the latent state to encode information directly useful for collision avoidance.

What would settle it

Vision-only baselines achieving equal or higher success rates and speeds than MAD agents on the same visual navigation and racing tasks would falsify the claimed benefit of map reconstruction.

Figures

Figures reproduced from arXiv: 2606.04534 by Boyu Zhou, Ding Yu, Fang Deng, Gang Wang, Jian Sun, Jie Chen, Runqing Wang, Xinhong Zhang, Yunfan Ren.

Figure 1
Figure 1. Figure 1: (a) Visualization of the flight scene together with the occupancy and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Mapping-Aware Dreamer (MAD) and the proposed MAD-based visual navigation system. (a) Self-supervised training of MAD in the GPU-based DiffAero simulator. DiffAero provides depth images, proprioceptive sensory information, and robocentric occupancy and visibility grid maps, and MAD is trained to infer latent states ht and zt that reconstruct these grid maps and sensory signals through a grid-map… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of local occupancy and visibility grid maps. Both maps are defined in the quadrotor-centric local frame. (a) OGMs encode whether each voxel is occupied by obstacles, while VGMs encode whether each voxel has been observed by the on-board depth camera. (b) The visibility re-marking mechanism. Voxels at time step t = k + 1 will be marked as visible, if they contain at least one anchor point of th… view at source ↗
Figure 4
Figure 4. Figure 4: Training pipeline of MAD. At each step, MAD encodes the depth observation and sensory inputs into a discrete latent representation zt. The encoder, RSSM state predictor, and decoders are jointly trained to reconstruct the designated targets from ht and zt, including the sensory observations, occupancy and visibility grid maps, continuation flags, and reward signals. C. Policy Learning MAD can be coupled wi… view at source ↗
Figure 5
Figure 5. Figure 5: Training curves of baseline algorithms and their MAD-based variants in the visual navigation task. Each curve shows the mean episode return and success rates over 8 random seeds, and the shaded regions indicate the standard deviations across runs. 3) Quality of Reconstructions: To demonstrate that our grid map reconstruction objective encourages the world model to extract odometry and attitude information … view at source ↗
Figure 6
Figure 6. Figure 6: Velocity and attitude reconstruction errors under different reconstruction strategies. Darker bars correspond to supervising MAD’s sensory decoder to estimate raw sensory inputs using non-detached ht and zt, allowing gradients to propagate through the entire model. Lighter bars use detached ht and zt, restricting gradients to the sensory decoder and preventing them from influencing the rest of MAD. The rep… view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of task transfer to the racing environment. (a) The vision-based racing task in DiffAero. (b) Top-down view of a flight trajectory controlled by a MAD-PPO agent with a transferred and frozen MAD, demonstrating smooth and agile maneuvers through narrow gates. consistent MAVLink interfaces, and accurate actuator and sensor emulation, thereby reducing the simulation-to-real dis￾crepancy. We eval… view at source ↗
Figure 7
Figure 7. Figure 7: As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of rollouts imagined by MAD. Top: depth images predicted by the prior, reconstructed by the posterior, and ground truth from the simulator. Middle: occupancy grid maps predicted by the prior, reconstructed by the posterior, and ground truth. Bottom: visibility grid maps predicted by the prior, reconstructed by the posterior, and ground truth. its grid map reconstruction task—carries richer sp… view at source ↗
Figure 10
Figure 10. Figure 10: Simulation performance of baseline algorithms in Gazebo with randomly generated cylindrical obstacles. (a) Example environment and flight trajectories of YOPO, EGO-Planner, PPO, and MAD-Dreamer for a representative test case. (b) Achieved peak flight speed as a function of the desired maximum speed (5-10 m/s) and environment sparsity (10-30 m2/corridor). (c) Success rates of YOPO, EGO-Planner, PPO, and MA… view at source ↗
Figure 11
Figure 11. Figure 11: Flight trajectories of MAD-Dreamer and three baseline algo￾rithms in an indoor corridor environment. Under the same sensing range, MAD-Dreamer exhibits markedly smoother and more agile trajectories than all baselines. outperforming EGO-Planner (14.33 s) and YOPO (16.02 s). In addition, our agent reaches a peak speed of 6.37 m/s, whereas EGO-Planner and YOPO attain 4.01 m/s and 3.71 m/s, respectively. PPO,… view at source ↗
Figure 13
Figure 13. Figure 13: Indoor flight trajectories of the proposed MAD. (a) Scenario I, a more cluttered environment where the quadrotor maintains relatively slow yet safe navigation (2.06 m/s).(b) Scenario II, where the obstacles are placed more coarse. Under the control of a MAD-Dreamer agent, the quadrotor flies at a higher speed (3.10 m/s). with irregular obstacle layouts and narrow corridors, achieving safe and consistent f… view at source ↗
Figure 12
Figure 12. Figure 12: Safe navigation visualization in a simulated dynamic en￾vironment. The dynamic scenario demonstrates the MAD-based agent’s robust capability for timely and preventive avoidance of moving obstacles (pedestrians). (a) (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
read the original abstract

Agile quadrotor flight in cluttered scenes requires more than a reactive mapping from a depth image to a control command: the vehicle must remember which regions have been observed, infer nearby occupied space, and act under partial visibility and tight latency. In this paper, we present Mapping-Aware Dreamer (MAD), a geometry-aware world model for vision-based quadrotor flight. Instead of using raw-image reconstruction as the main self-supervised objective, MAD learns recurrent latent dynamics that reconstruct robocentric occupancy and visibility grid maps together with proprioceptive states. This design forces the latent state to encode local geometry, visibility history, and ego-motion in a form that is directly relevant to collision avoidance. MAD is trained in DiffAero using a GPU-parallel map-construction module that provides high-throughput supervision for occupancy and visibility. The learned representation is used in three policy-learning modes: imagination-based MAD-Dreamer and feature-extractor variants based on PPO and SHAC. Across visual navigation and racing tasks, MAD-based agents achieve higher success rates, faster flight, and better cross-task transfer than corresponding vision-only baselines. The model also produces interpretable map predictions and accurate ego-motion estimates from depth observations. We further deploy the learned policy on a physical quadrotor with an Intel RealSense D435i and demonstrate safe indoor and outdoor flight under limited sensing, reaching 9.66 m/s in simulation and 5.05 m/s in real-world forest experiments. These results show that mapping-aware world models provide a practical middle ground between modular aerial navigation and end-to-end learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents Mapping-Aware Dreamer (MAD), a recurrent world model for vision-based quadrotor flight that replaces raw-image reconstruction with self-supervised prediction of robocentric occupancy and visibility grid maps plus proprioceptive states. The learned latent dynamics are used for policy learning via imagination (MAD-Dreamer), PPO, and SHAC variants. Across visual navigation and racing tasks the MAD agents are reported to achieve higher success rates, faster flight, and better cross-task transfer than vision-only baselines; the model also yields interpretable map predictions and ego-motion estimates. Real-world deployment on a quadrotor with an Intel RealSense D435i is demonstrated in indoor and outdoor forest settings, reaching 9.66 m/s in simulation and 5.05 m/s in hardware experiments.

Significance. If the performance advantages are causally linked to the mapping-aware objective, the work supplies a concrete middle ground between modular mapping pipelines and pure end-to-end vision policies for agile flight under partial observability. The combination of GPU-parallel map supervision in DiffAero, real-world transfer at 5 m/s, and cross-task generalization would constitute a useful empirical contribution to learning-based aerial robotics.

major comments (2)
  1. [Abstract] Abstract: the design claim that reconstructing robocentric occupancy/visibility grids 'forces the latent state to encode local geometry, visibility history, and ego-motion in a form that is directly relevant to collision avoidance' is presented as a direct consequence of the loss but is not supported by any ablation that isolates the map-reconstruction terms from proprioceptive prediction or architecture scale, nor by any probing of whether the policy actually conditions on the predicted maps during imagined rollouts. Without such evidence the reported gains in success rate and speed cannot be attributed to the mapping-aware component.
  2. Experiments section (implied by abstract claims): quantitative success rates, baseline numbers, and statistical tests are absent from the abstract and the provided summary; only peak speeds are stated. This makes it impossible to judge the magnitude or reliability of the 'higher success rates' and 'better cross-task transfer' assertions that constitute the central empirical claim.
minor comments (1)
  1. [Abstract] The simulator name 'DiffAero' appears without citation or brief description in the abstract.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on empirical validation and attribution of results. We address each major comment below, agreeing where the manuscript requires strengthening and outlining specific revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the design claim that reconstructing robocentric occupancy/visibility grids 'forces the latent state to encode local geometry, visibility history, and ego-motion in a form that is directly relevant to collision avoidance' is presented as a direct consequence of the loss but is not supported by any ablation that isolates the map-reconstruction terms from proprioceptive prediction or architecture scale, nor by any probing of whether the policy actually conditions on the predicted maps during imagined rollouts. Without such evidence the reported gains in success rate and speed cannot be attributed to the mapping-aware component.

    Authors: We agree that the abstract's design motivation would benefit from direct empirical support isolating the map-reconstruction objective. The manuscript compares full MAD agents against vision-only baselines, but does not include an ablation removing only the map terms while holding proprioceptive prediction and model scale fixed, nor explicit probing of map conditioning in imagined rollouts. In the revision we will add such an ablation (training a proprioception-only variant) and include analysis of policy dependence on predicted maps, e.g., via latent masking or rollout visualizations. These additions will allow clearer attribution of performance differences. revision: yes

  2. Referee: [—] Experiments section (implied by abstract claims): quantitative success rates, baseline numbers, and statistical tests are absent from the abstract and the provided summary; only peak speeds are stated. This makes it impossible to judge the magnitude or reliability of the 'higher success rates' and 'better cross-task transfer' assertions that constitute the central empirical claim.

    Authors: The abstract reports qualitative improvements without numerical success rates, baseline values, or statistical details. While the full manuscript contains tables with these metrics and comparisons, the abstract itself does not. We will revise the abstract to incorporate key quantitative results (success rates, mean speeds, transfer metrics) and note the statistical tests performed in the experiments section, making the central claims more precise and verifiable from the abstract alone. revision: yes

Circularity Check

0 steps flagged

No significant circularity; design claim is an assumption, not a self-referential derivation

full rationale

The paper's central claims rest on empirical outcomes from training in DiffAero and real-world deployment, with no equations, derivations, or fitted parameters presented in the provided text. The statement that the reconstruction objective 'forces the latent state to encode local geometry...' is a design assumption rather than a reduction of a prediction to its inputs by construction. No self-citations, uniqueness theorems, or ansatzes are invoked in a load-bearing way. Performance gains are reported as experimental results compared to baselines, without any renaming of known results or fitted-input predictions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no equations, training details, or modeling assumptions, so the ledger cannot be populated.

pith-pipeline@v0.9.1-grok · 5835 in / 1159 out tokens · 46434 ms · 2026-06-28T06:21:28.949741+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 3 linked inside Pith

  1. [1]

    Minimum snap trajectory generation and control for quadrotors,

    D. Mellinger and V . Kumar, “Minimum snap trajectory generation and control for quadrotors,” inProc. IEEE Int. Conf. Robot. Autom., Shanghai, China, Aug. 2011, pp. 2520–2525

    2011

  2. [2]

    Safety-assured high-speed navigation for MA Vs,

    Y . Ren, F. Zhu, G. Lu, Y . Cai, L. Yin, F. Kong, J. Lin, N. Chen, and F. Zhang, “Safety-assured high-speed navigation for MA Vs,”Sci. Robot., vol. 10, no. 98, p. eado6187, 2025

    2025

  3. [3]

    NavRL: Learning safe flight in dynamic environments,

    Z. Xu, X. Han, H. Shen, H. Jin, and K. Shimada, “NavRL: Learning safe flight in dynamic environments,”IEEE Robot. Autom. Lett., vol. 10, no. 4, pp. 3668–3675, 2025

    2025

  4. [4]

    Learning vision-based agile flight via differentiable physics,

    Y . Zhang, Y . Hu, Y . Song, D. Zou, and W. Lin, “Learning vision-based agile flight via differentiable physics,”Nat. Mach. Intell., vol. 7, no. 6, pp. 954–966, Jun. 2025

    2025

  5. [5]

    You only plan once: A learning-based one-stage planner with guidance learning,

    J. Lu, X. Zhang, H. Shen, L. Xu, and B. Tian, “You only plan once: A learning-based one-stage planner with guidance learning,”IEEE Robot. Autom. Lett., vol. 9, no. 7, pp. 6083–6090, 2024

    2024

  6. [6]

    Champion-level drone racing using deep reinforcement learn- ing,

    A. Loquercio, E. Kaufmann, S. Ramakrishnan, M. M ¨uller, and D. Scara- muzza, “Champion-level drone racing using deep reinforcement learn- ing,”Nature, vol. 620, no. 7976, pp. 982–987, 2023

    2023

  7. [7]

    Reach- ing the limit in autonomous racing: Optimal control versus reinforcement learning,

    Y . Song, A. Romero, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Reach- ing the limit in autonomous racing: Optimal control versus reinforcement learning,”Sci. Robot., vol. 8, no. 82, p. eadg1462, 2023

    2023

  8. [8]

    Dream to control: Learning behaviors by latent imagination,

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi, “Dream to control: Learning behaviors by latent imagination,” inProc. Int. Conf. Learn. Represent., 2020

    2020

  9. [9]

    Mastering Atari with discrete world models,

    D. Hafner, T. P. Lillicrap, M. Norouzi, and J. Ba, “Mastering Atari with discrete world models,” inProc. Int. Conf. Learn. Represent., 2021

    2021

  10. [10]

    Mastering diverse control tasks through world models,

    D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap, “Mastering diverse control tasks through world models,”Nature, pp. 1–7, 2025

    2025

  11. [11]

    TD-MPC2: Scalable, robust world models for continuous control,

    N. Hansen, H. Su, and X. Wang, “TD-MPC2: Scalable, robust world models for continuous control,” inProc. Int. Conf. Learn. Represent., 2024

    2024

  12. [12]

    Day- Dreamer: World models for physical robot learning,

    P. Wu, A. Escontrela, D. Hafner, P. Abbeel, and S. Levine, “Day- Dreamer: World models for physical robot learning,” inProc. Conf. Robot Learn., 2022

    2022

  13. [13]

    CORB-Planner: Corridor as observations for rl planning in high-speed flight,

    Y . Zhang, B. Gao, G. Wang, J. Sun, and Z. Li, “CORB-Planner: Corridor as observations for rl planning in high-speed flight,”IEEE/ASME Trans. Mechatron., 2025

    2025

  14. [14]

    Learning agility adaptation for flight in clutter,

    G. Zhao, T. Wu, Y . Chen, and F. Gao, “Learning agility adaptation for flight in clutter,”arXiv:2403.04586, 2024

    arXiv 2024

  15. [16]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,”arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  16. [17]

    Accelerated policy learning with parallel differentiable simulation,

    J. Xu, V . Makoviychuk, Y . Narang, F. Ramos, W. Matusik, A. Garg, and M. Macklin, “Accelerated policy learning with parallel differentiable simulation,” inProc. Int. Conf. Learn. Represent., 2021

    2021

  17. [18]

    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,”Sci. Robot., vol. 6, no. 59, p. eabg5810, 2021

    2021

  18. [19]

    Robust navigation for racing drones based on imitation learning and modularization,

    T. Wang and D. Eui Chang, “Robust navigation for racing drones based on imitation learning and modularization,” inProc. IEEE Int. Conf. Robot. Autom., 2021, pp. 13 724–13 730

    2021

  19. [20]

    Deep drone racing: From simulation to reality with domain randomization,

    A. Loquercio, E. Kaufmann, R. Ranftl, A. Dosovitskiy, V . Koltun, and D. Scaramuzza, “Deep drone racing: From simulation to reality with domain randomization,”IEEE Trans. Robot., vol. 36, no. 1, pp. 1–14, 2020

    2020

  20. [21]

    Learning vision- based flight in drone swarms by imitation,

    F. Schilling, J. Lecoeur, F. Schiano, and D. Floreano, “Learning vision- based flight in drone swarms by imitation,”IEEE Robot. Autom. Lett., vol. 4, no. 4, pp. 4523–4530, 2019

    2019

  21. [22]

    A hierarchical reinforcement learning algorithm based on attention mechanism for UA V autonomous navigation,

    Z. Liu, Y . Cao, J. Chen, and J. Li, “A hierarchical reinforcement learning algorithm based on attention mechanism for UA V autonomous navigation,”IEEE Trans. Intell. Transp. Syst., vol. 24, no. 11, pp. 13 309– 13 320, 2023

    2023

  22. [23]

    Depth transfer: Learning to see like a simulator for real-world drone navigation,

    H. Yu, C. D. Wagter, and G. C. H. E. de Croon, “Depth transfer: Learning to see like a simulator for real-world drone navigation,” arXiv:2505.12428, 2025

    arXiv 2025

  23. [24]

    Memory-based deep reinforcement learning for obstacle avoidance in UA V with limited environment knowledge,

    A. Singla, S. Padakandla, and S. Bhatnagar, “Memory-based deep reinforcement learning for obstacle avoidance in UA V with limited environment knowledge,”IEEE Trans. Intell. Transp. Syst., vol. 22, no. 1, pp. 107–118, 2021

    2021

  24. [25]

    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 Trans. Robot., vol. 40, pp. 3044–3067, 2024

    2024

  25. [26]

    Seeing through pixel motion: Learning obstacle avoidance from optical flow with one camera,

    Y . Hu, Y . Zhang, Y . Song, Y . Deng, F. Yu, L. Zhang, W. Lin, D. Zou, and W. Yu, “Seeing through pixel motion: Learning obstacle avoidance from optical flow with one camera,”IEEE Robot. Autom. Lett., vol. 10, no. 6, pp. 5871–5878, 2025

    2025

  26. [27]

    Dream to fly: Model-based reinforcement learning for vision-based drone flight,

    A. Romero, A. Shenai, I. Geles, E. Aljalbout, and D. Scaramuzza, “Dream to fly: Model-based reinforcement learning for vision-based drone flight,”arXiv:2501.14377, 2025

    Pith/arXiv arXiv 2025

  27. [28]

    SkyDreamer: Interpretable end-to-end vision-based drone racing with model-based reinforcement learning,

    A. Verraest, S. Bahnam, R. Ferede, G. de Croon, and C. D. Wagter, “SkyDreamer: Interpretable end-to-end vision-based drone racing with model-based reinforcement learning,”arXiv:2510.14783, 2025

    arXiv 2025

  28. [29]

    STORM: Efficient stochastic transformer based world models for reinforcement learning,

    W. Zhang, G. Wang, J. Sun, Y . Yuan, and G. Huang, “STORM: Efficient stochastic transformer based world models for reinforcement learning,” inProc. Adv. Neural Inf. Process. Syst., 2023

    2023

  29. [30]

    Mastering Atari, Go, Chess and Shogi by planning with a learned model,

    J. Schrittwieser, I. Antonoglou, T. Hubert, K. Simonyan, L. Sifre, S. Schmitt, A. Guez, E. Lockhart, D. Hassabis, T. Graepelet al., “Mastering Atari, Go, Chess and Shogi by planning with a learned model,”Nature, vol. 588, no. 7839, pp. 604–609, 2020

    2020

  30. [31]

    Mastering Atari games with limited data,

    W. Ye, S. Liu, T. Kurutach, P. Abbeel, and Y . Gao, “Mastering Atari games with limited data,” inProc. Adv. Neural Inf. Process. Syst., A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, Eds., 2021

    2021

  31. [32]

    Temporal difference learning for model predictive control,

    N. Hansen, X. Wang, and H. Su, “Temporal difference learning for model predictive control,” inProc. Int. Conf. Mach. Learn., 2022

    2022

  32. [33]

    Diffusion for world modeling: Visual details matter in atari,

    E. Alonso, A. Jelley, V . Micheli, A. Kanervisto, A. Storkey, T. Pearce, and F. Fleuret, “Diffusion for world modeling: Visual details matter in atari,” inProc. Adv. Neural Inf. Process. Syst., 2024

    2024

  33. [34]

    V oxgraph: Globally consistent, volumetric mapping using signed distance function submaps,

    V . Reijgwart, A. Millane, H. Oleynikova, R. Siegwart, C. Cadena, and J. Nieto, “V oxgraph: Globally consistent, volumetric mapping using signed distance function submaps,”IEEE Robot. Autom. Lett., vol. 5, no. 1, pp. 227–234, 2020

    2020

  34. [35]

    Occupancy grid mapping without ray-casting for high-resolution LiDAR sensors,

    Y . Cai, F. Kong, Y . Ren, F. Zhu, J. Lin, and F. Zhang, “Occupancy grid mapping without ray-casting for high-resolution LiDAR sensors,”IEEE Trans. Robot., vol. 40, pp. 172–192, 2024

    2024

  35. [36]

    ROG-Map: An efficient robocentric occupancy grid map for large-scene and high-resolution LiDAR-based motion planning,

    Y . Ren, Y . Cai, F. Zhu, S. Liang, and F. Zhang, “ROG-Map: An efficient robocentric occupancy grid map for large-scene and high-resolution LiDAR-based motion planning,” inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst., Abu Dhabi, United Arab Emirates, Oct. 2024, pp. 8119– 8125

    2024

  36. [37]

    Slope inspection under dense vegetation using LiDAR-based quadrotors,

    W. Liu, Y . Ren, R. Guo, V . W. Kong, A. S. Hung, F. Zhu, Y . Cai, H. Wu, Y . Zou, and F. Zhang, “Slope inspection under dense vegetation using LiDAR-based quadrotors,”Nat. Commun., vol. 16, no. 1, p. 7411, 2025

    2025

  37. [38]

    Training agents inside of scalable world models,

    D. Hafner, W. Yan, and T. Lillicrap, “Training agents inside of scalable world models,”arXiv:2509.10247, 2025

    Pith/arXiv arXiv 2025

  38. [39]

    EGO-Planner: An ESDF- free gradient-based local planner for quadrotors,

    X. Zhou, Z. Wang, H. Ye, C. Xu, and F. Gao, “EGO-Planner: An ESDF- free gradient-based local planner for quadrotors,”IEEE Robot. Autom. Lett., vol. 6, no. 2, pp. 478–485, 2021

    2021